Skip to content
Insights & Articles

Practical knowledge for business AI.

Research-informed perspectives, practical guidance, and useful business AI knowledge for US and EU audiences.

Articles that cover rules or practices tied to one region are labeled United States or European Union. Every article is reviewed before it is published.

Some articles are English translations of our Italian originals. Those link to the Italian version.

  • AI for Small Businesses
  • Business Automation
  • AI Governance & Responsible AI
  • Digital Productivity
  • Data Privacy & Cybersecurity
  • AI Tools & Business Applications

Data Privacy & Cybersecurity

Imagine investing heavily to lock the front door of your business, only to discover that the information you were protecting was stolen through a service entrance managed by someone else. The Maileva incident shows why vendor risk needs ongoing attention.

US & EU · NOVEMBER 15, 2026 · 13 MIN READ

Imagine investing heavily to lock the front door of your business, only to discover that the information you were protecting was stolen through a service entrance managed by someone else.

Your company’s perimeter is no longer just your office walls or your own servers. Your data lives in payroll systems, email marketing platforms, cloud storage, and customer relationship management software. When one of those providers suffers a breach, the consequences can reach your business, even if your own systems remain secure.

The Maileva incident in France illustrates that exposure. On November 14, 2026, La Poste confirmed to AFP that a cyberattack in mid-September had resulted in the theft of names and postal addresses belonging to approximately 22 million people. These individuals were recipients of communications sent by Maileva’s business customers. They did not necessarily choose the platform or know it was handling their information.

For small businesses across Europe and the United States, the lesson is clear: outsourcing your software does not mean outsourcing all your risk.

You may not control the provider’s servers. But when customers need an explanation, employees cannot access a service, or payments stop moving, your business still has to respond.

Information reviewed as of November 15, 2026. This article provides general business guidance, not legal or cybersecurity advice for a specific incident.

The Ripple Effect: From Exposed Data to Interrupted Business

Maileva is a business-mailing platform. Its work sounds routine: helping organizations send correspondence. Yet a routine task performed for many customers can create a substantial concentration of personal information.

According to the company statement reported by AFP, the confirmed exposure included first names, last names, and postal addresses. Banking details were not involved in that disclosure. The figure should not be described as 22 million passwords, bank accounts, or complete customer files, and the public material reviewed does not establish the precise technical entry point.

Those limits do not make the incident irrelevant. Names and addresses can help make impersonation more credible when combined with information obtained elsewhere. A caller who knows a customer’s details is not necessarily legitimate. That is a possible consequence of exposure, not evidence that the Maileva information has already been used in such attacks.

But confidentiality is only one part of vendor risk. The next question is what happens when the service itself stops.

The 2024 Change Healthcare attack demonstrated that problem in the United States. UnitedHealth Group’s public updates described disruption to claims, pharmacy, and payment services, alongside an investigation into exposed personal and health information.

For a smaller healthcare practice dependent on those services, the impact reached beyond privacy. Claims still needed processing. Staff still needed paying. Ordinary expenses continued while essential functions were being restored.

You do not need to operate a clinic to recognize that dependency. Your accounting platform, order-management system, or payroll service may occupy a similar position in your business.

A Secure Cloud Does Not Automatically Mean a Secure Account

A third example shows why the relationship cannot be reduced to whether the provider is reputable.

During the 2024 campaign against Snowflake customer environments, Mandiant traced the incidents it investigated to compromised customer credentials. It found no evidence that the access resulted from a breach of Snowflake’s enterprise environment. Contributing factors included accounts without multifactor authentication and previously stolen credentials that remained valid.

The distinction matters. The provider’s infrastructure and the customer’s account are not the same security boundary.

A platform may offer strong controls while leaving the customer responsible for enabling them, deciding who becomes an administrator, and approving connections to other applications. Buying a capable service does not automatically configure it safely.

Together, these cases reveal three different exposures:

  • Concentrated information: an ordinary service can hold data about many organizations’ customers.
  • Operational dependency: if a critical provider becomes unavailable, your ability to work or receive payment may be affected.
  • Customer-controlled access: a secure platform can still contain accounts or integrations that your business has not adequately protected.

These are not identical attacks, and they do not prove that outsourcing is inherently unsafe. They show why the relationship needs management after the subscription has been purchased.

The Three Traps Small Businesses Fall Into

The following are common organizational patterns, not findings about what caused the Maileva incident.

Digital Hoarding: Keeping Everything “Just in Case”

Storage is inexpensive, so deleting information rarely feels urgent. Old customer exports, identity documents, and outdated mailing lists remain in shared folders because somebody might need them later.

Over time, the business can lose track of why each copy exists. If an account or service is compromised, the available information may extend far beyond what is needed for current work.

The answer is not indiscriminate deletion. Some records must be retained for tax, employment, contractual, litigation, or other reasons. Start by separating information kept for a defined purpose from information kept because nobody has made a decision.

Reducing unnecessary retention will not prevent every intrusion. It can reduce what is available to steal.

Ghost Accounts: Access That Outlasts the Work

A contractor finishes a project, an employee leaves, or a tool is replaced. Their access remains because removing it never becomes urgent.

A former employee’s login may still work. An application connected for one campaign may continue reading customer records. A temporary administrator permission may become permanent.

Access management means keeping permissions aligned with current responsibilities, not simply creating accounts when people join.

The “Someone Else Is Handling It” Assumption

The owner assumes the IT contractor checks the settings. The contractor assumes the provider enforces them. The provider’s documentation explains that the customer must enable them.

Nobody necessarily intended to leave a gap. Responsibility was never made explicit.

A named internal owner helps close that gap. This person does not need to perform every technical task. They need to know who performs it, what remains unresolved, and whether the work has been completed.

Your Three-Step Vendor Survival Guide

You do not need an enterprise compliance department to begin. A small business can make meaningful progress by focusing on a few critical services and agreeing on a routine that people can actually maintain.

Step 1: Map Your “Life Support” Apps

Do not start by auditing every application your team has ever used. Begin with services holding sensitive information or supporting work that cannot easily stop: email, payroll, accounting, payments, customer management, and industry-specific software.

Create a short record for each one.

RecordWhat to Capture
Business purposeThe activity that depends on the service
Information heldThe customer, employee, financial, or other records involved
Internal ownerThe person responsible for the relationship
AccessAdministrators, users, and connected applications
Emergency contactVerified provider support or security details
FallbackAvailable exports, backups, and alternative working arrangements

For payroll, for example, record who administers the account, whether an outside accountant has access, and how payroll would be prepared during an outage.

The goal is not a perfect spreadsheet. It is to avoid discovering these dependencies during an emergency.

Once the important services are visible, review relevant provider evidence. Contracts, independent assurance reports, certifications, and technical documentation can help establish what the provider handles and what remains with your business.

A SOC 2 Type II report may be useful, but its scope, period, and exceptions matter. It is evidence to evaluate, not a guarantee against a breach.

Step 2: Clean House: Access and Data

Work with whoever manages your systems to review the critical accounts first.

Confirm appropriate multifactor authentication, prioritizing administrator, email, and remote-access accounts. This adds another proof beyond a password. Phishing-resistant methods are preferable where available and appropriate.

Then review former users, unnecessary administrator rights, and unused connections to other applications. Confirm an unfamiliar account’s purpose before disabling it: a rushed cleanup can interrupt legitimate operations.

Keep a brief record of changes. Give unresolved issues a responsible person and target date.

Apply the same discipline to information. Identify duplicate exports and outdated records, check retention requirements, and agree on approved deletion. Active files, archives, and backups may follow different schedules.

You cannot eliminate all exposure. You can avoid protecting unnecessary copies indefinitely.

Step 3: Rehearse the “What If They Go Offline?” Scenario

Choose one essential provider and work through a short outage.

For an appointment-based business, that could mean accessing tomorrow’s schedule and recording changes. For an accounting team, it could mean retrieving records and managing pending work without the main platform.

Open an export. Test a restoration in a safe environment. Confirm that the records are usable rather than assuming a successful download proves recovery is possible.

Then consider the people and process: who makes decisions, where temporary records are stored, and how they will be reconciled afterward.

The fallback must protect information too. Copying unnecessary personal details into an unsecured spreadsheet can create a second problem while trying to solve the first.

Backups Are Necessary, but Not the Whole Plan

Backups remain essential for recovering from deletion, corruption, or ransomware. They do not necessarily replace the provider’s software, integrations, or transaction network.

They also cannot retrieve information already copied by an attacker or prevent its publication.

That does not make backups obsolete. It means the business needs to distinguish three tasks:

  • Protect confidentiality: restrict access and reduce unnecessary information.
  • Recover data: maintain protected backups or exports and test them.
  • Continue operating: establish an alternative process while systems are unavailable.

Encryption has limits as well. It may protect data in particular circumstances, but an attacker using a legitimate account may be able to read information that the account is authorized to access.

No single safeguard covers every part of the problem.

What to Do When a Vendor Announces a Breach

A provider’s notice may arrive before the investigation is complete. Your business may need to make decisions while some details remain unknown.

1. Verify the Notice

Use an established support channel or access the provider’s website independently rather than relying on unexpected links. Confirm that the message is genuine, preserve it, and create an incident record.

A real breach can itself become material for phishing attempts.

2. Establish Your Specific Exposure

A headline may describe millions of affected people. You need to know which services and information relate to your organization.

Request the relevant dates, data categories, customer-specific findings, containment measures, and required actions. Record what remains unknown and when another update is expected.

Your technical contact should assess whether credentials, sessions, connected applications, or systems need attention. Preserve relevant evidence while containing the problem.

3. Bring in the Right Advisers

Legal, contractual, and insurance requirements may apply before every forensic detail is available. Do not assume the provider has handled all obligations on your behalf.

In the United States, notification requirements vary by state and sector. The information involved, contractual roles, and circumstances affect the analysis.

For organizations subject to the GDPR, controllers generally must notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of a personal data breach, unless it is unlikely to result in a risk to individuals’ rights and freedoms.

Communication to affected individuals follows a separate high-risk test and applicable exceptions. Processors also have obligations, including notifying controllers without undue delay.

The risk assessment informs whether notification is required; it should not be treated as a reason to postpone considering the deadline. Document the assessment and obtain advice appropriate to the incident.

4. Communicate Without Guessing

Where communication is required or appropriate, explain what is confirmed, what remains under investigation, what the company is doing, and what affected people should do next.

Advice must match the exposure. Do not recommend password changes when passwords were not involved, or promise that exposed information cannot be misused.

Restoring service does not establish that stolen information has been recovered from an attacker. Keep those messages separate.

A Short Checklist to Keep

  • Know the dependencies: identify providers holding important information or supporting essential work.
  • Name an owner: assign responsibility for each critical service.
  • Review access and retention: remove what is no longer needed, subject to operational and legal requirements.
  • Test the fallback: confirm usable records and a safe alternative working process.
  • Prepare the response: keep verified contacts and decision-making responsibilities accessible.

Conclusion

Outsourcing software is often smart business. Treating the provider relationship as something that needs no further attention is where the blind spot develops.

You may not own the servers, investigate the intrusion, or repair the platform. But you still have a relationship with the people whose information you collected and the employees who depend on that service to do their jobs.

Protecting that relationship begins with ordinary decisions: closing an account that should no longer work, removing an unnecessary export, testing a backup, and saving the contact needed during an incident.

Your front door still matters. So does knowing who manages the other entrances.

Sources

Data Privacy & Cybersecurity

Federal prosecutors allege that criminals impersonated executives and lawyers to persuade employees to transfer millions. For small businesses, the lesson is not simply to distrust unfamiliar messages. It is to verify payments outside the conversation that requested them.

UNITED STATES · NOVEMBER 15, 2026 · 12 MIN READ

Approximately $129 million is at the center of an alleged corporate impersonation scheme announced by federal prosecutors in New York on November 5, 2026. According to the Department of Justice, the defendants used technology to mimic executives’ voices or images and impersonated attorneys to persuade employees to transfer money for supposedly confidential financial transactions. Three people were charged. The defendants are presumed innocent unless and until proven guilty.

The alleged victims included healthcare organizations, financial businesses, and a pet food company. The exposure was not confined to one industry.

What makes the case particularly relevant to business owners is the structure of the deception. Employees were not simply approached by someone pretending to be the CEO. They were directed to supposed outside lawyers who reinforced the story and supplied payment instructions. The request and the apparent confirmation belonged to the same alleged scheme.

A small business may never handle transactions of this size. But a diverted supplier payment or fraudulent payroll change can still disrupt cash flow, delay orders, and leave the company paying the legitimate bill a second time. The defense must work even when the message sounds convincing.

Reporting reviewed as of November 15, 2026. This article provides general information, not legal, banking, or insurance advice.

The True Target: Your Approval Process

In the New York case, prosecutors describe a sequence built around apparent authority and confidentiality. An employee was approached by someone impersonating a senior executive, then directed to an apparent legal adviser involved in the transaction. According to the allegations, instructions discouraged discussion with colleagues.

The public announcement does not establish which AI products were used or how every impersonation was produced. It does describe the central mechanism: making a payment request appear credible through multiple trusted identities.

A company can lose money this way without an attacker taking over its bank account. A legitimate employee may enter the payment, an authorized signer may approve it, and the bank may receive instructions through the business’s normal portal.

