When AI Begins to Act, How Does Accountability Keep Up?

The discussion of artificial intelligence often begins with what a model knows and how it learned it. As AI agents gain access to tools, websites, and external systems, another question becomes increasingly important. How can organizations explain and control what those systems do with personal information?

On October 8, 2026, the UK Information Commissioner’s Office (ICO) published a report on its supervision of ten major foundation model developers and launched a call for evidence on agentic AI. According to the regulator, the developers have made, or committed to make, changes involving transparency, mechanisms for exercising information rights, and assessment of safeguards. The ICO is monitoring those commitments (ICO, 2026a).

Taken together, these developments suggest a widening focus. The information used to build an AI system remains important, but so does the behavior of the system once it begins interacting with the world.

From Training Data to Meaningful Control

The ICO’s report, Building trust and transparency into generative AI development: our work to create regulatory certainty, addresses how UK data protection law applies to foundation model development. Its expectations include identifying a lawful basis for processing, providing meaningful transparency, enabling people to exercise their rights, and demonstrating effective safeguards (ICO, 2026b).

The report also addresses whether models themselves contain personal data. The ICO calls for case-by-case assessment, including the possibility of extracting information or identifying people through singling out or linking records. Processing special category data requires both an Article 6 lawful basis and an applicable Article 9 condition. The regulator acknowledges technical difficulties while maintaining that these do not justify neglecting data protection obligations (ICO, 2026b).

For readers outside the UK, the report also raises a broader governance question. What does meaningful control look like when information has moved beyond an identifiable record and helped shape a model?

A person may understand how to delete a document from an account. Understanding whether information from that document remains represented in a trained model is more difficult. That difference makes the relationship between technical design and usable rights especially significant.

When Access Becomes Action

Agentic systems introduce another layer of complexity. A model connected to tools may retrieve information, communicate with other systems, and carry out steps toward a user’s goal. The relevant question extends beyond the information available to the model to the permissions and decisions surrounding its use.

In its earlier Tech Futures report, the ICO identified potential risks involving purpose limitation, data minimization, transparency, and accountability. It described how agents could pursue a goal in unexpected ways, seek additional personal information, or repurpose information. That report presents early thinking rather than formal regulatory expectations (ICO, 2026c).

Consider an assistant asked to arrange a meeting. Access to a calendar may support that task. Access to an entire mailbox could expose unrelated conversations, attachments, and information about other people. If the assistant can also send messages, a decision about access becomes a decision about possible disclosure.

This hypothetical illustrates why a useful result does not answer every governance question. What information was necessary? Which permissions were granted? Could the user understand the process or intervene before an action occurred?

The October call for evidence seeks input on security, transparency, accountability, automated decision-making, fairness, purpose limitation, and lawfulness. It closes on November 20, 2026. The ICO says the evidence will inform future guidance and its forthcoming statutory code of practice on AI and automated decision-making (ICO, 2026a, 2026d).

The consultation is an evidence-gathering process, not a completed code or a new blanket prohibition on AI agents.

As systems become more capable, accountability may depend increasingly on whether organizations can trace decisions, explain information flows, and provide practical routes for intervention. Greater autonomy does not, by itself, explain who should answer for an outcome.

The emerging question is how to preserve meaningful human and institutional responsibility when the steps between an instruction and its result become harder to observe.

This post reflects the ICO materials available on October 14, 2026. The consultation and subsequent guidance may develop further.

References

• Information Commissioner’s Office. (2026a, October 8). “ICO secures changes from leading AI developers as scrutiny extends to AI agents.”
• Information Commissioner’s Office. (2026b). Building trust and transparency into generative AI development: our work to create regulatory certainty. See “Data protection challenges and our expectations.”
• Information Commissioner’s Office. (2026c). ICO Tech Futures: Agentic AI. “Introduction” and “Data protection and privacy risks.”
• Information Commissioner’s Office. (2026d, October 8). “Agentic AI call for evidence.”