Skip to main content

DPIAs for AI agents: what to assess before go-live

A data protection impact assessment for an AI agent has to cover autonomy, tool access and inference, not just the data. What to assess before launch.

Share -
A woman in a navy jacket works at a laptop in a dark office while glowing blue holographic profile cards of several people float in front of her.

A data protection impact assessment for an AI agent should cover what the agent can do, not only what data it reads. Assess its tool access, the decisions it takes alone, what it might infer about people, where its records go and who answers for each step. Finish it before go-live and revisit it whenever the agent gains a tool.

Why a standard DPIA template falls short

Most DPIA templates were written for systems with a fixed flow: data comes in through a known form, is processed in a known way and is stored in a known place. An AI agent breaks each of those assumptions. It chooses which tools to call, it can pull data from several systems in one run, and the same request can produce a different path on a different day.

The UK Information Commissioner's Office made this point in its January 2026 Tech Futures report on agentic AI. The ICO is clear that the report is not guidance, but it points back to existing law: if deploying an agentic system is likely to result in a high risk to people's personal information, a DPIA is required. For most agents that act on customer or staff records, that test is easy to meet.

The practical problem is that a template answered for "the CRM" or "the chatbot" will describe the data and miss the behaviour. A useful DPIA for an agent describes both.

What to assess for an AI agent

Purpose and scope of each tool

Write down the purpose of the agent in one sentence, then list every tool it can call and why it needs it. Purpose limitation is where agents drift first: a support agent given a general search tool can read far more than a support task requires. Any tool that does not serve the stated purpose should be removed or narrowed before launch, and access should follow the rules you already use to stop agents reaching restricted data.

Decisions taken without a person

Identify every point where the agent's output changes something for a person: approving a refund, rejecting an application, changing an account. Under UK and EU data protection law, decisions based solely on automated processing that have legal or similarly significant effects carry extra duties. Record which decisions the agent takes alone, which go to a reviewer, and how a person can challenge the result. Where a human in the loop is the safeguard, describe what the reviewer actually sees.

Inference about people

Agents combine data from several sources, and a model can infer things nobody gave it, including health, religion or financial difficulty from ordinary records. Special category data created by inference is still special category data. The DPIA should say whether the agent could infer such details, whether that matters to its task, and what stops the inference being stored or acted on.

Records, memory and retention

List everything the agent writes down: transcripts, tool call logs, memory, embeddings and cached answers. Each is personal data if it contains personal data, and each needs a retention period and a deletion route. This is where earlier decisions on keeping personal data out of logs, where the data is processed and how erasure requests reach the agent belong in the assessment.

Controllers, processors and model providers

Name who is the controller for each part of the processing, which suppliers act as processors and what the contracts say about training on your data. If the agent calls third-party tools or other agents, each one is another recipient to list. Unclear roles are the gap regulators and auditors ask about first.

Transparency to the people affected

Decide what people are told and when: that an AI agent is involved, what it does with their information and how to reach a person. This links directly to when an AI agent must disclose that it is an AI.

Who should run it and when

The DPIA belongs to the business owner of the agent, with the data protection officer advising and the engineering team supplying the facts. Engineers matter more here than for most systems, because only they can confirm which tools the agent really has, what it logs and what it can reach in practice. Start the assessment at design time, when narrowing a tool is cheap, and sign it off before any real personal data flows through the agent.

Keep the evidence alongside the assessment: the tool list, the permission scopes, the logging configuration and test results from runs on realistic data. An assessment that matches the audit trail of the running system is far easier to defend than one written from the design document alone.

Keep it current after launch

An agent changes more often than most systems. New tools, a new model, a wider data source or a new group of users can each change the risk. Treat the DPIA as a living record: tie a review to your change approval process, so that adding a tool or widening access cannot ship until the assessment has been updated. Periodic access reviews give a natural point to check that the DPIA still describes what the agent can actually do.

If you are preparing an AI agent for launch and want engineering input on its data protection assessment, talk to our team.

Frequently asked questions

Does every AI agent need a DPIA?

Not every one. A DPIA is required where processing is likely to result in a high risk to individuals. An internal agent that only summarises public documents may not meet that test. An agent that reads customer records, makes decisions about people, or combines data from several systems usually will. If you decide a DPIA is not needed, record why, because that decision is itself something you may need to show a regulator.

Is a DPIA the same as an AI impact assessment?

No, although they overlap. A DPIA focuses on risks to people from the processing of their personal data, and in the UK and EU it is a legal requirement in high-risk cases. An AI impact assessment, such as one used under an AI management system, can also cover fairness, safety and wider social effects. Many teams run one combined assessment, provided the DPIA questions are answered in full.

When should the DPIA be updated?

Update it whenever the agent's processing changes in a way that could change the risk: a new tool, wider data access, a different model provider, new user groups or a new decision the agent takes alone. Tie the review to your change approval process so updates are not left to memory, and review it on a regular schedule even when nothing obvious has changed, because small changes add up.

Who signs off a DPIA for an AI agent?

The business owner accountable for the agent signs it off, after taking advice from the data protection officer where one is appointed. Engineering should confirm the technical facts, such as tool access, logging and retention, because those determine whether the stated controls exist. If high risks remain that cannot be reduced, UK and EU law require you to consult the data protection regulator before the processing starts.

Share this post

Contact us to build the right product

Talk to our engineers about your application, the systems it connects to, and what you want to build next.

Get in touch
Two people discussing work with a laptop