AI data confidentiality: 2026 best practices for SMEs
To keep control of your data when using AI, three decisions cover most of the risk: define what must never enter an AI tool, use only business plans whose contract rules out training on your data, and write short rules that anyone can apply tomorrow. The rest (impact assessment, records, supplier audits) follows, depending on how risky your uses are.
This guide is for managers and SME owners who see their teams using ChatGPT, Claude, Copilot or Gemini without always knowing where the data goes. It complements our guide to AI in business for SMEs and draws on the recommendations published by the CNIL, the French data protection authority.
Why confidentiality is now a leadership issue
Because every prompt is a data transfer to a third-party provider, and most leaks come from ordinary use rather than attacks. A salesperson pastes a customer file into a chatbot to draft an email, an HR officer summarizes an annual review, an accountant submits an invoice export. None of these acts is malicious, and each can amount to processing personal data under the GDPR.
CNIL figures give a sense of the general risk: 5,629 data breach notifications in 2024, 93% of them involving a loss of confidentiality, and 15% stemming from incidents at just 8 subcontractors (CNIL 2025 cybersecurity brochure). These figures are not specific to AI, but they make a useful point: a large share of incidents goes through service providers. An AI tool is one more provider in your chain.
On penalties, the GDPR provides administrative fines of up to 20 million euros or 4% of worldwide annual turnover, whichever is higher. For an SME, the real risk is rarely that ceiling. It is more often a customer losing trust, a large account cancelling a contract because it demands guarantees, or the time spent handling an incident. You can also read our article on AI Act sanctions for the AI-specific regulation.
Classify your data before choosing your tools
Start with three levels: free data, internal data, forbidden data. The classification fits on one page and avoids case-by-case debates.
- Free data: content that is already public (website, brochures, published job ads), generic text with no names or confidential figures.
- Internal data: procedures, anonymized minutes, tables without identifiable people. Usable only in a validated business tool.
- Data forbidden in any non-validated tool: identifying personal data of customers, employees or service users, health data, banking data, trade secrets, contracts under negotiation, sensitive source code, logins and passwords.
For a nonprofit, the "forbidden" category in practice includes beneficiary and donor files. For a firm bound by professional secrecy, it is broader still and deserves separate thought.
One limit to keep in mind: a classification only protects if people know it. A rule nobody has read is not a rule.
Choose tools whose contract protects your data
A free consumer tool does not offer the same guarantees as a business plan: check in writing what happens to your prompts. The points to verify are the same for every provider, and they appear in the terms of use or the data processing agreement (DPA).
- Training: is your data used to train models? Business plans generally exclude it by default; personal accounts may enable it by default or through a setting.
- Location and transfers: where is the data hosted and processed, and what mechanisms cover any transfer outside the European Union?
- Retention: how long are conversations kept, and can you delete them?
- Sub-processors: who else accesses the data?
- Security: encryption, logging, access management, strong authentication.
- Reversibility: can you export and delete your data when the contract ends?
Do not rely on a salesperson's verbal answer or a marketing page. Ask for the DPA, read it, file it.
Understand what the CNIL says about AI and the GDPR
The GDPR applies as soon as personal data is processed, whether to train an AI system or simply to use one. On 22 July 2025 the CNIL finalized its recommendations on developing AI systems and published a series of practical sheets covering, among other things, purpose, legal basis, legitimate interest, informing individuals, exercising rights, data annotation and security.
One point matters for AI users: according to the CNIL, drawing on the European Data Protection Board opinion of December 2024, the GDPR applies in many cases to models trained on personal data, because a model can memorize part of that data. In practice for you: the better you control what goes into the tool, the less you depend on what the model might reproduce.
Note the boundary: these sheets first target those who design and train systems. If you simply use a market tool, you remain the data controller for what you enter, and your provider is generally your processor. Compare with our analysis of AI Act vs GDPR to tell the two texts apart.
Know when an impact assessment (DPIA) is required
A DPIA is mandatory when processing presents a high risk to individuals, and AI reaches that threshold faster than a classic tool. On 11 October 2018 the CNIL adopted a list of fourteen types of processing for which it considers an assessment mandatory, and states that the list is not exhaustive: processing not on the list may also require one, based on nine criteria from European guidelines.
