# An AI Assistant on Your Company Documents: What GDPR Actually Requires
- **url:** https://easify-ai.com/ai-assistant-on-company-documents-gdpr-requirements/
- **published:** 2026-11-16
- **topics:** ai-automation-in-business
- **publisher:** Easify AI, part of Easify Technologies
Your team wants to ask questions of the company's contracts, policies and project files and get an answer with the source linked. The technology for this (often called RAG, retrieval-augmented generation) is mature. The question that stalls most projects is a different one: what does GDPR expect from us?
The good news is that GDPR does not treat an AI assistant as a special case. It asks the same questions it asks of any new system that handles personal data. This guide walks through them in the order you will meet them, written for the ops or IT lead who has to sign the project off. It is general guidance, not legal advice.
## Start with what the assistant will actually read
Company documents are full of personal data, even when nobody thinks of them that way. Contracts name contacts. Project files contain emails. HR folders contain salaries, sickness notes and performance reviews.
Before anything else, list the sources you plan to connect and sort them into three groups:
| Source type | Examples | Typical approach |
|---|---|---|
| Low personal data | Policies, manuals, product docs, templates | Connect first |
| Ordinary personal data | Contracts, project files, CRM notes, email archives | Connect with access control and retention rules |
| Special category or high risk | HR files, health data, disciplinary records, payroll | Leave out, or treat as a separate project with a DPIA |
> **Tip:** most of the value of a first assistant comes from the first row. Policies, procedures and manuals answer a large share of internal questions and carry little risk. Start there and expand once the controls are proven.
## Lawful basis: usually legitimate interests
Every processing activity needs a lawful basis under Article 6. For an internal assistant, the usual candidate is legitimate interests: helping staff find information faster is a real business interest, and the processing is close to what already happens when someone searches the shared drive.
Legitimate interests needs a short written balancing test. Record the purpose, why the assistant is necessary for it, and why it does not override the interests of the people named in the documents. Consent is rarely the right choice between employer and employee, because staff cannot easily refuse.
If you plan to include special category data such as health information, Article 6 is not enough on its own: you would also need a condition under Article 9. That is a strong reason to leave those folders out.
## Access control: the assistant must not see more than the user
This is where most real problems happen. A document assistant that indexes everything and answers anyone will happily summarise the CEO's salary review for a new starter. That is a confidentiality failure and, under GDPR, a breach of the integrity and confidentiality principle in Article 5(1)(f).
The rule to build to is simple: **each answer may only draw on documents the asking user could already open.** In practice that means:
1. **Carry permissions into the index.** When documents are indexed, store who may read each one (from SharePoint, Google Drive or your document system).
2. **Filter at retrieval time.** Before the model sees any text, drop every passage the user is not allowed to read.
3. **Keep permissions in sync.** When someone leaves a project or a folder is locked down, the index must follow within hours, not weeks.
4. **Test with real accounts.** Log in as a junior user and try to retrieve something you know they should not see.
> **Watch out:** the assistant will expose permission mistakes that already exist. If a folder of payslips is open to the whole company today, nobody may have noticed. An assistant that can find anything in seconds will. Clean up broad sharing before you connect a source.
Data minimisation (Article 5(1)(c)) points the same way. Connect the folders the use case needs, not the whole tenant.
## Processors and DPAs (Article 28)
Your model provider, your vector database host and any integration platform are processors: they handle personal data on your behalf. Article 28 requires a written contract with each one, covering what they may do with the data, security, sub-processors, deletion at the end, and help with rights requests and breaches.
The large providers all publish a data processing addendum for business and API customers. Check three things in each:
- **Training:** business and API data should not be used to train the provider's models by default. Confirm it in the contract, not just on a marketing page.
- **Retention:** how long prompts and outputs are kept for abuse monitoring, and whether a zero-retention option is available to you.
- **Sub-processors:** who else touches the data, and where they are.
## International transfers: where does the data go?
If personal data leaves the EU or EEA, Chapter V of GDPR applies. For US providers there are two common routes:
- **The EU-US Data Privacy Framework.** The Commission's 2023 adequacy decision lets you transfer to US companies certified under the framework. The EU General Court upheld it in September 2025; an appeal to the Court of Justice followed, so keep an eye on it.
- **Standard Contractual Clauses (SCCs).** Most provider DPAs include them as a fallback, with a transfer impact assessment on your side.
