Legal exposure
Privacy rules, the EU AI Act and your AI assistant: what applies

The short answer
For an assistant your staff use to look things up in company documents, the main rules are the privacy laws you already follow. If the documents hold personal information, the same duties apply when an AI reads them.
The AI-specific laws bite mainly in two cases: when customers talk to it, and when it helps make decisions about people, such as hiring or credit. The strictest of those rules start in 2027.
Think of building codes
An ordinary office has to meet the standard codes: fire exits, safe wiring. Use the same building as a school or a clinic, and stricter codes apply. What matters is what the building is used for, not how it was built.
The rules follow what the AI is used for.
What applies, by what you use it for
Staff look things up in company documents
The privacy laws you already follow, such as GDPR in Europe and state privacy laws and HIPAA in the US. In the EU, a duty to help staff understand the AI they use. Not "high-risk" under the EU AI Act.
The lightest rulesCustomers talk to it
Tell people they're talking to an AI, unless it's obvious. In the EU this has applied since 2 August 2026. Utah requires it when a customer asks, and up front in licensed professions.
Say it's an AIIt helps decide about people
Hiring, promotion, firing, credit or insurance. The EU's high-risk rules, plus California, Colorado and Illinois laws on automated decisions: notices, human review, risk checks.
The strictest rules
The dates that matter
Several of these dates moved in 2025 and 2026. In July 2026, a new EU law pushed back the AI Act's high-risk rules and softened its staff-training duty (European Commission). Colorado replaced its 2024 AI law with a new one in May 2026.
- Since February 2025: the EU's AI literacy dutyCompanies using AI must take steps to help staff understand it. Since July 2026 the wording is softer: support it, not guarantee it.
- Since January 2026: Illinois, AI in employment decisionsEmployers must tell employees when AI is used in decisions such as hiring or promotion, and may not use it in a way that discriminates (Public Act 103-0804).
- Since August 2026: the EU, "you're talking to an AI"AI that talks with people must say so, unless that's obvious.
- From January 2027: California and Colorado, automated decisionsNotices, the right to an explanation or human review, and risk checks, where AI helps make significant decisions about people (California Privacy Protection Agency; Colorado SB 26-189).
- From December 2027: the EU's high-risk rulesFor AI used in hiring, managing staff, credit and the other listed uses. Originally due August 2026. Employers must tell staff and their representatives before using it.
In the US, a December 2025 executive order set out to challenge state AI laws, but it doesn't cancel any of them by itself, and Congress hasn't passed a law that does (Executive Order 14365). The state laws above are in force or on their way.
Privacy law already covers the documents
Company documents are full of personal information: names in emails, HR files, customer records. When an AI system reads them, the privacy rules you already follow still apply. Under Europe's GDPR (General Data Protection Regulation), that usually means:
- A contract with the vendor saying what it may do with the data, as with any outside service.
- A legal route for data leaving Europe. The EU–US Data Privacy Framework is still valid, though an appeal against it is pending at the EU's top court.
- A risk assessment first where the use is likely to be high-risk for the people in the documents.
- Honoring people's rights, including deletion. If someone's data must go, it must go from the AI system's copies and catalogs too.
In US healthcare, HIPAA (the Health Insurance Portability and Accountability Act) requires a signed business associate agreement before a vendor handles patient information. Microsoft, Anthropic, OpenAI and Google offer one on business plans, not on free or personal ones (Microsoft, Anthropic).
An AI tool doesn't change what you may do with personal information. It just reads more of it, faster.
Three ways to set it up
Keep personal information out
Give the assistant policies, manuals and product documents, and leave out HR files, customer records and inboxes.
Best for: a first project. Most of the privacy work falls away.
Gives up: the questions that need those records, and someone must check what's really in the documents.
Let it in, with the privacy work done
Include records with personal information, with a vendor contract, a risk assessment and controls on who sees what.
Best for: HR, customer service and other teams whose questions are about people.
Gives up: time before launch, for your privacy team and your lawyer.
Use it in decisions about people
Let it help screen applications, assess performance or judge credit.
Best for: very few companies today, and only with legal advice from the start.
Gives up: simplicity. Notices, human review, testing for bias and records, under several laws at once.
What goes wrong
- Drifting into decisions about people. A recruiter starts asking the assistant which applicants look strongest. Now it's a hiring tool.
- Not knowing what's in the documents. Nobody checked the shared drive, and it holds payroll exports.
- Deleting from the source, but not the AI. The file is gone; the passages the system copied from it are not.
- Patient or client data in a personal account. No business associate agreement, no contract, no control.
- Reading the EU delay as "nothing applies yet". The literacy duty and the "it's an AI" rule already apply.
Questions to take to your lawyer and privacy team
- Which of the documents we plan to use contain personal information, and whose?
- Do we have a vendor contract that covers this, and where is the data processed?
- Do we need a privacy risk assessment before launch?
- Will customers, applicants or employees ever be on the other end of it, and what must we tell them?
- Could anyone use it to help decide about a person, and how would we know?
- If someone asks us to delete their data, can we delete it from the AI system's copies too?
Other names you may hear
- EU AI Act
- The European Union's law on AI, which sets rules by how risky a use is.
- Digital Omnibus
- The EU's 2025–26 package simplifying its digital laws. The AI part became law in July 2026; the privacy part is still being negotiated.
- High-risk AI
- Under the EU AI Act, AI used for listed purposes such as hiring, managing staff or credit, which carries the strictest duties.
- Provider, deployer
- EU terms for the company that makes an AI system and the company that uses it.
- AI literacy
- Helping staff understand the AI tools they use, and their limits.
- GDPR, General Data Protection Regulation
- Europe's main privacy law.
- DPIA, data protection impact assessment
- A written privacy risk assessment, required before risky uses of personal data.
- Automated decision-making, ADMT
- Software that makes or helps make decisions about people. California's rules call it automated decision-making technology.
- BAA, business associate agreement
- The contract HIPAA requires before a vendor handles patient information.