The ICO's Call for Evidence on Agentic AI: What It Asks and What Teams Building AI Agents Should Have Ready
Most teams building AI agents in the UK have spent the last year thinking about two questions: does the agent work, and who carries the loss when it does not? The Information Commissioner's Office has now added a third. Its call for evidence on agentic AI, which closes on 20 November 2026, asks how agents that act on personal data can be secure, transparent, accountable, fair and lawful, and how they handle automated decisions. That is not a new rulebook. It is a regulator signalling which questions it expects builders to be able to answer. This article reads the call as a practical checklist and sets out the evidence pack a product, risk or engineering team should keep, whether or not it sends a response.
This is a technology partner's reading of a public consultation, not legal advice. Read the ICO's own page, and take advice on your own position.
What the ICO has actually asked for
The ICO's consultation page describes a call for evidence on agentic AI covering six areas: security, transparency, accountability, automated decisions, fairness and lawfulness. It closes on 20 November 2026.
Three points about its status matter before anyone reads too much into it.
First, a call for evidence comes before guidance. The ICO is gathering views and examples; it has not set out new expectations specific to agents. Anyone describing it as "new ICO rules for AI agents" is overstating it.
Second, the existing law already applies. UK data protection law already covers personal data processed by software, including software that plans and acts on its own. It also already contains rules on solely automated decisions that have legal or similarly significant effects on people. An agent does not sit outside those rules because it is new.
Third, the six themes are a useful map even for teams that never respond. They describe the questions a regulator, an auditor, a bank partner or a large customer is likely to ask about any agent that touches personal data.
Why a data protection regulator is looking at agents now
The timing is not accidental. Through October and November, the public debate about agents moved from capability to control. Visa published a benchmark, designed and run by Visa itself, which found that 22 AI models failed roughly 43% of realistic shopping test runs, and that only 23% of US consumers trust AI to pay for them. Bank of America told customers not to share banking credentials with AI assistants, and United Airlines barred agents from completing bookings, according to reporting by TechCrunch and PYMNTS. In the UK, the FCA published "AI in financial services: good practice and areas for improvement", which says existing rules apply and frames AI use as support, automate or abstain.
The ICO's call widens that picture from financial conduct to personal data. For a UK firm building an agent that reads a customer's account, books on their behalf or triages their complaint, the two regulators are now asking overlapping questions from different directions.
The six themes, read as design questions
The most useful way to read the call is to turn each theme into a question you could answer in a design review today. If you cannot answer one, that is your gap, regardless of what the ICO eventually publishes.
Security: what can the agent reach, and what stops it reaching further?
An agent is a piece of software with credentials and tools. The security question is about scope. Which systems can it call? Which personal data can it read, and which can it change? What happens if a prompt, a web page or a document tries to redirect it?
Evidence to keep: a list of every tool and data source the agent can use, the permission each one carries, and the tests you have run against attempts to make the agent act outside that scope.
Transparency: does the person know an agent is acting, and on what?
Transparency for agents has two audiences. The customer needs to know when an agent, rather than a person, is acting on their data or their account. Your own staff need to know what the agent did and why, in a form they can read.
Evidence to keep: the customer-facing notice that explains agent involvement, and a plain-language description of what the agent does with personal data, written so a non-engineer can follow it.
Accountability: who owns the agent's actions?
Accountability is the theme most teams find hardest, because agents blur the line between a tool and a decision-maker. The industry is converging on a simple principle. In a recent paper on agentic payments in wholesale finance, Deutsche Bank and Ant Digital set out that agents need verifiable identity bound to an accountable owner, according to Finextra's reporting. The same principle applies to an agent that processes personal data: someone in your organisation must own what it does.
Evidence to keep: a named accountable owner for each agent, a record of who approved its deployment and on what evidence, and a decision log that shows what the agent did in each case.
Automated decisions: where does the agent decide, and where does a person?
This is where existing UK law already bites. If an agent makes a decision about a person that has a legal or similarly significant effect, such as declining an application, closing an account or refusing a refund, the rules on automated decision-making are engaged. Teams need to know which of the agent's actions fall into that category.
The FCA's support, automate or abstain framing is helpful here. For each task, decide whether the agent supports a person who decides, automates the decision itself, or abstains and hands over. Write that down per task, not per agent.
Evidence to keep: a task-by-task map showing which actions the agent may take alone, which need human approval, and which it must refuse, together with the route a customer has to challenge an automated outcome.
Fairness: does the agent treat comparable people comparably?
Fairness is often discussed for models in isolation. Agents add a new layer, because the same model can produce different outcomes depending on the tools it calls and the order it calls them in. Visa's benchmark is a reminder that consistency, not raw capability, was the failure point in its tests.
Evidence to keep: repeat-run tests on representative cases, with results broken down by the customer groups that matter for your product, and a record of what you changed when outcomes diverged.
Lawfulness: what is your basis for each use of personal data?
An agent can combine data in ways its designers did not plan. Each purpose for which it processes personal data needs a lawful basis, and data collected for one purpose should not quietly be repurposed because an agent found it useful.