Many teams prefer to avoid the question by keeping processing in Europe. As of November 2026, this is what the main providers offer:
| Provider | EU option | What to check |
|---|---|---|
| OpenAI | European data residency for eligible API projects and ChatGPT Enterprise/Edu workspaces; must be chosen when the project or workspace is created | Eligibility, and which features are covered |
| Anthropic (Claude API direct) | Inference location can be set to "global" or "us" only; workspace storage is US | Use an EU region on a cloud platform instead |
| Claude via AWS Bedrock or Google Cloud | Regional endpoints and EU inference profiles, depending on the model | Model availability per region |
| Azure OpenAI (Microsoft Foundry) | "Data Zone" deployments in the EU zone keep processing within the EU Data Boundary; "Standard" keeps it in one Azure geography | Avoid "Global" deployment types, which may process anywhere |
| Google Vertex AI | Regional EU endpoints, and an "eu" multi-region endpoint that keeps processing within EU member states | Global endpoints give no residency guarantee; per-model support varies |
> **Note:** "EU region" usually covers stored data and model processing, but not always support access or every feature. Read the provider's own residency page for the exact scope, and check it again before go-live, because these options change often.
The other half of the pipeline matters too. Your vector database, file storage and logs should sit in the EU as well, or the residency of the model alone does little. Our [guide to a GDPR-safe RAG pipeline on European servers](/gdpr-safe-rag-pipeline-on-european-servers-guide/) covers the architecture.
## When you need a DPIA (Article 35)
A data protection impact assessment is required when processing is likely to result in a high risk to people. Regulators' lists commonly flag new technologies, large-scale processing, and monitoring or evaluating employees.
An assistant over policies and manuals is unlikely to need one. An assistant over HR files, customer records at scale, or one used to assess staff probably does. If you are unsure, do a short DPIA anyway: it forces the decisions this article describes, and it is the record a regulator would ask to see.
## Retention, logging and rights requests
An assistant creates new personal data: the questions people ask, the answers, and the logs of which documents were retrieved. Decide up front:
- **Prompt and answer history.** Keep it only as long as it is useful, for example 30 to 90 days for troubleshooting, then delete.
- **Audit logs.** Who asked what and which sources were used. These help you investigate misuse; keep them for a defined period with restricted access.
- **Index freshness.** When a source document is deleted under your retention policy, remove it from the index too.
Rights requests then become manageable. Under Articles 12 and 15, you normally have one month to answer an access request. The assistant can help you find documents that mention someone, but a person should check what is disclosed. Stored prompts and answers that mention the requester are in scope as well.
## A launch flow that keeps a person in charge
```flow
trigger | Source approved | Folder, purpose and owner agreed | SharePoint
tool | Index with permissions | Who may read each document | PostgreSQL
ai | Answer with sources | Only from documents the user can open | Azure OpenAI
human | Owner reviews | Access tests, logs and feedback weekly
output | Expand or fix | Next source, or tighten access | Microsoft Teams
```
Each new source goes through the same loop: approve, index, test with real accounts, review. This is how we build document assistants in [Bison](/systems/bison/), our operations system, where every answer links to the document it came from.
## Summary
- **Know your sources:** list what the assistant will read, and leave special category data out at first.
- **Pick a lawful basis:** usually legitimate interests, with a written balancing test.
- **Enforce permissions:** each answer may only use documents the user could already open.
- **Sign DPAs and settle transfers:** use EU processing where you can, and SCCs or the Data Privacy Framework where you cannot.
- **Run a DPIA** when HR, health or large-scale customer data is involved.
- **Set retention** for prompts, answers and logs before launch, and plan for rights requests.
## FAQ
### Do we need consent from employees to run an AI assistant on company documents?
Usually not. Consent is a weak basis between employer and employee because it is rarely freely given. Most companies rely on legitimate interests, backed by a short written assessment, and tell staff how the assistant works in their privacy notice.
### Is it enough that our AI provider has EU data residency?
It helps, but it is not the whole answer. You still need a data processing agreement, you should check whether support staff or sub-processors outside the EU can access data, and you need access control and retention rules on your side.
### Do we need a DPIA for an internal document assistant?
Not always, but often. If the documents include HR files, health data, customer records at scale, or the assistant is used to evaluate people, a DPIA is likely required. Many regulators list innovative technology as a factor, so doing one is the safer default.
### Can the assistant answer a subject access request for us?
It can help you find documents that mention a person, but a person should review what is disclosed. Remember that prompts and answers stored by the assistant may themselves be personal data that falls within the request.