The deception occurs earlier, when the company decides that the payment is genuine. Security that confirms who signed into the account does not necessarily establish whether the transaction they approved was commercially legitimate.

The Italian Parallel

A similar pattern appeared in reporting from Italy.

In September 2026, Italian news organizations described a February fraud involving Fideuram, part of Intesa Sanpaolo. According to ANSA’s account, drawing on reporting by Corriere della Sera, an apparent message from a senior executive was reinforced by a phone call impersonating a known lawyer.

Reports described approximately €95 million in outgoing transfers and a substantial partial recovery. Published figures for the remaining loss differed, so they should not be treated as a definitive accounting.

The Italian case is a reported investigation. The sources used for this article do not establish a connection to the New York defendants. Its relevance is operational: a familiar executive identity and an apparent professional adviser can reinforce each other. Hearing from two people is not an independent check if both contacts originate inside the same fraudulent conversation.

What AI Changes

Business impersonation fraud existed before generative AI. Criminals have long used lookalike email addresses, altered invoices, and compromised accounts to redirect legitimate payments. These schemes are commonly grouped under business email compromise, or BEC.

AI can strengthen parts of the deception. Written messages can be fluent and tailored to a business context. Synthetic audio can resemble a familiar speaker. Manipulated video can make someone appear to participate in a meeting.

The important change is not that callbacks have become useless. It is that a familiar voice is not, by itself, proof of identity. An incoming call or a call placed to a number supplied in the suspicious request can be another part of the attack. A separately initiated call to an established contact creates a different verification path.

Not every fraudulent message uses AI. Caller-ID spoofing, for example, changes the number displayed on the recipient’s screen and does not necessarily involve artificial intelligence. A genuine email account can also be compromised, allowing a criminal to send instructions from an address the business already trusts.

The FBI’s 2025 Internet Crime Report recorded approximately $3.05 billion in reported BEC losses. That is a broader fraud category, not a measure of losses caused exclusively by AI. The payment control needs to work whether the message was written by a person, generated by AI, or sent through a compromised account.

The Illusion of Independent Confirmation

A second conversation often feels reassuring. A supposed lawyer explains the transaction, a vendor’s “new billing department” confirms the account, or someone who appears to be the owner repeats the instruction on a call.

But a contact supplied by the requester is not automatically independent evidence. If the email includes a verification number, calling it may connect the employee to another participant in the fraud. The same applies to a meeting link or an adviser introduced inside the suspicious conversation.

Independent verification means choosing a trusted contact without relying on the request itself.

For an established vendor, that usually means using information recorded before the change request arrived. For an executive, it means using the company’s existing directory or speaking in person. Switching from email to a call is not enough if the attacker controls the destination.

Pause, Verify, Approve, Record

A small business does not need to treat every invoice as an emergency investigation. It needs a repeatable process for changes and exceptions: new recipients, revised banking instructions, unusual executive requests, and transactions outside normal purchasing arrangements.

StepWhat the business does
PauseKeep the request unverified and avoid immediately changing stored payment details.
VerifyConfirm the instruction through independently selected, previously verified contacts.
ApproveReview the business purpose, recipient, and banking details before releasing funds.
RecordPreserve a short account of the verification and approval process.

These are recommended organizational controls. They should be adapted to the company’s banking services, staffing, and payment risks.

Vendor Updates: Verify Before Editing

A routine invoice can become dangerous when accompanied by new remittance instructions.

For US businesses, the change may involve an ACH routing number, bank account number, or wire instructions. ACH is the electronic payment network commonly used for vendor payments and direct deposits; wires follow a different payment process. Recovery options and deadlines are not interchangeable.

The first safeguard is to leave the existing vendor record unchanged while the new instructions are checked.

Keep the request separate

Save the email or document, but do not immediately overwrite the accounting system’s bank details. Once an unverified account becomes the default, later payments may be sent there without anyone revisiting the original change.

Use an established contact

Retrieve the number from the vendor file, original onboarding documents, or another established source. Do not use a “new accounts receivable number” supplied in the same message requesting the banking change.

Reach a person authorized to confirm the vendor’s payment instructions. Confirm that the change was requested and verify the replacement details. A previously stored contact is stronger than an incoming call, but not infallible; unusual circumstances may require another established contact.

Review before activation

Where staffing permits, one person verifies the request and another approves the banking update. The reviewer should inspect the verification record, not simply accept that someone else “spoke to the vendor.”

For a new vendor, collect contact and payment information through a deliberate onboarding process. Tax forms, contracts, and professional-looking documents may support that process, but they do not independently prove ownership of a bank account.

Record the change

A short log should identify:

  • the vendor and date of the request;
  • the employee who performed the check;
  • the contact information used;
  • the person who confirmed the change;
  • the result and approving person.

Access to vendor records should be restricted so contact numbers and banking details cannot be changed without an appropriate record.

If verification fails, hold the change and escalate it. Do not automatically pay the old account either: the vendor may genuinely have closed or restricted it.

Handling the Owner’s “Urgent Wire” Request

Payment controls should apply to the owner’s requests as well as everyone else’s. An employee receiving an unusual instruction should have explicit authority to pause the conversation and contact the executive through a known number or in person.

With AI voice cloning, a familiar-sounding caller is no longer sufficient proof of identity. Follow the verification procedure even when the voice sounds exactly like the owner.

The confirmation must cover the transaction itself: amount, recipient, purpose, and banking instructions. If the executive is unavailable, the company needs an existing escalation route, such as a designated co-owner, finance lead, or other authorized reviewer. Without an approved alternative, the payment remains on hold.

Confidentiality may justify limiting who sees the information. It should not eliminate verification. Identify trusted reviewers or an existing legal adviser before an exceptional transaction arises. An adviser introduced during the suspicious exchange is not, by that introduction alone, a trusted verification source.

The owner must support the rule when it slows a genuine request. Otherwise, employees learn that bypassing the process is what leadership actually expects.

Two Approvals, Two Actual Checks

A second approval offers little protection when both people rely on the same fraudulent message.

The reviewer needs supporting records and should check what the payment purchases or settles, whether the recipient is expected, where the banking instructions originated, and whether any recent change was independently confirmed.

Ask the bank which separation-of-duty features are available. Some business services allow one user to prepare a payment and another to release it. Confirm what those settings actually enforce rather than assuming an internal policy is reflected in the portal.

Credentials and authentication devices should remain individual. Sharing the owner’s login or approval phone can undermine the intended separation.

Make Controls Work for Small Teams

A three-person business may not have a controller and a separate accounts payable department. The company can still distinguish established payments from changes and exceptions, focusing additional verification on new recipients, revised accounts, and unusual instructions.

If the owner handles everything, a pause followed by a separately initiated vendor check provides useful friction. An external accountant or adviser may assist where their role, access, and responsibility have been agreed in advance.

There is no universal dollar threshold. A payment that is routine for one business may threaten another’s operating cash. An automatic overnight delay is not a complete safeguard either: waiting helps only if someone uses the time to verify the instruction.

Protect Payroll Changes

An employee’s direct-deposit change can also be impersonated.

Where a protected payroll self-service portal is available, employees should normally update their details through that established process rather than through an unexpected email attachment. For manually administered changes, use an existing employee contact or an in-person process, not a new number supplied alongside the instructions.

Where supported, send a change notification through an established channel so the employee can identify an unauthorized modification.

Suspicious changes near payroll cutoff need prompt escalation. Do not introduce arbitrary delays that could interfere with wage-payment obligations. Establish a verified alternative process that preserves both security and timely payment.

A Real Email Address Is Not Enough

Checking the sender’s address is useful, but it cannot reveal whether a legitimate account has been taken over.

Protect email, payroll, and banking access with multifactor authentication, using phishing-resistant methods where available. The person managing company systems should also be able to review active sessions, unexpected forwarding rules, connected applications, recovery settings, and permission changes.

These controls reduce account-compromise risk. They do not make every message from the account trustworthy. Bank-detail verification should still apply when the request arrives inside a familiar email thread.

Damage Control: If the Payment Already Left

Act immediately. Do not wait to establish which technology was used.

Contact the sending bank

Use the bank’s established fraud or business-support channel. Explain that the payment may have resulted from impersonation fraud and provide the amount, date, time, recipient, account information, and transaction reference.

Ask what action is available for that payment type, including a wire recall, recipient-bank contact, or an applicable ACH process. Follow the bank’s instructions and supply requested documentation promptly. Recovery is not guaranteed.

Report to IC3

The FBI’s Internet Crime Complaint Center accepts reports of cyber-enabled crime. Include accurate transaction and banking information, save a copy of the submitted complaint, and retain supporting evidence.

Do not delay because the amount seems small or some details remain unclear. For urgent cases, contact the appropriate law-enforcement office directly as well; an online report is not a promise of immediate personal contact.

Preserve records and contain access

Keep original electronic emails where possible, together with attachments, messages, call logs, bank confirmations, and a timeline. Screenshots can help, but should not replace original records.

If an account may be compromised, involve the IT provider to contain access and preserve relevant information. Do not leave an attacker connected simply to collect evidence, but avoid uncoordinated deletion or device wiping.

Notify advisers and the insurer

Review insurance reporting requirements promptly. Coverage depends on the policy’s terms, exclusions, limits, and conditions.

For funds sent abroad, FinCEN’s Rapid Response Program may support recovery through law enforcement and foreign partners. Victims do not activate it by contacting FinCEN directly: official guidance directs them to their financial institution and law enforcement.

There is no safe waiting period. References to a 24- or 72-hour window should never be interpreted as permission to postpone action.

Payment Fraud Is Not Always a Data Breach

An impersonation attempt can occur without anyone accessing company systems. A compromised mailbox, however, may expose customer, employee, or other protected information as well as facilitate payment fraud.

Assess both where the money went and whether systems or information were accessed improperly. Notification obligations depend on the facts and applicable law. Neither the absence of ransomware nor the presence of a fraudulent transfer settles that question.

Do Not Assume Reimbursement

Commercial payment disputes depend on the payment method, agreements, applicable law, and circumstances. A payment entered by an authorized employee under deception should not automatically be treated as legally identical to an outsider issuing an unauthorized payment order.

Obtain advice on the specific incident rather than relying on a general promise of reimbursement.

Insurance also needs attention before an incident. Ask the broker whether the policy addresses social-engineering fraud, fraudulent transfer instructions, and payments made by deceived employees. A smaller coverage limit may apply even when the overall cyber policy has a much higher headline limit.

Test Before the Next Request

A useful exercise does not require an expensive simulation.

Give the team a clearly fictional situation: a regular vendor sends new bank details shortly before payment is due, but the usual contact cannot be reached. Walk through the actual response. Locate the trusted number, identify who can approve a change, and confirm who handles the unresolved request.

The exercise succeeds when it exposes a missing step before a real payment depends on it. Owners must participate: employees will not consistently enforce a control that leadership routinely overrides.

Conclusion

The alleged $129 million scheme illustrates why a convincing request can survive more than one conversation. An apparent executive introduces an apparent lawyer, documents reinforce the explanation, and confidentiality discourages colleagues from comparing notes.

For a small business, the answer is not to become expert at spotting synthetic voices. It is to keep payment authorization separate from the story presented by the requester.

Verify changes through established contacts. Give the second approver something real to check. Protect vendor and payroll records. Define who can stop an unusual payment.

The voice may sound familiar. The payment still needs verification.

5-Point Small Business Protection Checklist

  • Control vendor changes: Restrict edits to payment records and document verification before activating new ACH or wire instructions.
  • Verify independently: Confirm changes using an established contact selected outside the request. Never rely solely on numbers supplied in the new message.
  • Set an executive protocol: Agree in advance how unusual owner or CEO requests will be verified, including confidential and after-hours transactions.
  • Separate preparation and approval: Use banking permissions where available; for very small teams, define a documented alternative check.
  • Practice AI-aware verification: Teach staff that familiar voices, video calls, and genuine email addresses are not conclusive proof, and authorize them to pause a payment.

Sources

AI Governance & Responsible AI

The EU and the US both regulate AI, but through different legal structures. This guide uses recent enforcement cases to show how an SME can build one AI governance program that works on both sides of the Atlantic, from shadow AI and the EU AI Act timetable to US employment rules, Colorado and the NIST AI RMF.

US & EU · NOVEMBER 12, 2026 · 21 MIN READ

In February 2025, the US Federal Trade Commission finalized an order against DoNotPay, a company that had promoted its chatbot as “the world’s first robot lawyer.”

According to the FTC, the company had not adequately tested whether the service could perform at the level of a human lawyer. The final order required DoNotPay to pay $193,000, notify former subscribers and stop making claims that its service could replace a professional unless it possessed sufficient evidence to support them.

Across the Atlantic, European authorities were examining AI services through a different regulatory lens.

In April 2025, the Italian Data Protection Authority fined Luka, the US company behind the Replika chatbot, €5 million. The authority identified problems concerning the legal basis for processing personal data, transparency, data minimization, accountability, privacy by design and age verification.

In July 2026, the same authority fined Character Technologies, the operator of Character.AI, €158,000. It also ordered the company to strengthen age-verification measures, prevent minors with frozen accounts from immediately registering again and make minors’ profiles private by default.

