The ICO wants evidence on agentic AI, and your data team should answer

What happened

The UK Information Commissioner's Office (ICO) has opened a call for evidence on how to apply data protection law to agentic AI. Its consultation page gives a start date of 8 October 2026 and a closing date of 20 November 2026, and shows the status as open. The ICO describes the aim as informing its thinking on how to apply data protection law to agentic AI's distinctive capabilities and risks, so it can identify where to give more clarity or practical support before publishing final guidance.

The survey is organised into sections: about you and your organisation, data security, transparency, accountability, automated decision-making, fairness and purpose limitation, lawfulness of processing, and additional questions. The page lists these topics and does not reproduce the individual questions. Responses go through the ICO's Citizen Space portal.

On publication, the ICO may publish organisational responses in full or as a summary, depending on what the respondent prefers. It removes personal contact details where appropriate, and tells respondents not to include anything they would not want made public. Late responses may not be considered.

Key dates, and the survey sections set against what we would check in your own agents:

Why it matters

The ICO is asking before it writes. A call for evidence comes ahead of guidance. Organisations that supply concrete examples have a better chance of seeing guidance that fits how agents are actually built. Those that stay silent should expect to adapt to it afterwards.

The section headings are a preview of the questions you will face. Data security, transparency, accountability, automated decision-making, fairness and purpose limitation, and lawfulness of processing cover most of what a regulator or a customer's data protection officer will ask about an agent. They are a useful audit frame even if you never respond.

Agents change the data question. A report that reads data is one thing. An agent that reads data, calls tools and acts on the result is another. The question moves from "who can see this?" to "what can this system do with it, on whose authority, and who would know?" Most data governance programmes were built for the first question.

Your data foundations will decide the answer. Whether you can say what an agent used and why depends on lineage, access control and logging that sit at the data layer. Teams that have invested there can respond with evidence. Teams that have not will find the gaps when they try.

What we would do

1. List your agents and what personal data each can reach. Include pilots, vendor-provided agents and anything a team has built on a platform you did not procure centrally. You cannot respond, or govern, what you have not counted.
2. Map each agent against the ICO's survey sections. For each, write a one-line answer on security, transparency, accountability, automated decision-making, fairness and purpose, and lawful basis. Mark the ones you cannot answer.
3. Decide who responds. Pair your data protection lead with an engineer who has built an agent. Practical detail on what worked, what was ambiguous and what support would help is what a regulator can use.
4. Respond before 20 November, and assume it may be published. Use the organisational option if you want the response published as a summary, and keep anything commercially sensitive out of it.
5. Check the decision points. Where an agent's output affects a person, record whether a human reviews it and how. Take legal advice on whether the automated decision-making rules apply to your use cases.
6. Fix logging at the data layer. Make sure access by agents is attributed to the agent and the task, so you can answer what it touched.

This is not yet guidance, and we would not treat it as such. It is a clear signal of where the regulator's attention is going, and an invitation that is cheap to accept.

It is also worth being realistic about the effort. Responding well takes a few hours of someone's time, not a project. The larger cost is discovering, while you try to answer, that nobody can say which agents are live or what they can read. That discovery is better made now, on your own terms, than during a customer audit or a regulator's enquiry.