AI agents and UK GDPR

Picture a practice manager on a Tuesday afternoon. She has a folder of client letters, a deadline on Thursday, and an agent that can read the shared drive, draft replies and send email. She sets it going. By Wednesday it has done a week's work. Did anyone check what it touched, who it wrote to, or what it kept?

That is the situation the Information Commissioner's Office is now asking about. On 8 October it opened a call for evidence on agentic AI. These are systems that take a goal, use tools and carry out several steps without anyone watching each one. The call runs until 20 November 2026, and what comes back will shape the ICO's guidance on how UK data protection law applies to agents.

What the ICO has said so far

The call is short on new rules and long on signals. They are hard to miss.

The ICO is blunt. The fact that AI agents act with autonomy "is not an excuse for poor compliance." If an agent handles personal data, the duties apply in full. Nobody gets a lighter touch because the software picked its own next step.

The regulator is also looking at what has happened in testing. It has made enquiries with OpenAI, Anthropic, Meta and the UK AI Security Institute. It describes reports of agents that bypassed protections, used unauthorised communication channels and reached external systems. Those are reports, not findings. But the ICO is paying attention, and so should you.

This is the start, not the end. The call feeds a statutory code of practice on AI and automated decision-making that is still to come. The ICO is also updating its automated decision-making guidance for the Data (Use and Access) Act 2025.

The six areas

The call is split into six areas. Each one is a question you should be able to answer about any agent you run. The wording below is mine, not the ICO's.

• Security. What can the agent reach, with what permissions, and who can change them?
• Transparency. Do the people whose data it handles know an agent is involved?
• Accountability. Who owns the agent, who signs off what it produces, and where is the record?
• Automated decisions. Does the agent make a decision about someone, or shape one? If it does, can a person review it?
• Fairness and purpose. Is it only used for what the data was collected for?
• Lawful basis. What is the basis for each kind of data it touches, special category data included?

Six checks before an agent touches personal data

None of this is new in principle. Agents just make it easy to skip. An agent can do a month's work before anyone reads the log.

1. Map the data. Include what the agent generates about people, not only what it is given.
2. Write down the purpose and the lawful basis. One sentence per agent is enough to start. If you cannot write it, the agent is not ready.
3. Give it the least access that does the job. A summariser does not need to send email.
4. Put a person in front of anything with an outside effect. Sending, paying, publishing, deleting.
5. Keep a record you can actually produce. What it did, on which data, with which model, and who approved it.
6. Know where the model runs and who else processes the data. Which provider, which region, and what the contract says about training and retention.

If an agent's output feeds a decision about a person, such as credit, hiring or a claim, the automated decision questions bite hardest. Those are the agents to govern first.

What does not change yet

The call is not guidance, and the final guidance is months away. You do not have to wait for it. The principles already in UK GDPR are lawfulness, transparency, security and accountability. The ICO will test you against those, so design for them now. This is general information, not legal advice. Check your position with your own adviser.

If you have EU clients too, there are two dates. The Article 50 duties, which include telling people they are talking to an AI system, have applied since 2 August 2026. The obligations for stand-alone high-risk systems under Annex III moved to 2 December 2027 under Regulation (EU) 2026/1744. Our AI governance health check covers EU AI Act exposure if you want to see where you stand.

Where Odokai fits

A platform will not make you compliant. The lawful basis, the impact assessment and the call on what an agent may do stay with you. What a platform can do is make those decisions stick.

• Where the data sits. Agents run in your own AWS account or on our shared platform. You choose.
• Who can do what. Roles, approvals and an audit trail show who could act and what they did. How Odokai works has the detail.
• Which model. You pick it per task, local models included, so sensitive work does not have to go wherever everything else goes.

Use the time before 20 November. Take each agent you run, or plan to run, through the six questions. The ones you cannot answer are the ones to fix first.

You can respond to the call on the ICO's consultation page. If you are working out where client data should sit with AI, read our earlier piece on whether client data is safe in ChatGPT. And if you want to take a plan through these questions with us, get in touch.

The agent can't answer the ICO's letter. You can.