Neither the Replika nor the Character.AI decision was an enforcement action under the EU AI Act. Both were primarily based on European data-protection law. Their significance lies in demonstrating that AI services were already subject to enforceable European rules before the AI Act’s main transparency and governance provisions became applicable.

These cases do not involve identical products or legal violations. Together, however, they illustrate the different routes through which AI-related enforcement develops in the United States and the European Union.

In the United States, regulators often begin with the result: Was the marketing deceptive? Did the system discriminate? Was sensitive information misused? Were consumers or workers harmed?

The European Union considers many of the same questions but increasingly adds a formal governance structure around the AI system itself: its intended purpose, risk category, data processing, transparency, human oversight and position within the AI value chain.

For small and medium-sized enterprises, the lesson is not that one region regulates AI while the other does not. Both do. They simply reach the business through different legal structures. An SME operating on both sides of the Atlantic must be prepared to work within both models.

AI is already operating inside the average SME

An SME does not need to develop a proprietary model to face AI-related responsibilities.

Artificial intelligence may already be present in:

  • customer-service chatbots;
  • email and document assistants;
  • meeting transcription tools;
  • applicant-screening platforms;
  • CRM and marketing automation;
  • fraud-detection systems;
  • accounting and invoice-processing software;
  • productivity and workforce-management tools;
  • dynamic pricing applications;
  • recommendation systems;
  • cybersecurity services;
  • image, video and advertising tools.

Some of these functions may have appeared through ordinary software updates rather than a deliberate AI project.

A CRM provider may activate automated lead scoring. A recruitment platform may introduce candidate ranking. A helpdesk may begin generating replies or summarizing customer conversations. Employees may also use public generative AI services independently, without notifying management.

This creates the first shared problem for American and European SMEs: a business cannot govern an AI system it does not know it is using.

1. Control shadow AI before classifying risk

“Shadow AI” refers to AI systems, features or accounts used without appropriate approval, visibility or governance.

It may involve an employee using a public chatbot to summarize a confidential contract, but the problem is broader than unauthorized standalone tools.

Shadow AI can also include:

  • AI functions activated inside approved software;
  • browser extensions with access to company content;
  • meeting assistants invited to confidential calls;
  • customer information uploaded to writing or translation tools;
  • marketing platforms that automatically generate and target content;
  • AI agents connected to email, storage or business applications;
  • personal AI accounts used for company work;
  • software features enabled by default after an update.

The risk is not created simply because a product contains AI. It depends on what information the tool receives, what it produces, which decisions depend on it and how the provider handles data.

IBM’s 2025 breach research reported that one in five surveyed organizations had experienced a breach connected with shadow AI. The survey involved organizations that had suffered data breaches, so the result should not be presented as a universal rate for all businesses or all SMEs. It nevertheless shows that uncontrolled AI use has become a cybersecurity and data-governance issue, not merely an internal policy concern.

Build a proportionate AI inventory

The first operational step is to create an inventory covering both approved and discovered tools.

Inventory fieldQuestion to answer
ToolWhich product or AI-enabled function is being used?
ProviderWho supplies and operates it?
PurposeWhich business task does it perform?
UsersWhich employees, contractors or departments use it?
DataWhat information enters the system?
OutputWhat content, recommendation or decision does it produce?
ImpactWho may be affected if the output is incorrect?
IntegrationCan it access email, files, CRM records or other systems?
Human reviewWho checks or approves its output?
ContractWhat do the terms and data agreement permit?
GeographyIn which countries or US states is it used?
StatusIs it approved, restricted, under review or prohibited?

Employee questionnaires can help, but they rarely reveal everything.

The company should also review:

  • SaaS administration panels;
  • single sign-on and identity-provider records;
  • browser extensions;
  • expense reports;
  • corporate card purchases;
  • integrations between business systems;
  • AI functions enabled in existing software.

The goal is not to monitor every employee action. It is to identify company systems, data flows and unmanaged exposure.

2. The European model: determine the system’s role and risk

The EU AI Act follows a harmonized, risk-based structure. It does not apply the same obligations to every AI tool.

A grammar assistant, customer chatbot and recruitment-ranking system may all use artificial intelligence, but they can occupy very different regulatory positions.

An SME should first determine:

  1. whether the technology falls within the AI Act’s definition and scope;
  2. whether the business is a provider, deployer, importer, distributor or another operator;
  3. the system’s intended purpose;
  4. whether a prohibited AI practice is involved;
  5. whether the system creates a transparency risk;
  6. whether it may qualify as high-risk;
  7. which other laws, including the GDPR, remain applicable.

Most SMEs purchasing an AI service for internal or commercial use will normally act as deployers. However, the role can change.

A company may assume provider responsibilities if it places a system on the market under its own name, changes its intended purpose or makes certain substantial modifications. Calling a solution “third-party software” does not automatically settle the legal classification.

3. SMEs are supported, but not generally exempt

The AI Act includes measures designed to support SMEs and start-ups. These include regulatory sandboxes, dedicated communication channels, proportionality and simplified requirements in certain areas.

They do not create a general exemption.

A small company may still have obligations when it provides or deploys an affected system. The precise burden depends on the company’s role, the use case and the system’s classification.

It would therefore be inaccurate to say that every SME using AI must produce the complete documentation required from a provider of a high-risk system.

Most ordinary drafting, translation and internal productivity tools will not become high-risk merely because they use AI. Other requirements may still apply, particularly when personal, confidential or protected information is involved.

Proportionality should affect how an SME designs its governance process, but it should not be confused with an automatic exclusion from the law.

4. A US company can still fall within the EU framework

The AI Act is not limited to businesses established in the European Union.

Depending on the circumstances, it may also apply to providers and deployers outside the EU when they:

  • place an AI system on the European market;
  • put a system into service in the EU;
  • use the system within the Union;
  • produce output that is used in the EU.

An American SME should therefore examine its customers, users and deployment context rather than relying only on its place of incorporation.

For example, a US software company that provides an AI recruitment tool used to assess applicants for European jobs cannot assume that its physical location keeps it outside the European framework.

This extraterritorial reach also makes contractual clarity important. A European customer may request information from an American provider about system capabilities, intended purpose, data, monitoring, transparency and human-oversight functions.

5. The AI Act did not become applicable all at once

The EU AI Act follows a phased implementation timetable.

As of November 2026:

  • prohibited AI practices and AI literacy requirements are applicable;
  • rules for general-purpose AI models have entered into application;
  • enforcement powers apply to the provisions currently in force;
  • Article 50 transparency obligations have applied since August 2, 2026;
  • Annex III high-risk rules are scheduled to apply from December 2, 2027;
  • rules for certain high-risk AI systems embedded in regulated products are scheduled to apply from August 2, 2028.

The timetable was adjusted through the 2026 AI Omnibus. Older articles, compliance calendars and vendor summaries may therefore contain outdated dates.

It is misleading to say that the whole AI Act became fully applicable in August 2026. A substantial part of the framework is now enforceable, but important high-risk obligations follow later dates.

For an SME, the correct approach is to build readiness around the actual timetable instead of waiting for every provision to apply at the same time.

6. EU transparency does not mean labeling every use of AI

Article 50 introduces transparency obligations for specified systems and situations.

For example:

  • providers of AI systems designed for direct interaction with individuals must generally design the systems so people are informed, from the beginning of the first interaction, that they are interacting with AI, unless this is already obvious to a reasonably well-informed person;
  • providers of systems that generate synthetic content must support machine-readable marking and detection, subject to the relevant conditions and exceptions;
  • deployers of emotion-recognition or biometric-categorization systems must inform affected individuals;
  • deployers must disclose deepfakes;
  • certain AI-generated or manipulated text published to inform the public on matters of public interest requires disclosure when it has not received the relevant human review or editorial control.

This does not mean that every AI-assisted email, internal summary, translated paragraph or edited sentence needs a public label.

The correct analysis depends on:

  • what the system generated or altered;
  • whether the system interacts directly with a person;
  • whether the interaction with AI is already obvious;
  • how substantially the content was changed;
  • why and where the content is published;
  • whether meaningful human review occurred;
  • whether editorial responsibility is assigned;
  • whether an exception applies.

A generic statement hidden inside a privacy policy may not provide a sufficiently clear notice for a customer-facing chatbot. The information should appear at the point at which it becomes relevant.

What an SME should record

For each customer-facing or content-generating tool, document:

  • who provides the AI system;
  • whether the user interacts directly with it;
  • who is responsible for the notice;
  • where and when the disclosure appears;
  • whether generated content is technically marked;
  • whether the business publishes deepfakes or public-interest information;
  • how human review is performed;
  • which exception, if any, is being relied upon.

7. AI literacy must reflect the actual use case

The AI literacy obligation has applied since February 2025, with supervision and enforcement in place since August 2026.

Following the 2026 amendments, the law does not require every organization to achieve one universal or predefined level of AI literacy. Providers and deployers must take measures to support the development of appropriate AI literacy among staff and other people operating AI systems on their behalf.

The provision does not prescribe:

  • one mandatory course;
  • a universal examination;
  • a specific number of training hours;
  • a formal certificate;
  • a particular governance board;
  • a compulsory AI officer.

Training and guidance should reflect:

  • employees’ technical knowledge;
  • the AI systems they use;
  • the context and purpose of use;
  • foreseeable limitations and errors;
  • the sensitivity of the data involved;
  • the people affected by the system;
  • the degree of human oversight required.

A marketing employee using generative AI to prepare a campaign needs different instruction from an HR manager reviewing algorithmic candidate scores.

A practical internal record may include:

  • date;
  • participants;
  • relevant systems;
  • subjects covered;
  • practical examples;
  • materials provided;
  • person responsible;
  • future review date.

The purpose is to support competent use, not to manufacture training certificates without operational value.

8. The US model: regulate conduct, claims and consequences

The United States still has no single comprehensive federal AI statute equivalent to the EU AI Act.

That does not mean AI is unregulated.

Existing federal laws can apply when AI is used to:

  • deceive consumers;
  • discriminate against workers or applicants;
  • make unsupported claims;
  • misuse personal information;
  • create unfair commercial practices;
  • operate improperly in regulated sectors;
  • produce false reviews or misleading promotional content.

States and cities are introducing additional requirements concerning automated decisions, employment tools, chatbots, biometric information, privacy and high-impact uses.

The DoNotPay case demonstrates one characteristic of the American approach: calling a service “AI-powered” does not excuse unsupported claims about its performance.

The FTC alleged that DoNotPay had not tested whether its chatbot performed at the level of a human lawyer. The final order restricted future professional-substitution claims unless the company could support them with adequate evidence.

Claims that require particular caution

An SME should not describe an AI system as:

  • “completely unbiased”;
  • “fully compliant”;
  • “100% accurate”;
  • “equivalent to a professional”;
  • “guaranteed to select the best candidate”;
  • “incapable of human error”;
  • “fully autonomous and safe”;
  • “protecting all confidential information”;

unless the claim is precise, appropriately qualified and supported by evidence.

A business should also avoid repeating a vendor’s promotional language automatically. Once those claims appear in the SME’s own website, proposal or customer communication, they may become the SME’s representations as well.

9. US employment AI: local rules can reach ordinary employers

Employment is one of the most sensitive AI use cases in both regions.

New York City’s Local Law 144 prohibits covered employers and employment agencies from using an automated employment decision tool unless:

  • a bias audit has been completed within the previous year;
  • a summary of the results has been made publicly available;
  • required notices have been provided to affected candidates or employees.

The law does not cover every digital recruitment function.

Its definition focuses on tools using machine learning, statistical modeling, data analytics or artificial intelligence that substantially assist or replace discretionary decision-making in hiring or promotion.

Federal anti-discrimination law also remains relevant. The US Equal Employment Opportunity Commission has warned that automated employment tools may:

  • screen out applicants with disabilities;
  • create barriers to reasonable accommodation;
  • prompt prohibited disability-related inquiries;
  • disadvantage protected groups;
  • reproduce discrimination even when a third-party vendor supplies the tool.

Purchasing the software from an external provider does not automatically transfer employment-law responsibility away from the employer.

10. The European comparison for recruitment AI

Under the EU AI Act, systems intended for specified recruitment and employment purposes can fall within the high-risk category.

Relevant uses may include:

  • recruitment and selection;
  • filtering applications;
  • evaluating candidates;
  • decisions affecting work relationships;
  • task allocation based on personal characteristics or behavior;
  • monitoring or evaluating employee performance.

The principal Annex III high-risk obligations are scheduled to apply from December 2027 rather than August 2026.

The EU framework is expected to require a structured compliance chain involving providers and deployers. Depending on the system and circumstances, this can involve instructions for use, logs, human oversight, monitoring and controls over how the system is operated.

The US approach remains more geographically fragmented. An employer may need to consider federal discrimination law, state requirements and local rules such as New York City’s audit and notice regime.

What an employer should do now

Before using AI to screen, score or rank applicants:

  1. determine whether the system materially influences the decision;
  2. identify where applicants and positions are located;
  3. request technical and compliance information from the vendor;
  4. understand the data and criteria used;
  5. examine whether the system may exclude protected groups;
  6. provide a reasonable-accommodation route;
  7. check notice and audit requirements;
  8. train reviewers not to accept scores automatically;
  9. record the human reasons supporting the final decision;
  10. monitor the system after deployment;
  11. create a correction or reconsideration process;
  12. establish conditions for suspending its use.

“Human in the loop” is not a complete control by itself.