Evidence to keep: a data protection impact assessment that covers the agent specifically, not just the wider product, and a record of the purposes and lawful basis for each category of personal data it touches.
The evidence pack: one folder, kept current
Pulling those themes together gives a short, practical evidence pack. It is worth keeping for any agent that processes personal data, whether or not you respond to the ICO.
- Scope register. Every tool, system and data source the agent can reach, with the permission level for each.
- Task map. For each task: support, automate or abstain, and the human approval points.
- Accountable owner. A named person, the approval record for deployment and the evidence it relied on.
- Decision log. What the agent did, on what inputs, with what outcome, in a form staff can read.
- Customer notice. How people are told an agent is involved, and how they can challenge an outcome.
- Test record. Scope-escape tests, repeat-run consistency results and fairness checks by customer group.
- Impact assessment. A data protection impact assessment specific to the agent, with purposes and lawful basis.
- Stop rules. The conditions under which the agent must stop and hand over, and who owns the handover queue.
None of this is exotic. Most of it is what a careful team would build anyway. The difference is having it written down, dated and owned before someone asks for it.
Should you respond before 20 November?
Responding is a business decision, not an obligation. There are reasonable arguments either way.
A response makes sense if you have practical experience to share, for example a specific difficulty with consent, transparency or automated decisions in a live agent. Regulators use calls for evidence to understand real constraints, and concrete examples tend to be more useful than general positions.
It may make less sense if your agent work is still at proof-of-concept stage, or if you would mainly be repeating points others will make. In that case, the more valuable step is to use the six themes as an internal review and close your own gaps.
Either way, the closing date is fixed. If you intend to respond, check the ICO's consultation page for the response format and deadline in its own words.
What changes for teams already building agents
For most teams the call does not change what to build. It changes what to be able to show. Three practical shifts follow.
Design evidence moves earlier. Decision logs, scope registers and task maps are much cheaper to build into an agent from the first sprint than to reconstruct after launch. Treat them as product features with acceptance criteria, not compliance paperwork.
Data protection and conduct reviews converge. For a UK financial services firm, the ICO's themes and the FCA's good-practice document overlap heavily. A single review that answers both sets of questions avoids duplicated work and contradictory answers.
Accountability gets a name. "The AI team" is not an accountable owner. Each agent needs a person whose job includes its behaviour, with the authority to pause it.
Where Innovify fits
Innovify is a London-based technology partner with deep expertise in banking, financial services, fintech and insurance. We design compliance in from day one, which for agents means building the scope register, task map, decision log and stop rules into the product rather than adding them before an audit. Through our AI Consulting work we help teams turn regulator themes like the ICO's into a concrete design review for a specific agent, and our AI / ML Development teams build those controls into production systems. If you are deciding what your agent needs to evidence, that review is a sensible place to start.
FAQ
What is the ICO's call for evidence on agentic AI?
It is a public consultation in which the Information Commissioner's Office asks industry for evidence about agentic AI. It covers security, transparency, accountability, automated decisions, fairness and lawfulness, and it closes on 20 November 2026. It gathers views ahead of any guidance; it does not create new rules.
Does UK data protection law already apply to AI agents?
Yes. Existing UK data protection law applies to personal data processed by AI agents in the same way it applies to other software. That includes the existing rules on solely automated decisions with legal or similarly significant effects. The call for evidence is about how those rules apply to agents, not whether they apply.
Do we have to respond to the ICO's call for evidence?
No. Responding is voluntary. It is most useful when you have concrete experience to share, such as a specific difficulty in a live agent. If you do not respond, the six themes still make a practical internal review for any agent that handles personal data.
What evidence should a team building an AI agent keep?
Keep a scope register of the tools and data the agent can reach, a task map of what it may do alone or with approval, a named accountable owner, a readable decision log, the customer notice, test records and an agent-specific data protection impact assessment. Add written stop rules and a named owner for the handover queue.
How does the ICO call relate to the FCA's AI good-practice document?
They come from different angles but overlap. The FCA document, published in October 2026, says existing conduct rules apply and frames AI as support, automate or abstain. The ICO call looks at personal data. A UK financial services firm can answer both sets of questions in one design review.
Conclusion
The ICO has not written new rules for AI agents. It has told the market which questions matter: security, transparency, accountability, automated decisions, fairness and lawfulness. Teams that can answer those questions today, with a dated, owned evidence pack, are in a stronger position with regulators, bank partners and customers alike, whatever guidance follows. If you plan to respond, do it before 20 November. If you do not, use the six themes as this month's design review.
Related reading
- What the FCA's October AI Statement Means for Teams Building Financial Products: Good Practice, Poor Practice and Disclaimers on AI Advice
- Regulators Are Watching Agentic AI: A Practical FCA and Bank of England Briefing for Product Leaders
- AI Governance for Fintech Platforms: What Changes the Day AI Becomes Infrastructure
- Data Governance for Enterprise AI: What Changes When Models Consume Your Data











