AI and business data privacy: which AI tools can you trust with company data?

To use AI without exposing your company’s data, pick a tool whose plan excludes training on your content and states clearly how long data is kept and where it is stored, then limit what you send to it. A business plan backed by a contract, the right settings and a few simple internal rules protect you far better than a personal subscription chosen at random.

This article provides general information, current as of October 22, 2026. It does not replace advice from your data protection officer, privacy lead or a lawyer for a specific situation.

Why privacy is a real issue with AI

When an employee pastes a contract, a customer spreadsheet or meeting notes into an AI assistant, that text leaves the company and lands on the vendor’s servers. Four questions decide how risky that is.

1. Training. Can your conversations become training data for future models? If so, confidential information could, in theory, shape answers given to other users.
2. Retention. How long are conversations kept, and can people at the vendor review them?
3. Location. In which country is the data stored and processed? That affects which laws apply and whether data is transferred abroad.
4. Access. Who inside your company and at the vendor can see the data? A personal account sits outside any employer control.

The most common risk is not a dramatic breach. It is “shadow AI”: staff using free personal accounts because no approved tool is available.

What the rules say: GDPR, the ICO and the FTC

Whenever information about people is involved (customers, job applicants, employees, patients), privacy law applies. In Europe, the GDPR has applied since May 25, 2018, including to American and Chinese vendors serving people located in the EU. The United Kingdom has its own UK GDPR, overseen by the Information Commissioner’s Office (ICO). In the United States there is no single federal privacy law of this kind; the Federal Trade Commission (FTC) enforces against unfair or deceptive data practices, and several states have their own privacy laws.

In practice, the organization that decides to use an AI tool remains responsible for how it uses the data. Core principles to remember: a clear purpose, the minimum data needed, security, transparency toward the people concerned, and a data processing agreement when a vendor processes personal data on your behalf. In the EU, the AI Act adds to the GDPR rather than replacing it.

Personal plans versus business plans: the real dividing line

The key line does not run between free and paid, but between individual plans and team or enterprise plans. Here is what vendors state, according to the site’s profiles as of October 22, 2026.

Assistant	Individual plans	Business plans
ChatGPT	Training on by default, opt-out under “Improve the model for everyone”; temporary chats kept up to 30 days	Business: no training by default; data residency in Europe on Enterprise and Edu
Claude	Training can be turned off; incognito chats never used for training	Team and Enterprise: no training by default
Gemini	With “Keep Activity” off, chats kept 72 hours and not used for training	Not detailed in the profile
Le Chat (Mistral Vibe)	Data hosted in the EU by default; training on for Free, Pro and Student, can be turned off	Team and Enterprise: separate business terms
Lumo (Proton)	No training, “zero-access” encryption, servers in Europe	Not detailed in the profile
DeepSeek	Data stored in China, per its privacy policy	No paid plan for the assistant

Two more points matter for technical teams. According to GitHub, since April 24, 2026, interaction data from GitHub Copilot Free, Pro and Pro+ accounts is used for training unless users opt out in settings. Notion says it does not train models on customer data without explicit consent, with zero data retention at model providers on its Enterprise plan.

The profiles of Mistral’s Le Chat and Proton’s Lumo detail two Europe-based options. To compare features across assistants, try the AI assistant comparison tool.

Match the tool to how sensitive the data is

Rather than hunting for one “private AI”, sort your uses into three tiers. Each tier calls for different tools and precautions.

Tier 1: public or low-stakes information. Rewriting a job ad that is already published, brainstorming headlines, explaining a concept. A consumer assistant works, with training switched off as a habit.

Tier 2: internal information. Meeting notes, procedures, unpublished figures, sales documents. Use a business plan approved by your organization, with a no-training commitment and accounts managed by the company. Strip out anything the task does not need.

Tier 3: sensitive personal data or trade secrets. Health data, HR files, financial details, trade secrets, critical source code. Here the decision belongs to leadership and your privacy lead. Options range from an enterprise plan with controlled hosting to an open-weights model run on your own servers, which sends nothing outside. Sometimes the right answer is not to use AI for that task at all.

Check yourself A colleague wants an AI tool to summarize the 40 résumés received for a job opening. What is the right approach?

Résumés contain personal data, so this is tier 2, possibly tier 3. She should use a company-approved tool on a plan that excludes training, and remove names, contact details and photos where possible before uploading. Applicants should have been told how their data is handled, and the final decision must stay with a human. Your privacy lead can confirm what rules apply.

Set up simple rules in 6 steps

An AI acceptable use policy does not need to run fifty pages. For a small business or a team, these six steps cover the essentials.

1. Take inventory. Ask, without blame, which tools people already use and for what. You will often find useful habits worth making official.
2. Pick one or two approved tools. Compare training, retention, hosting and stated certifications using the vendor’s official pages, not just its marketing.
3. Read the contract. Check for a data processing agreement (DPA), the subprocessors listed and where processing happens. Get help from your privacy lead or counsel.
4. Configure accounts. Company-managed accounts, training off, retention set as short as practical, access removed when someone leaves.
5. Write a short policy. One page is enough: approved tools, the three data tiers, what is off-limits, who to ask when unsure.
6. Train and review. Explain the rules with concrete examples, then review them every six months: plans and settings change often.

For bringing your people on board, see our guide on how to train your team on AI. If you work solo or run a very small business, our article on AI for freelancers and small businesses offers a lighter version of these rules.

The special case of AI agents

An AI agent does more than answer: it reads your email, opens files, connects to your tools and takes action. It therefore sees far more data than a simple chatbot, and for longer.

Two precautions matter. First, grant the minimum access: one specific folder rather than the whole drive, read rather than write. Second, watch out for prompt injection: an instruction hidden in a web page or document can hijack the agent and make it send information elsewhere. Requiring human confirmation before any message or payment goes out limits that risk. Our look at AI agents for business shows where they genuinely help.

To place these rules within a broader approach, our practical guide to AI at work covers the everyday uses that save the most time.