A reviewer who approves every recommendation without understanding, questioning or having authority to override it does not provide meaningful oversight.

11. Colorado shows how US regulation is becoming more structured

Colorado demonstrates how some US states are moving toward more detailed AI governance while maintaining a different structure from the EU model.

Legislation signed in 2026 regulates automated decision-making technology used to materially influence consequential decisions. It creates requirements for developers and deployers and provides consumers with rights concerning inaccurate personal data used by the technology.

The provisions are scheduled to take effect on January 1, 2027.

Colorado also adopted a Chatbot Safety Act covering certain conversational AI services. It includes disclosure, youth-protection, privacy and safety requirements and is also scheduled to take effect in January 2027.

The chatbot legislation addresses areas such as:

  • informing users that they are interacting with AI rather than a human;
  • estimating users’ ages;
  • protecting younger users from sexually explicit material;
  • reducing simulated emotional dependence;
  • providing privacy and account-management tools;
  • developing self-harm and suicide-response protocols;
  • avoiding representations that chatbot outputs are equivalent to licensed professional services.

This model is more prescriptive than relying only on general consumer-protection law, but it is still not equivalent to a nationwide American AI Act.

An SME must determine where the system operates, who is affected and whether it contributes to consequential decisions. The relevant location may not be the company’s headquarters.

12. NIST AI RMF: a useful baseline, not automatic compliance

The NIST AI Risk Management Framework is one of the most useful resources for organizing an American AI governance program.

It is structured around four functions:

FunctionPractical SME question
GovernWho owns AI risk and which rules guide its use?
MapWhat is the system’s context, purpose and potential impact?
MeasureHow are accuracy, bias, privacy, security and reliability assessed?
ManageWhich controls, responses and monitoring processes are applied?

NIST describes the framework as voluntary, non-sector-specific and adaptable to organizations of different sizes.

Following it does not automatically:

  • establish legal compliance;
  • create a universal statutory safe harbor;
  • replace sector-specific rules;
  • satisfy every state requirement;
  • demonstrate that an individual system is safe or fair.

It is nevertheless valuable because it helps an SME create consistent governance evidence:

  • defined responsibilities;
  • documented use cases;
  • risk assessments;
  • testing criteria;
  • approval decisions;
  • incident procedures;
  • monitoring;
  • corrective action;
  • decisions to restrict or retire a system.

The framework should be mapped to the laws that actually apply. It should not be treated as a substitute for them.

13. Data minimization versus AI’s appetite for context

Modern AI applications often perform better when given more context. Privacy principles frequently push in the opposite direction: the company should use only the information necessary for a defined purpose.

This conflict is especially visible with AI agents connected to company systems.

An agent may be technically capable of searching every email, document, calendar entry and CRM record. That does not mean it should receive unrestricted access.

A sensible permission review asks:

  • Which repository does the agent genuinely need?
  • Which records are relevant to its task?
  • Does it require complete documents or only selected fields?
  • Can personal identifiers be removed?
  • Can sensitive data be masked?
  • How long are prompts and outputs retained?
  • Can the provider use the information for model training?
  • Which subprocessors receive it?
  • Can the agent send messages or modify records?
  • Which actions require human approval?
  • Are access and actions logged?
  • Can the integration be stopped quickly?

A Data Processing Agreement is important when the relationship and applicable law require one. It is not a technical safeguard by itself.

A contract cannot prevent an employee from uploading information improperly. It cannot stop an agent from accessing an unnecessary folder if permissions have been configured incorrectly.

Contracts should be combined with:

  • role-based access;
  • approved company accounts;
  • retention controls;
  • training-data settings;
  • authentication;
  • logging;
  • data-loss prevention where proportionate;
  • employee guidance;
  • periodic technical testing.

14. One governance program with two regulatory mappings

A transatlantic SME does not need two completely separate AI programs.

It needs one operational foundation with distinct legal mappings.

Governance areaEuropean emphasisUS emphasis
System classificationAI Act role, intended purpose and risk categoryUse case, sector, state and affected decision
TransparencyArticle 50 duties for specified systems and contentConsumer disclosures, state chatbot rules and deceptive-practice risks
EmploymentHigh-risk classification and deployer obligations from the applicable dateFederal discrimination law plus state and local requirements
PrivacyGDPR legal basis, minimization, rights and DPIAs where requiredState privacy laws, sectoral rules and FTC enforcement
Marketing claimsConsumer law and applicable AI Act information dutiesFTC substantiation and deceptive-practice enforcement
TrainingExpress, context-based AI literacy obligationRisk-based employee training and evidence of reasonable governance
FrameworkHarmonized regulation supported by standards and guidanceNIST AI RMF and sector-specific controls
EnforcementEU and national authorities operating within a common frameworkFederal, state and local authorities applying different laws

The shared foundation should include:

  • AI inventory;
  • assigned owners;
  • approved-use rules;
  • vendor assessment;
  • data mapping;
  • legal classification;
  • risk review;
  • output testing;
  • human-oversight design;
  • incident reporting;
  • employee training;
  • ongoing monitoring;
  • retained evidence.

Final SME checklist

  • Every approved AI system has an identified owner.
  • Embedded AI features are included in the inventory.
  • Employees can report unauthorized tools without being discouraged from speaking up.
  • Each system has a documented business purpose.
  • The company’s role under the EU AI Act has been considered.
  • The relevant EU risk category has been reviewed.
  • Extraterritorial EU applicability has been considered.
  • Relevant US federal, state and local rules have been mapped.
  • Customer-facing AI provides appropriate notice.
  • Synthetic content is reviewed for applicable disclosure requirements.
  • Recruitment tools have been checked for audit, notice and accommodation duties.
  • Vendor claims have been independently reviewed.
  • Confidential and personal information is restricted.
  • Provider training-data and retention settings have been checked.
  • AI agents use limited permissions.
  • Human reviewers understand how and when to intervene.
  • Employees receive role-specific AI literacy or training.
  • Unexpected outputs and incidents can be escalated.
  • Relevant decisions and evidence are retained.
  • Every higher-impact system has a review date.
  • The company can restrict, suspend or retire a system when necessary.

Conclusion

The European Union and the United States are not moving toward identical AI regulation.

The EU has created a horizontal, risk-based framework that assigns roles, classifies systems and establishes common obligations across the Union. That framework operates alongside the GDPR, consumer law, employment rules and sector-specific legislation.

The United States continues to regulate AI through a combination of consumer protection, civil-rights law, sectoral rules and increasingly detailed state and local legislation.

The operational message, however, is remarkably similar.

An SME must know which AI systems it uses, understand the data they receive, test important claims, protect affected people, supervise consequential decisions and preserve evidence of what it has done.

The DoNotPay case illustrates the danger of promoting AI capabilities that have not been substantiated. The Replika and Character.AI decisions show how privacy, accountability, age verification and system design can become enforcement issues in Europe even when the AI Act is not the legal basis of the case.

AI experimentation has not ended. It remains valuable, particularly for smaller businesses that use technology to compete with larger organizations.

What has ended is the assumption that experimentation can remain invisible, undocumented and disconnected from ordinary business responsibility.

The SMEs best prepared for 2027 will not necessarily be those with the longest policies. They will be those capable of answering five practical questions:

  1. What AI are we using?
  2. What data does it receive?
  3. Who may be affected by its output?
  4. Who checks and controls it?
  5. What evidence shows that the process works?

Sources

  1. EU AI Act, consolidated text, European Union, EUR-Lex: eur-lex.europa.eu
  2. EU AI Act implementation timeline, European Commission AI Act Service Desk: ai-act-service-desk.ec.europa.eu
  3. When AI Act enforcement starts, European Commission AI Act Service Desk: ai-act-service-desk.ec.europa.eu
  4. Transparency obligations under Article 50, European Commission: digital-strategy.ec.europa.eu
  5. Guidelines on transparency obligations, European Commission: digital-strategy.ec.europa.eu
  6. AI literacy questions and answers, European Commission: digital-strategy.ec.europa.eu
  7. FTC final order concerning DoNotPay, Federal Trade Commission: www.ftc.gov
  8. Replika enforcement decision, European Data Protection Board: www.edpb.europa.eu
  9. Character.AI enforcement decision, Italian Data Protection Authority: www.garanteprivacy.it
  10. NIST AI Risk Management Framework, National Institute of Standards and Technology: www.nist.gov
  11. New York City Automated Employment Decision Tools FAQ, NYC Department of Consumer and Worker Protection: www.nyc.gov
  12. AI and disability discrimination in employment, US Equal Employment Opportunity Commission: www.eeoc.gov
  13. Colorado automated decision-making and chatbot rulemaking, Colorado Attorney General: coag.gov
  14. IBM report on AI governance and shadow AI, IBM: newsroom.ibm.com

Data Privacy & Cybersecurity

Privacy enforcement in the US now looks at what a business actually collects and transmits. This guide explains how an SME can prepare for 2027, from California’s expanded deletion right and the growing state-law patchwork to website trackers, vendors and basic cybersecurity.

UNITED STATES · NOVEMBER 11, 2026 · 19 MIN READ

On November 10, 2026, five US states joined a Federal Trade Commission lawsuit accusing telehealth provider Hims & Hers of deceptive privacy practices. The complaint alleges, among other things, that sensitive health information was shared with advertising platforms despite assurances given to consumers. Hims & Hers disputes the allegations, and the case has not yet resulted in a final determination.

For small and medium-sized enterprises, the importance of the case extends well beyond healthcare. It illustrates how privacy enforcement increasingly works in the United States: one data practice can attract federal scrutiny, parallel action by several states, contractual disputes with technology providers and serious reputational damage.

An SME does not need to operate a hospital or hold millions of customer records to encounter similar risks. Website pixels, analytics services, advertising platforms, customer-support software and AI tools can all transmit personal information to third parties.

The first compliance question is therefore no longer simply whether the company has a privacy policy. It is whether the business knows what information actually leaves its systems, where it goes and whether those transfers match what customers, employees and business partners have been told.

The challenge will become more significant in 2027. New state laws will enter into force, existing privacy rights will expand and businesses will face increasingly detailed requirements concerning deletion, sensitive information, automated decision-making, cybersecurity and vendor oversight.

This guide explains how an SME can prepare without treating every state law as identical or assuming that every legal requirement applies to every business.

Privacy and cybersecurity are connected, but they are not the same

Data privacy concerns the rules governing how personal information is collected, used, disclosed, retained and deleted. It also includes the information provided to individuals and the rights available to them.

Cybersecurity concerns the safeguards used to protect systems and information against unauthorized access, theft, alteration, disruption and loss.

A company can perform poorly in one area while appearing competent in the other.

For example, an SME may use encryption, multifactor authentication and secure backups but still collect more customer information than necessary or disclose it to an advertising provider without adequate notice. Conversely, a company may publish a detailed privacy policy while leaving customer accounts protected by weak passwords and excessive employee permissions.

A credible compliance program must address both questions:

  1. Is the company permitted to collect and use the information in this way?
  2. Is the information protected appropriately throughout its lifecycle?

Why the US privacy landscape is difficult for SMEs

The United States does not have one comprehensive federal privacy law equivalent to the European Union’s General Data Protection Regulation.

Instead, businesses must navigate several overlapping layers:

  • federal consumer-protection enforcement;
  • state comprehensive privacy laws;
  • sector-specific federal and state requirements;
  • data-breach notification laws;
  • biometric, health, financial and children’s privacy rules;
  • employment and anti-discrimination requirements;
  • contractual obligations imposed by clients and technology providers.

Applicability may depend on revenue, the number of consumers whose data is processed, whether personal information is sold or shared, the company’s industry and the state in which affected individuals live.

This means that a business cannot determine its obligations by looking only at its registered office.

An Italian company selling products to California residents, for example, may need to assess California law even if it has no physical office in the United States. A US business based in one state may also become subject to the law of another state when it reaches the relevant consumer or revenue thresholds.

The correct starting point is not “Which law applies to companies like ours?” but a more precise set of questions:

  • Whose data do we process?
  • In which states do those individuals live?
  • How much data do we process?
  • What types of information are involved?
  • Do we sell, share or use it for targeted advertising?
  • Do we make significant decisions using automated systems?
  • Are we subject to sector-specific regulation?

California remains the most important starting point

For many SMEs, California is the first jurisdiction to examine because of the size of its economy and the reach of the California Consumer Privacy Act.

The CCPA does not apply automatically to every website accessible from California. A for-profit business must satisfy the applicable statutory conditions and at least one threshold.

As of 2026, the principal thresholds include:

  • annual gross revenue above the inflation-adjusted threshold of $26.625 million;
  • buying, selling or sharing the personal information of at least 100,000 California consumers or households annually;
  • deriving at least 50% of annual revenue from selling or sharing consumers’ personal information.

The analysis must also consider whether the company is controlled by, controls or shares branding with another covered business.

A small company below these thresholds may still face other privacy, security, contractual or sector-specific obligations. Falling outside the CCPA does not mean that personal information can be processed without restrictions.

Rights available under the CCPA

Depending on the circumstances and applicable exceptions, California consumers may have rights to:

  • know what personal information a business collects;
  • request access to specific information;
  • request deletion;
  • correct inaccurate information;
  • opt out of the sale or sharing of personal information;
  • limit certain uses of sensitive personal information;
  • receive equal treatment when exercising their rights.

