The ICO has secured changes from ten AI developers, and scrutiny now extends to AI agents

The ICO has stated that ten leading foundation-model developers made or committed to changes after its scrutiny. The changes cover clearer transparency information, stronger ways for people to exercise their rights, and tougher safeguard assessments. The ICO says it is monitoring progress.

The ICO also states that one developer’s engagement is paused and that a formal ICO investigation into Grok is ongoing.

The announcement does not state new binding obligations. It reports regulatory engagement and its results, which tells customers what the regulator now expects from model suppliers.

The call for evidence has eight sections, including data security, transparency, accountability, automated decision-making, fairness and purpose limitation, and lawfulness.

The ICO says the autonomy of AI agents is not an excuse for poor compliance. The point is that an organisation cannot treat an agent’s independent action as outside its data protection responsibilities.

Final guidance on agentic AI is expected in autumn 2026, but that expectation comes from the ICO’s plans page, which was not opened for this article, so it should be confirmed.

When an AI agent browses, retrieves, writes or acts on a person’s behalf, personal data moves through several parties: the model supplier, the platform that hosts the agent, the business deploying it and sometimes the end user.

Each party’s data protection role depends on who decides the purposes and means of the processing. The ICO’s call for evidence includes accountability as a section, which suggests role allocation is a live question.

A contract that describes the supplier as a processor while the supplier in practice decides how data is used for its own purposes has a role problem that no clause can fix. The roles should be mapped to the actual data flows before the paper is drafted.

The ICO’s developer changes concern clearer transparency information and stronger ways for people to exercise their rights. Customers who build on those models inherit the downstream obligation to tell their own users what happens to their data.

Where an agent contributes to decisions about people, automated decision-making is one of the call for evidence’s sections. A customer needs enough information from the supplier to explain the decision and to support a request for human review.

The practical contract point is information and assistance: the supplier should commit to provide the documentation, logs and cooperation the customer needs to meet its own obligations.

No binding new obligations were stated in the ICO’s announcement. The changes by developers were made voluntarily or by commitment, and the call for evidence is a consultation.

The names of the developers and the detail of each change should be read on the ICO’s pages before being relied on. The ICO’s statement about the paused engagement and the Grok investigation is the ICO’s own description of its work, and no finding has been made public in what was read.

The sensible reading is that regulatory attention is moving from models to agents, and that contract terms agreed now will be tested against the final guidance.

The following is drafting advice, not a statement of what the ICO requires. It should be adjusted once the final agentic AI guidance is published.

• Role allocation, stating whether the supplier is a controller, joint controller or processor for each agent function, with the processing purposes written down.
• Transparency and rights support, requiring the supplier to provide the information and assistance the customer needs to explain processing and handle rights requests.
• Automated decision terms, covering which decisions an agent may make, when human review applies, and what records are kept.
• Change notice, requiring notice of material changes to a model, an agent’s capabilities or its data use, with a right to suspend or exit.
• Audit and incident terms, giving the customer access to evidence of safeguards and a defined route and timetable for incidents involving personal data.