Warning signs for an SME: using AI to assess or rank people (candidates, employees, customers), processing sensitive data or data at large scale, monitoring employee activity. In these cases, do not start before documenting risks and measures. For a simple writing use with no personal data, a DPIA is generally not needed.
Write short rules people can apply
A good AI charter fits on one or two pages and answers five questions: which tools, which data, which uses, who validates, what to do when in doubt. If your document is longer, it will not be read.
Here is the outline we recommend to the SMEs we train:
- The list of authorized tools, with the business account to use (and a ban on using a personal account for work).
- The three data categories and what is allowed for each.
- The verification rule: any output used externally is reviewed by a human.
- A contact person for questions, even if it is the manager at first.
- The procedure when something goes wrong: report immediately, with no sanction for the person who reports.
That last rule matters more than people think. A team afraid of being punished hides its mistakes, and you learn about the leak too late. For the wider framework, see how to write an AI policy for your company and our AI policy template.
Deal with shadow AI
Banning without offering an alternative pushes your teams toward personal tools you can no longer see. That is the mechanism behind shadow AI: the usage already exists, it just happens outside any framework.
What works: first, ask your teams, without judgment, what they already use. Then provide a validated business tool for the most common uses. Finally, block only what carries real risk, rather than everything. A limit worth acknowledging: a full technical block is hard to set up in an SME without a dedicated IT team, and it can be bypassed from a personal phone. Training remains the most reliable lever.
Train teams: an obligation as much as a safeguard
The European AI regulation (Regulation EU 2024/1689) requires any organization deploying AI to ensure a sufficient level of AI literacy among its staff; that is Article 4. Beyond the obligation, it is also your best defense: most confidentiality incidents come from a good-faith act by someone who did not know.
For the details of the obligation, read our article on Article 4 of the AI Act and our guide to AI training for business. A half-day awareness session is often enough to cover the basic rules; the length of AI training then depends on the jobs involved.
Prepare your response to a leak
If personal data has been exposed, notification to the CNIL must happen within 72 hours when the breach presents a risk to individuals. Do not discover this procedure on the day of the incident.
Prepare a one-page sheet in advance: who to alert internally, how to document the incident (date, data involved, number of people, tool involved), who decides whether to notify, who contacts the provider to have the data deleted. Keep an internal incident log, even for minor events. If you have no data protection officer, designate at least one person who knows the procedure. For an overall view of your risks, an AI audit for SMEs can map real usage before an incident does it for you.
Frequently asked questions
Can I put customer data into ChatGPT or another AI assistant?
Not in a consumer or personal account. With a business plan whose contract excludes training on your data, and subject to a legal basis and proper information of the individuals, it is possible for defined uses. When in doubt, anonymize first.
Is anonymization enough to protect the data?
It greatly reduces the risk, provided it is real. Removing the name is not enough if the job title, city and date let someone identify the person. Test it by asking whether a colleague would recognize the person when reading.
Do I need a data protection officer (DPO) to use AI?
Not systematically. Appointment is mandatory in certain cases set by the GDPR (public authorities, regular large-scale monitoring, large-scale sensitive data). For an SME outside those cases, an internal contact person is often enough.
Does an AI service hosted in Europe solve the issue?
It simplifies the question of transfers outside the European Union, but it settles neither training, retention nor access security. Hosting is one criterion among others.
What should I do if an employee pasted sensitive data into a personal tool?
Document what was entered, ask the provider to delete the conversation and data, assess the risk to individuals, and decide with your contact person whether a CNIL notification is needed. Do not punish straight away: you want future reports to arrive quickly.
Does this approach also apply to nonprofits?
Yes, with particular care for beneficiary, donor and volunteer files, which are often sensitive by nature.
How GrowthPerf helps
GrowthPerf is a Qualiopi-certified training provider specializing in AI, no-code and automation for SMEs and nonprofits. We work on three points: a short awareness session for your teams, centered on confidentiality rules and exercises on your own cases; drafting a usage charter with you, adapted to your tools; and an audit of existing usage to spot shadow AI.
To leave with something concrete, request our free charter: it follows the outline of this article, ready to adapt to your organization. We are available to discuss it on growthperf.com.
To continue reading, see the complete guide to AI in business for SMEs and our article on AI governance in SMEs.