A covered business needs more than a policy describing these rights. It needs an operational process capable of receiving, verifying, documenting and completing requests within the required timeframe.

California’s expanded deletion right

A particularly important change takes effect on January 1, 2027.

California’s deletion right will extend to personal information obtained from third parties, rather than applying only to data collected directly from the consumer. Online-only businesses will also need to offer an online method, such as a web form, for submitting deletion requests.

This change has practical consequences.

A company may hold information about the same person in its ecommerce platform, CRM, newsletter system, advertising audience, customer-support archive and data-enrichment service. Deleting the record from only one system may not satisfy the request.

Businesses should therefore identify:

  • where personal information originates;
  • which systems contain copies;
  • whether it was obtained from a third party;
  • which records must be retained under an exception;
  • how deletion instructions reach service providers;
  • how deleted information is prevented from being imported again.

California expressly permits the use of a suppression list for this purpose. Such a list should contain only the minimum information necessary to ensure that the deleted record is not unintentionally restored.

California’s new risk, cybersecurity and automated-decision rules

California’s finalized regulations concerning risk assessments, cybersecurity audits and automated decision-making became effective on January 1, 2026, with several obligations introduced through phased deadlines.

Not every business is automatically required to conduct every assessment or audit. Applicability depends on whether the business is covered by the CCPA and meets the conditions established for the specific obligation.

For affected businesses, the regulations introduce requirements concerning:

  • risk assessments for processing presenting significant privacy risks;
  • cybersecurity audits for businesses meeting the relevant criteria;
  • information and rights connected with certain uses of automated decision-making technology;
  • documentation and regulatory submissions.

The first compliance step is therefore a scope assessment, not the immediate purchase of audit or privacy software.

An SME should determine:

  1. whether it is a covered business;
  2. which processing activities fall within the regulations;
  3. whether it makes significant decisions using automated systems;
  4. whether its revenue or processing activities trigger cybersecurity-audit requirements;
  5. which implementation and submission deadlines apply.

Cybersecurity-audit certification deadlines are phased according to revenue, beginning in 2028 for the largest covered businesses and extending to 2030 for covered businesses with less than $50 million in annual revenue. Risk-assessment compliance begins earlier, with attestations and summaries due in 2028 for businesses within scope.

The state-law patchwork will continue expanding

California is not the only jurisdiction that matters.

During 2026, additional states adopted comprehensive privacy legislation scheduled to become effective over the following years. Louisiana and Oklahoma laws are expected to take effect in 2027, followed by Alabama and later Vermont. The precise applicability thresholds, exemptions, cure periods and consumer rights differ among them.

Existing state laws are also being amended.

Delaware, for example, approved changes scheduled for January 1, 2027 that expand protections for sensitive data and narrow parts of the employee-data exclusion. The changes are particularly relevant when employee information is used by AI-assisted recruitment, interview-scoring or workforce-management tools.

The operational lesson is important: compliance cannot be based on a static table created once and forgotten.

At least annually, and before entering a new US market, an SME should review:

  • the states in which it has customers, employees or monitored website users;
  • current applicability thresholds;
  • exemptions relevant to its sector;
  • newly protected categories of sensitive information;
  • opt-in and opt-out requirements;
  • rules governing targeted advertising and profiling;
  • obligations concerning automated decisions;
  • cure periods and enforcement dates.

Website traffic is not the same as a legal threshold

One recurring mistake is to treat the number of website visitors as if it automatically represented the number of consumers whose personal data is processed under a state privacy law.

The two figures may be different.

A website visit may generate identifiers, IP addresses, cookie data, device information and advertising signals. Whether each visitor counts toward a particular statutory threshold depends on the law, the information collected, the company’s role and how the data is processed.

Businesses should not assume that a monthly analytics figure proves either that a law applies or that it does not.

Instead, they should document:

  • the categories of data collected from visitors;
  • whether visitors can be associated with individuals or households;
  • which tracking technologies are active;
  • which third parties receive the information;
  • whether data is used for targeted advertising;
  • how duplicate or repeat visitors are counted;
  • whether the relevant law excludes particular data or activities.

Advertising pixels deserve special attention

The recent FTC action demonstrates why marketing technology cannot be managed only by the marketing team.

A tracking pixel or analytics tag may transmit page visits, identifiers, form events, purchase activity or other information to an external platform. The risks increase when the page concerns health, finance, employment, children or another sensitive subject.

Before activating a new tracking tool, an SME should answer five questions:

  1. What exact information does the tool collect?
  2. Is any of it sensitive or capable of revealing a sensitive characteristic?
  3. Which provider receives it?
  4. What does the provider do with the information?
  5. Does the company’s privacy notice accurately describe the transfer and its purpose?

Consent banners should not be treated as decorative compliance elements. Their settings must correspond to the technologies actually present on the website. A banner that rejects advertising cookies while advertising tags continue firing creates a gap between the interface and the underlying processing.

Periodic technical testing is therefore necessary.

Build one privacy program, not a separate process for every state

Trying to create an entirely separate operational system for each state can become unmanageable for a small business.

A more practical approach is to establish one defensible privacy program and then add state-specific variations where necessary.

Create a data inventory

The inventory should identify:

  • categories of personal information;
  • sources;
  • purposes;
  • systems;
  • responsible departments;
  • recipients and service providers;
  • retention periods;
  • security controls;
  • states or countries connected with the data;
  • whether information is sold, shared or used for targeted advertising.

It does not need to begin as a complex database. A controlled spreadsheet may be sufficient for a smaller organization, provided that it has an owner and is updated when systems or processing activities change.

Map vendors and data flows

List the providers used for:

  • hosting;
  • email marketing;
  • analytics;
  • advertising;
  • payments;
  • customer support;
  • payroll;
  • recruitment;
  • CRM;
  • cloud storage;
  • AI-assisted services.

For each provider, record the type of data received, contractual role, permitted purposes, retention practices, location, subprocessors and available deletion or export functions.

Establish a rights-request procedure

The procedure should define:

  • available submission channels;
  • how requests are recorded;
  • identity-verification methods;
  • applicable deadlines;
  • relevant exceptions;
  • systems that must be searched;
  • who approves the response;
  • how service providers are instructed;
  • what evidence is retained.

The company should test the process. A policy may look complete while failing in practice because no one can delete a contact from a legacy system or identify the owner of an archived mailbox.

Control sensitive information

Sensitive information may include precise location, health details, biometric identifiers, account credentials, racial or ethnic origin and other categories defined by applicable law.

The safest operational rule is not merely to protect sensitive data more carefully, but to avoid collecting it when it is unnecessary.

Set retention rules

“Keep everything in case it becomes useful” is not a sound retention policy.

Each category should have:

  • a business or legal justification;
  • a defined retention period;
  • a deletion or anonymization method;
  • an identified owner;
  • documented exceptions for disputes, tax records, security or other legitimate needs.

Cybersecurity obligations extend beyond privacy laws

Even when a comprehensive state privacy law does not apply, an SME may still face cybersecurity duties under contracts, sector-specific rules, breach-notification statutes or the FTC Act.

The FTC has repeatedly treated unreasonable security practices and misleading privacy or security claims as potential unfair or deceptive practices.

A business should avoid making absolute statements such as:

  • “Your information is completely secure.”
  • “We never share data with third parties.”
  • “All information is encrypted.”
  • “Your data is deleted immediately.”

Such statements are risky if they cannot be demonstrated across every relevant system and provider.

More accurate notices describe the controls and purposes without promising that no incident can occur.

When the FTC Safeguards Rule applies

The FTC Safeguards Rule does not apply to every small business. It applies to covered non-bank financial institutions under the Gramm-Leach-Bliley Act.

Depending on the activity performed, covered organizations may include certain lenders, mortgage brokers, financial advisers, tax-preparation firms, collection agencies and other businesses significantly engaged in financial activities.

Covered institutions generally need a written information-security program proportionate to their size, complexity, activities and the sensitivity of the information handled.

Relevant requirements can include:

  • designating a qualified individual;
  • conducting risk assessments;
  • implementing safeguards;
  • controlling access;
  • using encryption where required;
  • monitoring service providers;
  • testing security controls;
  • preparing an incident-response plan;
  • reporting to the governing body.

Institutions maintaining customer information concerning fewer than 5,000 consumers may qualify for exemptions from certain specified requirements, but not from the entire Safeguards Rule.

Covered financial institutions must also assess whether a security event triggers the FTC notification requirement. Qualifying events involving the unencrypted information of at least 500 consumers generally must be reported to the FTC as soon as possible and no later than 30 days after discovery.

Use the NIST Cybersecurity Framework as an operating model

For SMEs that need a practical structure, the NIST Cybersecurity Framework 2.0 offers a voluntary and adaptable model. NIST also provides a Quick-Start Guide designed for smaller organizations.

The framework is organized around six functions:

FunctionPractical meaning for an SME
GovernAssign responsibility and establish policies and priorities.
IdentifyUnderstand systems, information, vendors and risks.
ProtectApply safeguards such as access controls, training and backups.
DetectIdentify suspicious activity and security events.
RespondContain incidents and coordinate communications.
RecoverRestore operations and learn from the event.

Following the framework does not automatically establish legal compliance. It does, however, help a company replace disconnected security products with a documented risk-management process.

Minimum cybersecurity measures for an SME

The appropriate controls depend on the organization, but a basic program commonly includes:

  • multifactor authentication for email, administration and critical cloud services;
  • unique accounts rather than shared credentials;
  • least-privilege access;
  • prompt removal of access when roles change;
  • secure password management;
  • device and software updates;
  • encrypted backups;
  • restoration testing;
  • endpoint protection;
  • phishing and social-engineering training;
  • vendor-security review;
  • logging and monitoring;
  • an incident-response plan;
  • cyber-insurance review where appropriate.

Backups should be protected from the same credentials and systems they are intended to recover. A backup that ransomware can encrypt with the production environment may not provide meaningful resilience.

Prepare an incident-response plan before a breach

A useful incident plan should identify:

  • who receives the first report;
  • who has authority to isolate systems;
  • which external technical and legal contacts are available;
  • how evidence is preserved;
  • how affected data and individuals are identified;
  • which jurisdictions may require notification;
  • who communicates with customers, employees, insurers and authorities;
  • how operations are restored;
  • how lessons are recorded after the event.

US breach-notification requirements vary by state. Deadlines, definitions of personal information, regulator-notification duties and exceptions are not uniform.

The company should therefore avoid publishing a fixed promise such as “We will notify every affected person within 30 days” unless that commitment has been legally and operationally verified.

European SMEs using US providers

A European SME using American cloud, marketing or AI services must consider both US requirements and the GDPR.

The EU–US Data Privacy Framework can provide a transfer mechanism when the US recipient is actively certified for the relevant type of data. Certification should be checked through the official framework list rather than assumed from a vendor’s marketing page.

If the recipient is not covered, another valid transfer mechanism may be required, such as Standard Contractual Clauses, together with the appropriate assessment and safeguards.

Certification does not eliminate the need to consider:

  • the Article 28 data-processing agreement when the provider acts as a processor;
  • the instructions given to the provider;
  • purpose limitation;
  • security;
  • retention;
  • subprocessors;
  • data-subject rights;
  • onward transfers.

A provider’s certification also does not automatically extend to every subsidiary, product or subprocessor. Each relevant entity and service must be checked.

A 10-step action plan for 2027

1. Identify applicable jurisdictions

Determine where customers, employees and other individuals live. Do not rely solely on the company’s office location.

2. Test legal thresholds

Assess revenue, processing volume, sales or sharing, targeted advertising, sensitive information and sector-specific activities.

3. Inventory personal information

Record what is collected, why, where it is stored and who receives it.

4. Audit website technologies

Test cookies, pixels, analytics, forms, chatbots and advertising integrations. Compare the results with the privacy notice and consent interface.

5. Review vendors

Check contracts, security information, deletion options, subprocessors and transfer mechanisms.

6. Update deletion workflows

Prepare for California’s broader deletion right and ensure that third-party data can be located and removed where required.

7. Review automated decisions

Identify systems used in recruitment, lending, insurance, housing, pricing, employee management or other consequential contexts.

8. Strengthen basic security

Prioritize identity protection, access control, backups, patching and incident response before purchasing advanced tools.

9. Train employees

Training should reflect real roles. Marketing staff need guidance on tracking and advertising tools; HR needs rules for applicant data; IT needs escalation procedures; management needs incident and governance responsibilities.

10. Document decisions

Record why a law was considered applicable or inapplicable, how vendors were assessed and why specific safeguards were selected.

Documentation cannot cure a deficient practice, but it helps demonstrate that decisions were based on an organized process rather than guesswork.

Final checklist

  • ☐ We know which states our customers and employees are located in.
  • ☐ We have assessed the applicability thresholds of relevant privacy laws.
  • ☐ We maintain an inventory of personal information.
  • ☐ We know which vendors receive personal data.
  • ☐ Our website trackers have been technically tested.
  • ☐ Our privacy notice reflects actual data practices.
  • ☐ Our cookie and opt-out controls work as described.
  • ☐ We can receive, verify and document privacy requests.
  • ☐ We can locate data obtained from third parties.
  • ☐ Our deletion process covers relevant service providers.
  • ☐ Sensitive information is collected only when necessary.
  • ☐ Retention periods are defined.
  • ☐ Critical accounts use multifactor authentication.
  • ☐ Access is removed promptly when roles change.
  • ☐ Backups are protected and restoration is tested.
  • ☐ We have an incident-response plan.
  • ☐ We have identified applicable breach-notification rules.
  • ☐ Automated decision tools have been reviewed.
  • ☐ Employees receive role-specific privacy and security training.
  • ☐ European-to-US transfers use an appropriate mechanism.
  • ☐ Compliance decisions and exceptions are documented.

