The ICO is asking how businesses control AI agents. Can you answer?

· · AI

If somebody in your business has set up an AI tool that does things, sends the follow-up emails, books the meetings, looks up a customer and updates the record, then personal data is being handled by software that makes its own choices about the steps. Under UK data protection law the responsibility for that sits with you. It does not sit with the tool, and for the most part it does not sit with the company that built it.

The regulator has just said so in as many words, and has started collecting evidence on how businesses are coping.

What the ICO announced

Two things came out of the Information Commissioner's Office on 8 October, as The Register reports.

The first concerns the companies that make the models. Ten large AI developers, Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI, and Stability AI, have made changes or committed to them after the ICO examined how they handle personal data. The commitments cover clearer explanations of how personal information is used to train models and better ways for people to exercise their rights. The ICO is not calling this finished. It says it will check that the promises are kept, and it accepts that some problems remain open, such as how a person gets their details removed from a model that has already been trained on them.

The second is the part that reaches ordinary businesses. The ICO has turned to agentic AI, the term for AI that carries out tasks with limited supervision, not just answering questions. It has contacted several developers after reports of agents getting round safeguards. Richard Nevinson, its director of technology regulation, put the position in one line: "the fact AI agents act with autonomy is not an excuse for poor compliance."

Alongside that, it has opened a call for evidence on agentic AI. Responses are due by 20 November. The ICO says what it learns will shape new guidance and its forthcoming statutory code of practice on AI and automated decision-making.

Why a call for evidence matters to you

A call for evidence is not a rule. Nothing changes on 20 November. But it is the clearest view you will get of what the regulator intends to ask about later, and the ICO addresses it to "developers, deployers and other experts". A deployer is any organisation that puts one of these tools to work. That is a much longer list than the ten companies above.

The call is organised around six themes. Read as questions to your own business, they make a fair self-check.

  1. Data security. What can the agent reach? Could something it reads, an email or a web page, persuade it to send information somewhere it should not go?
  2. Transparency. An agent may pass data through several suppliers to finish one task. Do you know which, and could you tell a customer?
  3. Accountability. When it gets something wrong, who answers for it: you, the tool's maker, or the firm that set it up? Is that written down anywhere?
  4. Automated decisions. Is the agent making decisions about people, such as who gets a refund, an interview, or a credit limit, without a person involved? The law has specific safeguards for that.
  5. Fairness and purpose. Is it using personal data only for the job you collected it for, or wandering into whatever it can find?
  6. Lawfulness. For each thing it does with personal data, do you have a lawful reason, the same as you would need if a member of staff did it?

Most small businesses will not have tidy answers to all six, and the ICO knows that. It says it wants to hear what organisations are doing now and "what further guidance and support they want from us".

What to do

  • Find out what is already running. Ask around. Agent features have arrived inside tools you already pay for, including office software, customer systems, and browsers. Staff may have switched them on without it feeling like a decision.
  • For each one, write down three things. What it can access, what it can do without asking, and who set it up.
  • Narrow the access. An agent that drafts replies does not need the whole customer database. Give it what the task needs.
  • Put a person in front of anything that affects someone. Decisions about money, jobs, or access to a service should be approved by a human until you are confident about the rest.
  • Update what you tell people. If an AI tool handles customer details in a way your privacy notice does not mention, add it.
  • Consider responding. If you have tried to use these tools properly and found the rules unclear, that is the evidence the ICO is asking for. Small firms are heard least in these exercises and affected most by the guidance that follows.

How Steelwise can help

Working out which AI tools are acting on your behalf, what they can reach, and what you would say if asked is the starting point of our AI advisory work. Get in touch.

Further reading

← All filings

We use cookies to understand how visitors use this site. More information.