Conclusion

The latest US enforcement actions show that privacy compliance is moving from policy language to operational evidence.

Regulators increasingly examine what data a business actually collects, what its website transmits, how vendors use the information and whether public statements correspond to real practices.

At the same time, state requirements are becoming more detailed. California is expanding deletion rights and phasing in obligations concerning risk assessments, cybersecurity audits and automated decisions. Other states are introducing or amending their own rules, often with different thresholds and exemptions.

An SME does not need to reproduce the compliance department of a multinational company. It does need to understand its data, control its technology providers, maintain reasonable security and respond to individual requests consistently.

The most effective approach is incremental:

  1. map the data;
  2. identify applicable rules;
  3. correct the highest-risk practices;
  4. establish repeatable procedures;
  5. document the work;
  6. review the program as laws and business activities change.

Privacy and cybersecurity are no longer separate legal and technical projects. Together, they form part of ordinary business management.

AI Governance & Responsible AI

A small or medium-sized enterprise may already be using AI without calling it that. This guide compares how the EU AI Act and the US patchwork of federal, state and local rules apply to everyday automation, and sets out a practical compliance plan.

US & EU · NOVEMBER 11, 2026 · 18 MIN READ

A small or medium-sized enterprise may already be using artificial intelligence without describing its systems as “AI.”

A customer service platform may generate suggested replies. Accounting software may extract information from invoices. A recruitment platform may rank applications. A CRM may predict which leads are most likely to convert. Even familiar office applications increasingly include generative AI, automated classification, summarization, and decision-support features.

These tools can reduce repetitive work and help employees process information more efficiently. They can also introduce errors, bias, privacy risks, and regulatory responsibilities.

In 2026, the central question for an SME is therefore no longer simply:

“Can this activity be automated?”

The more useful questions are:

  • What role does the AI system play?
  • Which data does it process?
  • Does it support or influence a decision about a person?
  • Can a human detect and correct an error?
  • Which laws apply in the countries where the business operates?
  • What evidence can the company provide if a customer, employee, auditor, or authority asks how the system is used?

The European Union and the United States are addressing these questions differently. The EU has adopted a single, risk-based framework through the AI Act. The United States continues to rely on a combination of federal laws, state legislation, local rules, sector-specific requirements, and voluntary risk-management frameworks.

For SMEs operating across borders, the challenge is not choosing one model over the other. It is building an internal process capable of identifying and managing both.

1. How SMEs Are Using AI for Business Automation

Most SMEs are not developing their own foundation models. They are using AI functions provided by external software vendors or incorporated into applications they already use.

The most common use cases can be divided into several operational areas.

Customer and vendor communication

Natural language processing and generative AI can be used to:

  • classify incoming requests;
  • suggest email responses;
  • answer frequently asked questions;
  • summarize conversations;
  • translate messages;
  • route tickets to the appropriate department;
  • extract information from quotations or supplier communications.

A chatbot that only provides opening hours presents a different level of risk from one that recommends products, handles complaints, discusses refunds, or accesses customer account information.

The technology may be similar, but the purpose, data, and possible consequences are not.

Back-office automation

AI-assisted tools can support:

  • invoice data extraction;
  • document classification;
  • expense categorization;
  • duplicate detection;
  • contract summarization;
  • meeting transcription;
  • preparation of minutes and action lists;
  • email prioritization;
  • reconciliation of records.

These use cases may appear low risk, but they still require controls. An incorrectly extracted amount, tax code, payment date, or bank account number can produce a real operational or financial error.

Marketing and content production

Businesses may use generative AI to:

  • create first drafts;
  • adapt content to different channels;
  • generate images;
  • summarize market research;
  • analyze reviews;
  • prepare product descriptions;
  • identify recurring customer questions.

The main risks include inaccurate claims, copyright issues, disclosure requirements in specific circumstances, misuse of personal data, and publication of content that no employee has verified.

Recruitment and workforce management

AI systems may:

  • screen résumés;
  • rank candidates;
  • analyze recorded interviews;
  • recommend applicants;
  • monitor employee activity;
  • forecast performance;
  • support promotion or termination decisions.

These applications require particular caution. They can affect employment opportunities, reproduce historical discrimination, disadvantage people with disabilities, or make it difficult for an individual to understand how a decision was reached.

Credit and other consequential decisions

Some systems support decisions concerning:

  • creditworthiness;
  • lending;
  • insurance;
  • housing;
  • education;
  • healthcare;
  • access to essential services.

These are among the most sensitive applications because their output can significantly affect a person’s rights, opportunities, or financial situation.

2. The Operational Limits of AI

AI should not be treated as a digital employee with independent judgment. It is a technical system operating within a defined workflow.

Its limitations must be reflected in the design of that workflow.

Hallucinations and unreliable outputs

Generative AI can produce plausible but inaccurate information. It may invent a contractual clause, refund condition, legal reference, product specification, source, or deadline.

The system’s confidence or fluency is not evidence of accuracy.

If a chatbot gives customers incorrect information, the exact legal consequences will depend on the facts, applicable consumer and contract law, the way the message was presented, and whether the customer relied on it. It is therefore inaccurate to state that every incorrect chatbot response automatically creates a binding obligation.

Nevertheless, inaccurate automated communications can expose the company to:

  • customer complaints;
  • refunds or remediation costs;
  • allegations of misleading conduct;
  • contractual disputes;
  • regulatory scrutiny;
  • reputational damage.

A business should define what the chatbot is permitted to say, when it must escalate a request, and who is responsible for reviewing its performance.

Bias and discrimination

AI systems trained on historical data may reproduce patterns found in that data.

A recruitment system, for example, might learn from previous hiring outcomes that reflect unequal opportunities rather than genuine differences in candidate quality. A performance-monitoring tool could disadvantage workers whose roles, disabilities, schedules, or working methods are not represented accurately in its metrics.

Purchasing a tool from an external vendor does not remove the need for the company using it to assess its effects.

Limited context

AI systems do not understand a customer, employee, or business situation in the same way a responsible person does.

They may fail to recognize:

  • exceptional circumstances;
  • ambiguous instructions;
  • emotional distress;
  • accessibility needs;
  • conflicting evidence;
  • outdated records;
  • commercial sensitivities;
  • legal or ethical implications.

For this reason, a workflow should not be designed around the assumption that the AI will always recognize when it is wrong.

Automation bias

Employees may accept an automated recommendation because it appears objective or technically sophisticated.

Human review is ineffective when the reviewer:

  • lacks access to the underlying information;
  • does not understand the system;
  • has too little time to challenge the output;
  • is expected to approve nearly every recommendation;
  • cannot override the result;
  • fears consequences for disagreeing with the system.

Meaningful human oversight requires authority, competence, time, information, and a clear escalation process.

3. Human Oversight: Important, but Not a Universal Legal Rule

“Human in the loop” is a useful design principle, but it should not be presented as a single global legal requirement applying to every AI use.

The necessary level of oversight depends on the context.

A human review may be unnecessary for a low-impact function such as removing spam, provided that users can recover legitimate messages. It becomes much more important when an AI system influences:

  • recruitment;
  • dismissal;
  • credit;
  • insurance;
  • healthcare;
  • access to essential services;
  • contractual commitments;
  • customer complaints;
  • legal or regulatory communications.

Under the EU AI Act, specific human-oversight obligations apply to high-risk systems. Under the GDPR, individuals also have protections in relation to certain decisions based solely on automated processing that produce legal or similarly significant effects.

In the United States, existing employment, disability, civil-rights, consumer-protection, and privacy laws can apply even when a rule does not explicitly use the term “artificial intelligence.”

The practical principle is therefore:

4. The European Union: One Framework Based on Risk

The EU AI Act establishes an EU-wide regulatory framework based primarily on the purpose and risk of an AI system.

An SME that uses an AI system under its authority will often be considered a deployer. Its obligations depend on the type of system, the way it is used, and whether the company substantially modifies the system or uses it outside its intended purpose.

The AI Act does not impose the same requirements on every AI-enabled application.

Prohibited AI practices

Certain uses are prohibited because they present unacceptable risks. The relevant prohibitions began applying on February 2, 2025, with some later additions introduced through subsequent amendments.

An SME should therefore verify not only whether a tool is commercially available, but also whether the intended use is lawful.

AI literacy

Article 4 of the AI Act has applied since February 2, 2025.

Providers and deployers must take measures to ensure, to their best extent, a sufficient level of AI literacy among employees and other people dealing with AI systems on their behalf. The appropriate level depends on factors such as:

  • the employees’ role;
  • their technical knowledge;
  • the context in which the system is used;
  • the people or groups affected;
  • the risks associated with the application.

This does not mean that every employee needs to become an AI engineer. It does mean that a company should provide training appropriate to the actual use.

For example, employees using generative AI for marketing should understand:

  • that outputs may be inaccurate;
  • how to verify factual claims;
  • which company or customer data must not be entered;
  • when human approval is required;
  • how to report an incident;
  • what internal policy applies.

The company should retain reasonable evidence of the measures taken, such as training materials, attendance records, internal instructions, or role-specific guidance.

Transparency requirements

Article 50 of the AI Act has applied since August 2, 2026.

The requirements are more specific than a general obligation to label everything created with AI.

Examples include:

  • people interacting directly with certain AI systems, such as chatbots, must be informed that they are interacting with AI unless this is already obvious to a reasonably well-informed and observant person;
  • deepfake image, audio, and video content must generally be disclosed as artificially generated or manipulated;
  • AI-generated or manipulated text published to inform the public on matters of public interest may require disclosure, subject to the conditions and exceptions in the legislation;
  • providers have obligations concerning machine-readable marking and detectability of certain AI-generated content.

An ordinary internal draft prepared with AI and reviewed by an employee is not automatically subject to the same disclosure rules as a deepfake or AI-generated public-interest article.

The purpose, publication context, type of content, role of the organization, and applicable exceptions must be assessed.

High-risk AI systems

The AI Act identifies particular high-risk areas, including certain systems used for:

  • recruitment and candidate selection;
  • decisions affecting employment relationships;
  • access to education;
  • creditworthiness;
  • insurance risk and pricing in specified contexts;
  • access to essential private and public services;
  • biometrics;
  • critical infrastructure;
  • law enforcement, migration, and justice.

Following the 2026 amendments to the implementation timetable, the rules for high-risk systems listed in Annex III will apply from December 2, 2027. Requirements for high-risk systems embedded in regulated products under Annex I will apply from August 2, 2028.

Preparation should begin before those dates.

A business considering AI-based recruitment, workforce management, credit assessment, or another sensitive application should determine:

  1. whether the system falls within a high-risk category;
  2. whether the business is acting as deployer, provider, importer, distributor, or more than one of these;
  3. which documentation the vendor supplies;
  4. whether the system can be used according to its instructions;
  5. how human oversight will work;
  6. which data will be entered;
  7. how outputs and incidents will be monitored;
  8. how affected people will be informed where required.

GDPR obligations continue to apply

Compliance with the AI Act does not replace compliance with the GDPR.

When personal data is involved, the business may still need to address:

  • lawful basis;
  • transparency;
  • purpose limitation;
  • data minimization;
  • retention;
  • access controls;
  • processor agreements;
  • international transfers;
  • data-protection impact assessments;
  • security;
  • data-subject rights.

Article 22 GDPR also protects people from certain decisions based solely on automated processing that produce legal or similarly significant effects, subject to defined exceptions and safeguards.

Not every classification, recommendation, or AI-assisted decision falls under Article 22. The degree of human involvement and the real effect on the individual must be evaluated.

Penalties and SMEs

The highest AI Act fine, up to €35 million or 7% of worldwide annual turnover, concerns prohibited AI practices.

Other violations, including specified operator and transparency obligations, fall under different maximum levels.

For SMEs, the AI Act generally applies the lower of the fixed maximum or percentage-based maximum. Authorities must also consider proportionality, the circumstances of the infringement, the interests of SMEs, and their economic viability.

This does not make minor transparency errors harmless, but it is more accurate than suggesting that every SME faces the maximum headline fine.

5. The United States: Federal Laws Plus State and Local Rules

The United States still does not have one general federal AI statute equivalent to the EU AI Act.

Instead, businesses must consider a combination of:

  • federal consumer-protection law;
  • civil-rights and anti-discrimination law;
  • employment law;
  • disability law;
  • financial-services rules;
  • sector-specific requirements;
  • state privacy legislation;
  • state AI laws;
  • local ordinances;
  • contractual obligations.

Federal agencies have repeatedly emphasized that existing laws continue to apply when AI is used. There is no general “AI exemption” from rules prohibiting deception, unfair practices, or discrimination.

The NIST AI Risk Management Framework offers a voluntary method for identifying and managing AI risks. It is useful for organizations of different sizes, but following it does not automatically establish legal compliance.

Colorado

Colorado revised its approach in 2026 through legislation governing automated decision-making technology used to materially influence consequential decisions.

The updated requirements will take effect on January 1, 2027.

They cover developers and deployers in defined circumstances and include protections connected with algorithmic discrimination, notices, data correction, and consequential decisions.

Colorado also adopted separate chatbot-safety requirements, effective January 1, 2027, addressing certain conversational AI services and protections for minors.

New York City

New York City Local Law 144 has been enforced since July 5, 2023.

It restricts the use of covered automated employment decision tools unless:

  • an independent bias audit has been conducted within the previous year;
  • a summary of the audit results has been made publicly available;
  • required notices are provided to affected candidates or employees.

The law does not apply to every digital recruitment tool. Whether it applies depends on the definition of an automated employment decision tool and the extent to which it substantially assists or replaces discretionary decision-making.

Illinois

Illinois’ Artificial Intelligence Video Interview Act applies specifically when employers request recorded video interviews and use AI to analyze those videos for positions based in Illinois.

Among other requirements, an employer must:

  • notify the applicant that AI may be used;
  • explain generally how it works and the characteristics it evaluates;
  • obtain the applicant’s consent;
  • restrict sharing of the video;
  • delete the video within the required period when requested.

It should not be described as a general law covering every use of AI in employment.

California

California’s completed regulations address automated decision-making technology, risk assessments, cybersecurity audits, and CCPA rights.

The regulations became effective on January 1, 2026, with phased compliance deadlines. ADMT-specific obligations concerning significant decisions, including applicable notice, access, and opt-out rights, must generally be implemented by January 1, 2027.

Not every California business or automated process is covered in the same way. CCPA applicability, the nature of the decision, the data involved, and the regulatory definitions must be checked.

Texas

The Texas Responsible Artificial Intelligence Governance Act took effect on January 1, 2026.

It includes:

  • specific disclosure requirements for certain government and healthcare uses;
  • prohibitions involving unlawful discrimination;
  • restrictions on particular biometric and manipulative uses;
  • enforcement provisions;
  • an opportunity to cure alleged violations;
  • possible reliance on recognized risk-management frameworks in specified enforcement circumstances.

It does not create one universal consumer opt-out rule for every private-sector use of automated profiling.

6. EU and US Approaches Compared

Regulatory featureEuropean UnionUnited States
StructureEU-wide AI framework supplemented by GDPR, consumer, employment and sector-specific lawFederal, state, local and sector-specific requirements
Main approachClassifies obligations according to the system’s purpose and level of riskFrequently focuses on harmful outcomes, discrimination, privacy and consumer protection
AI literacyExpress obligation under Article 4 of the AI ActGenerally a risk-management practice rather than one universal statutory obligation
Chatbot transparencySpecific Article 50 requirements apply from August 2, 2026Depends on state, sector, context and representations made to users
Employment AICertain uses are classified as high risk under the AI ActFederal anti-discrimination law plus state and local rules, including NYC and Illinois requirements
Automated decisionsAI Act and, where personal data is involved, GDPR protectionsVaries across privacy, employment, credit, consumer and state AI laws
EnforcementNational authorities and, in specific areas, the European Commission or AI OfficeFederal agencies, state attorneys general, specialized regulators and local authorities
Risk-management referenceMandatory requirements for covered systems, supplemented by standards and guidanceNIST AI RMF is voluntary, although legislation may recognize it in particular contexts
SME treatmentProportionality and lower maximum fine calculations apply in specified casesThresholds, exemptions and penalties vary significantly by law and jurisdiction

7. A Practical AI Compliance Plan for SMEs

An SME does not need to begin with a complex compliance platform. It needs a reliable internal process.

Step 1: Create an AI inventory

Record all AI systems used in the business, including features embedded in existing software.

For each system, document:

  • tool and vendor;
  • business purpose;
  • department;
  • responsible owner;
  • users;
  • input data;
  • output;
  • people affected;
  • integrations;
  • level of human review;
  • retention;
  • countries involved;
  • contract and privacy documents;
  • date of the most recent review.

The inventory should also cover tools adopted informally by employees. These are often referred to as “shadow AI.”

Step 2: Classify the use, not just the product

The same AI service may present different risks in different workflows.

A language model used to improve grammar in an internal memo is not equivalent to the same model being used to rank job applicants or advise customers about medical products.

Assess:

  • the actual purpose;
  • possible consequences;
  • personal and confidential data;
  • whether the output affects a person;
  • whether the decision can be reversed;
  • whether a human can identify an error;
  • applicable countries and sectors.

Step 3: Review the vendor

Before approving a tool, request information about:

  • intended use;
  • limitations;
  • training and customer data;
  • retention;
  • subprocessors;
  • international transfers;
  • security controls;
  • incident notification;
  • logging;
  • output monitoring;
  • export and deletion;
  • contractual responsibilities;
  • changes to models or functionality.

Marketing claims should not substitute for technical and contractual evidence.

Step 4: Assign meaningful human oversight

For sensitive workflows, identify:

  • who reviews the output;
  • what evidence the reviewer sees;
  • which errors the reviewer must look for;
  • when the output must be rejected;
  • who can override it;
  • when the case must be escalated;
  • how the decision is documented.

Adding an approval button is not sufficient if the reviewer cannot meaningfully challenge the recommendation.

Step 5: Train employees

Training should reflect the employee’s role and the systems used.

At a minimum, relevant employees should understand:

  • approved and prohibited uses;
  • hallucinations;
  • confidentiality;
  • personal data;
  • verification procedures;
  • bias and discrimination risks;
  • disclosure requirements;
  • incident reporting;
  • human-review rules.

Step 6: Introduce transparency at the correct point

A company should not pretend that a chatbot is a human operator.

A clear notice such as “You are interacting with an AI assistant” may be appropriate for a customer-facing system, together with an accessible way to contact a person where needed.

For AI-generated content, the business should determine whether disclosure is legally required, contextually appropriate, or unnecessary after genuine editorial review. A single universal label for every AI-assisted sentence is not a reliable compliance strategy.

Step 7: Test normal cases and failures

Testing should include:

  • correct inputs;
  • incomplete information;
  • unusual cases;
  • biased or inconsistent outputs;
  • unavailable services;
  • incorrect permissions;
  • duplicate actions;
  • escalation failures;
  • attempted prompt manipulation;
  • unauthorized access;
  • recovery procedures.

A successful demonstration is not evidence that the system will remain reliable in production.

Step 8: Monitor and review

AI systems and embedded software features change.

The company should periodically review:

  • output accuracy;
  • complaints;
  • overrides;
  • incidents;
  • discriminatory effects;
  • vendor updates;
  • access rights;
  • integrations;
  • legal requirements;
  • continued business need.

8. Questions to Ask Before Automating a Process

Before deployment, the business should be able to answer:

  • What problem are we solving?
  • Is AI necessary, or would a rules-based automation be more reliable?
  • What happens when the system is wrong?
  • Which data enters the system?
  • Are personal or confidential data involved?
  • Is the output used to make or support a consequential decision?
  • Can an affected person request an explanation, correction, or human review?
  • Who approves the output?
  • Can the system be stopped safely?
  • Is there a manual alternative?
  • What evidence will we retain?
  • Which EU, US federal, state, local, or sector-specific rules apply?
  • Has the vendor provided enough information to support our assessment?
  • When will the system be reviewed again?

Conclusion

AI remains one of the most accessible automation technologies available to SMEs. It can accelerate document processing, improve access to internal information, support employees, and reduce repetitive work.

But adoption should not begin with the software.

It should begin with the process, the data, the people affected, and the consequences of an error.

The European Union has created a structured, risk-based framework in which an SME may have specific responsibilities as a deployer. The United States relies on a more fragmented combination of federal enforcement, state legislation, local rules, and sector-specific protections.

Despite these differences, the operational lesson is similar on both sides of the Atlantic:

For an SME, responsible AI governance does not require unnecessary bureaucracy. It requires proportionate controls that become stronger as the potential impact of the system increases.

That is what allows automation to create sustainable efficiency without turning an avoidable technical error into a legal, financial, or reputational problem.

Primary Sources

  1. European Commission, AI Act overview and implementation timeline: digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  2. European Commission, Navigating the AI Act: digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act
  3. European Commission, Article 50 transparency obligations: digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
  4. EUR-Lex, Consolidated EU AI Act: eur-lex.europa.eu
  5. EUR-Lex, General Data Protection Regulation, including Article 22: eur-lex.europa.eu
  6. Colorado Attorney General, Automated decision-making and chatbot legislation: coag.gov/ai
  7. New York City DCWP, Automated Employment Decision Tools FAQ: nyc.gov
  8. Illinois General Assembly, Artificial Intelligence Video Interview Act: ilga.gov
  9. California Privacy Protection Agency, ADMT and CCPA regulations: cppa.ca.gov
  10. Texas Legislature, Texas Responsible Artificial Intelligence Governance Act: capitol.texas.gov
  11. NIST, AI Risk Management Framework: nist.gov
  12. Federal Trade Commission, Enforcement involving deceptive AI claims: ftc.gov
  13. US Equal Employment Opportunity Commission, Artificial Intelligence and the ADA: eeoc.gov

AI for Small Businesses

Leggi in italiano →

AI is no longer reserved for large multinationals. A small or medium-sized business can use it to speed up repetitive work, organize information and draft documents, as long as it starts from a concrete problem, protects its data, trains its people and keeps humans in control.

EUROPEAN UNION · NOVEMBER 11, 2026 · 14 MIN READ

Artificial intelligence is no longer a technology reserved for large multinationals or tech companies. Even a small or medium-sized business can now use AI-based tools to speed up repetitive tasks, organize information, prepare draft documents and support administrative, sales and operational processes.

This does not mean that AI automatically solves problems of organization, management or strategy. No tool can make up for confusing procedures, messy data, unclear goals or undefined responsibilities.

For an SME, artificial intelligence can create value when it is introduced with a method: starting from a concrete problem, protecting data, training people and keeping human control over important decisions.

Research and verified sources

SourceWhat it supportsLimits and clarifications
ISTAT — Enterprises and ICT, 2025Among Italian businesses with at least 10 employees, AI adoption rises from 5.0% in 2023 to 8.2% in 2024 and 16.4% in 2025.The survey does not include micro-enterprises with fewer than 10 employees.
Eurostat — Use of artificial intelligence in enterprisesIn 2025, 20% of EU enterprises with at least 10 employees use AI technologies. Small enterprises record a rate of 17%, against 55% for large ones.Image, video and audio generation was recorded as a separate category only from 2025.
OECD — data and reports 2025–2026Across the OECD countries with available data, AI adoption by businesses reaches 20.2% in 2025, compared with 14.2% in 2024 and 8.7% in 2023.Survey methodologies may vary between countries.
OECD — AI adoption by SMEsLack of skills is a significant barrier: in a survey across four G7 countries, 50% of SMEs said their employees do not have sufficient skills to use generative AI.The figure does not automatically represent all European or Italian SMEs.
Italian Data Protection Authority (Garante) — decision no. 342 of 14 May 2026It recalls the limits on using AI to analyze work chats on Slack or Teams in order to infer employees’ stress or emotional state.The case concerns a specific configuration, but it recalls principles that are relevant for businesses.
AI Act — EU Regulation 2024/1689Introduces obligations related to AI literacy, human oversight, prohibited practices and high-risk systems.Obligations vary depending on the type of system, the company’s role and the context of use.
MIMIT — training measures for SMEs in Southern ItalyThe research points to a €50 million measure for training the staff of SMEs in Southern Italy.Calls, requirements, rankings and time windows must be checked at the time of application.

Data to read with caution

Not all statistics on AI adoption describe the real situation of SMEs. Some studies aggregate the figure by number of employees, which gives large companies much more weight in the final result.

For this reason, it is not correct to say that “over 70% of companies use AI” without checking how the figure was calculated, which businesses it includes and which year it refers to.

Likewise, it is not correct to promise guaranteed savings, fixed productivity percentages or automatic replacement of staff. The effectiveness of AI depends on the quality of the data, the process chosen, the training of the team and the company’s ability to check the results.

1. What is artificial intelligence

Artificial intelligence is a set of computing technologies that allows software and digital systems to process complex information, recognize patterns, classify data, make predictions or generate content.

Unlike traditional programs, which follow rigid instructions defined in advance, many AI systems analyze data and examples to produce a result consistent with a goal.

In practice, AI can:

  • recognize recurring patterns in large amounts of data;
  • process text, images, sounds or numerical data;
  • classify documents and requests;
  • suggest forecasts based on historical data;
  • detect anomalies;
  • optimize time, resources or routes;
  • generate text, images, summaries, emails, presentations or code.

One point needs to be made clear, however: AI has no consciousness, intentions, personal experience, common sense or legal responsibility. It does not “understand” the world the way a person does. It produces results through mathematical, statistical and probabilistic models.

Many people already use AI systems every day, even without calling them that:

  • spam filters in email;
  • automatic translators;
  • purchase suggestions in e-commerce stores;
  • real-time maps and navigation;
  • bank anti-fraud systems;
  • automatic recognition of text and documents.

Key concepts

ConceptSimple explanation
AlgorithmA series of instructions and calculations used by software to process information and solve a problem.
AI modelA structure trained to recognize patterns, classify data or generate answers.
TrainingThe process in which the model analyzes large numbers of examples to learn statistical relationships.
Machine learningAn approach that lets software improve its performance by analyzing data and examples, without receiving rigid instructions for every single case.
Deep learningAn advanced form of machine learning based on multi-layer artificial neural networks, especially useful for images, sounds and language.
PromptThe instruction, question or request that the user sends to an AI tool.
OutputThe result produced by the AI: text, image, summary, table, code or classification.

2. The main types of artificial intelligence

Generative AI, the kind used to write text or create images, is only one part of a broader ecosystem. An SME may come across different AI technologies, often built into the software it already uses.

Type of AIWhat it doesExamples of use in an SME
Generative AIProduces new outputs such as text, images, emails, summaries, presentations or code from natural-language instructions.First drafts of emails, posts, product pages, minutes and newsletters.
Predictive AIAnalyzes historical data to estimate future events or trends.Sales forecasts, stock estimates, seasonal demand, risk of late payments.
Classification and recognition AISorts and assigns categories to documents, images, invoices or requests.Email routing, ticket classification, document filing, reading supplier invoices.
Conversational AIInteracts with users through text or voice.Chatbots for frequently asked questions from customers or staff.
Recommendation AISuggests products, content or actions based on available behavior and data.Cross-selling in e-commerce and prioritizing opportunities in the CRM.
Computer visionAnalyzes images and video to recognize objects, writing or anomalies.Quality control in factories, shelf monitoring, checking safety equipment.
Anomaly detection AIIdentifies unusual behavior or deviations from historical data.Flagging suspicious transactions, cyber intrusions or equipment failures.
Optimization AICompares many variables to suggest the best combination of time, resources or routes.Planning shifts, deliveries, routes and production sequences.

Example: a trading company

A company that distributes seasonal products can use a predictive system to analyze sales history and estimate the volumes to buy. It can then use a recommendation engine on its B2B e-commerce site to suggest complementary products to customers. Finally, the marketing department can use a generative assistant to prepare a first draft of an email campaign for different customer segments.

3. How generative AI works

Tools such as ChatGPT, Claude, Gemini, Microsoft Copilot and Perplexity have different architectures, features, levels of integration and settings. However, the general principle is similar: they receive a request, analyze the context available and generate an output that is probabilistically coherent.

These systems are trained on large amounts of text and multimedia data. During training, they learn statistical relationships between words, images, concepts and contexts.

When a person enters a prompt, the model:

  1. analyzes the request received and the information available in the context;
  2. calculates, element by element, which response is most coherent from a statistical point of view;
  3. produces an output that is grammatically and semantically plausible.

This means a text can look professional, fluent and convincing even when it contains errors.

The limits of generative AI

  • It does not automatically verify facts. It can produce a plausible answer without checking whether a figure, a rule or a quotation is correct.
  • Its information may be out of date. Some tools can access up-to-date sources and others cannot. This depends on the version, the configuration and the features turned on.
  • It depends on the quality of the input. A vague prompt with no context tends to produce generic or unhelpful results.
  • It can be convincingly wrong. An error can be presented with a confident tone, formal language and an apparently correct structure.

Generative AI can be compared to a very fast collaborator who can prepare a first draft in a few seconds. However, it does not automatically know your company, does not always check sources and cannot take responsibility for the result.

ConceptSimple explanation
AI hallucinationFalse information, a non-existent quotation, an invented number or a wrong rule presented as plausible.
ContextThe information, constraints, data and instructions that help the AI produce a more useful output.
Human reviewThe check carried out by a competent person before the result is used or sent to customers, suppliers or colleagues.

4. What AI can be used for in an SME

AI can support activities across different departments, especially where there are repetitive tasks, long documents, recurring requests or large amounts of information to organize.

Business areaExamples of use
AdministrationExtracting data from invoices, classifying documents, preparing draft entries.
SalesDraft quotes, summaries of tender specifications, preliminary analysis of calls for tender and requests.
Customer careFAQ chatbots, message routing, first replies on orders, shipping and opening hours.
Human resourcesOnboarding checklists, draft job descriptions, internal guides and procedures.
MarketingProduct pages, posts, newsletters, copy variations and editorial plans.
OrganizationMeeting summaries, writing procedures, operational planning and document analysis.

Administration and documents

AI can help reduce the time spent manually entering data from invoices, receipts and unstructured documents. Recognition and classification tools can extract items such as VAT number, taxable amount, due dates and withholding taxes, preparing a working base to be checked in the management software.

Email and internal communication

Writing assistants can prepare first replies to recurring questions from customers and suppliers. They can also summarize long email conversations, identifying decisions, open tasks and critical points.

Marketing and content

Generative AI can help overcome the blank page: it can produce first drafts of e-commerce product pages, newsletters, posts and message variations. The result must still be checked against tone of voice, the real characteristics of the product, copyright and brand positioning.

Sales and quotes

A language model can support the preliminary reading of tender specifications, requests for proposal and calls for tender, extract requirements and deadlines and organize the information in a table. It can also help produce a draft of the descriptive sections of a quote, to be validated before sending.

Customer care

A conversational assistant can handle standard questions about opening hours, policies, order status or product availability. Complaints, complex problems, negotiations and sensitive situations must instead stay in the hands of people.

Research and analysis

AI can read long documents, industry reports, studies or regulations and help extract the main concepts, differences between versions and comparison tables. This does not replace reading the original sources when the decision is legal, tax-related, technical or strategic.

Training and internal procedures

Starting from notes, transcribed voice memos or fragmented documents, AI can help create drafts of manuals, checklists, operating procedures and onboarding materials.

Work organization

Optimization systems can help plan shifts, deliveries, production activities and calendars based on operational variables and available data.

Three concrete examples

Professional firm. A labor consulting firm uses an AI assistant to compare an earlier and a new version of a collective agreement. The tool prepares a table of possible changes to minimum pay, overtime and seniority increases. The consultant checks every figure against the official texts before sending a circular to clients.

Made-to-measure furniture e-commerce. A company adds a conversational assistant to its website to answer frequent questions about shipping, materials, maintenance and indicative timing. Complaints about damaged products are handed over to human customer care.

Manufacturing company. A metalworking company uses cameras and computer vision to spot micro-cracks, scratches or defects on metal sheets. The system flags suspect pieces and the operator carries out the final check.

What AI can do and what must stay under human control

What AI can doWhat must stay under human control
Create the first draft of a sales email.Check tone, technical data, recipients and the correctness of the message.
Summarize a report or a manual.Check for omissions, errors and critical passages.
Generate copy variations for a campaign.Choose the version consistent with the brand and the commercial goal.
Analyze historical data and suggest trends.Assess data quality and make strategic decisions.
Answer standardized FAQs.Handle complaints, negotiations and situations that require empathy.
Prepare a draft procedure.Validate safety, feasibility and operational responsibilities.

5. What AI cannot do in place of the company

AI can assist, speed up and organize work. It cannot replace management’s responsibility, the experience of people who know customers and processes, or the judgment of qualified professionals.

There are three main limits.

It does not know context that is not provided

AI does not automatically know that a customer is strategic, that a negotiation has a delicate history or that a team is going through a difficult situation. This information has to be explained, but it is not always appropriate to enter it into an external tool.

It has no legal responsibility

If a quote contains an error, if a tax document is wrong, if a text infringes third-party rights or if personal data is mishandled, responsibility remains with the company and the people who used the result.

It does not replace qualified professionals

AI does not replace accountants, lawyers, labor consultants, doctors, DPOs, health and safety officers (RSPP) or other professionals who are called upon to give qualified opinions.

6. Risks to know before you start

Personal data and GDPR

When an SME uses AI tools with data of customers, suppliers or employees, it must know which data is entered, in which environment it is processed, which plan or license is used and which contractual guarantees are available.

Uploading personal data, contracts, tax information, health data, CVs, price lists or trade secrets to personal accounts or unapproved tools can expose the company to privacy and security risks and to a loss of control over its data.

Shadow AI

Shadow AI is the unauthorized or uncontrolled use of AI tools by employees, often through personal accounts.

It can happen when a person uploads a confidential contract, a customer list, a file with prices or proprietary code to an external tool in order to work faster. The risk is not only technological: the company can lose control over data, access, documents and responsibility.

Copyright, intellectual property and hallucinations

An AI output should not be used without review. A text or an image may turn out too similar to protected material, while an answer may contain non-existent rules, invented quotations or wrong data.

Connectors, plug-ins and excessive permissions

Connecting an AI tool to email, cloud storage, a CRM or company documents can be useful, but it requires prior assessment. Overly broad permissions or uncontrolled integrations can create risks of improper access, data exfiltration or cyberattacks.

AI Act, work and sensitive areas

The AI Act prohibits the use of systems to infer people’s emotions in the workplace, except for limited exceptions for medical or safety reasons.

The Garante’s decision of 14 May 2026 reprimanded a start-up that had developed a plug-in to analyze communications on Slack and Teams and infer workers’ stress indicators.

AI tools used for staff selection, CV screening, performance evaluation, worker monitoring or creditworthiness analysis can also fall into high-risk areas. In these cases much greater caution, specific assessments and effective human oversight are required.

7. How to start with AI in your company: 5 concrete steps

1. Choose one operational problem

The starting point is not buying a platform. It is identifying a repetitive, well-defined, low-risk activity.

Examples:

  • turning a recording into minutes;
  • preparing the first draft of an email reply;
  • summarizing a meeting;
  • sorting requests received through a form;
  • translating an internal manual;
  • creating an operational checklist.

2. Start with simple, low-risk activities

There is no need to begin with complex automations, proprietary projects or expensive platforms. An SME can test tools already included in its work suite, with a small group of users and clearly defined tasks.

3. Define the data that cannot be entered

Before authorizing the use of AI, the company should create a simple but clear internal policy.

The policy must state which information cannot be entered into external tools, for example:

  • personal data that is not necessary;
  • health data;
  • confidential contracts;
  • patents;
  • production costs;
  • credentials;
  • non-public price lists;
  • sales strategies;
  • trade secrets.

4. Train the team on limits, prompts and checks

The AI Act requires providers and deployers to take adequate measures to ensure a sufficient level of AI literacy among the people who work with these tools.

In an SME, this means teaching people to:

  • write clear requests;
  • recognize possible errors;
  • verify sources and data;
  • understand which information can be entered;
  • tell a draft from a final document;
  • know when to turn to a manager, the DPO, the consultant or IT.

The research also mentions public measures dedicated to SME training. However, calls and incentives have specific time windows and requirements: they must be checked before planning investments.

5. Measure time, quality and errors

A pilot should have a defined period, for example 90 days.

It is useful to measure:

  • time spent before and after using AI;
  • quality of the drafts produced;
  • number of corrections needed;
  • errors or hallucinations detected;
  • level of adoption by the team;
  • activities for which AI did not produce value.

Only after checking the result of the first use case does it make sense to extend AI to other areas.

8. How to choose AI tools without looking for “the best one”

There is no single best tool for every company. ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity and similar solutions have different features, privacy levels, integrations, prices and ways of working.

The choice should depend on the business context.

CriterionQuestion to ask
CompatibilityDoes the tool integrate with the email, documents, cloud or suite the company already uses?
Data securityAre there business plans, contractual guarantees and suitable settings?
Task to performDo you need to write, analyze documents, do research, organize meetings or optimize processes?
BudgetIs the cost of licenses proportionate to the real value of the use case?
Internal skillsCan the team use the tool and check the outputs?
GovernanceAre authorized users, permitted data, roles and review procedures clear?

Conclusion

Artificial intelligence is not a goal to be reached at all costs and it is not a shortcut for replacing people.

For an SME, the real value of AI lies in reducing the time spent on repetitive tasks, preparing better drafts, organizing information and freeing up space for activities that require experience, customer relationships, judgment and decision-making.

The result depends less on how complex the chosen tool is and much more on the quality of the organization: clear goals, carefully managed data, trained people and constant human control.

Checklist: first steps with AI in your company

  1. Select a repetitive, standardized activity to test.
  2. Appoint an internal person responsible for the pilot.
  3. Map the AI tools already being used informally by colleagues.
  4. Create an AI policy with permitted and prohibited data.
  5. Use approved company accounts, avoiding personal accounts for work materials.
  6. Check the processing of personal data with the DPO, a privacy consultant or legal counsel when necessary.
  7. Exclude prohibited systems, such as analysis of employees’ stress or emotions.
  8. Train the people involved on prompts, limits, data and output checks.
  9. Apply mandatory human review to numbers, rules, technical data and documents intended for external use.
  10. Evaluate the outcome of the test after 90 days before extending the use of AI.

Disclaimer

This article is for information purposes only and does not constitute legal, privacy, tax, technical or organizational advice. Before using AI tools with personal data, confidential documents or processes that affect people, you should assess the specific case with competent professionals.

Main sources

  • ISTAT — Enterprises and ICT, 2025
  • Eurostat — 20% of EU enterprises use AI technologies
  • OECD — AI use by individuals surges across the OECD as adoption by firms continues to expand
  • OECD — AI adoption by small and medium-sized enterprises
  • Italian Data Protection Authority (Garante) — decision no. 342 of 14 May 2026
  • EU Regulation 2024/1689 — AI Act
  • MIMIT — training and digitalization measures for SMEs