Reasonable steps, for a senior manager overseeing AI agents, means being able to show at the time, not afterwards, that you knew what each agent in your area could do, approved its limits on evidence, received regular reports on its behaviour, acted on exceptions, and could stop it. A contemporaneous record of those steps is the defence.
What does the reasonable steps standard require, in plain words?
Under the Senior Managers and Certification Regime, each senior manager is responsible for a defined area. If the firm breaches a requirement in that area, the regulator can take action against the senior manager personally where they did not take the steps a reasonable person in their position would have taken to prevent or stop the breach.
The standard is judged against what the senior manager knew, or should have known, at the time. It considers the size and complexity of the firm, the risk in the area and what information was available. It does not demand perfection. It demands that the person acted like a competent manager who took the risk seriously.
The FCA and PRA have not written agent-specific guidance on this standard. It applies to agents in the same way it applies to any process a senior manager oversees.
Who has to meet it?
Every approved senior manager whose responsibilities include an area where agents do regulated work. That includes the obvious owners, such as the head of retail lending whose credit agents draft decisions, and less obvious ones, such as the senior manager responsible for operations when an agent reconciles client money.
The first reasonable step is knowing the agents are yours. Our guide to which senior manager owns an AI agent covers how to allocate each agent, and who is the accountable senior manager works through disputed cases.
Why does an agent that acts raise the bar?
An assistant that only answers questions affects few outcomes directly. An agent that sends letters, closes complaints or moves money produces regulated outcomes at speed, often without a person seeing each one. The EU debate on what changes when the system acts makes the same distinction.
That changes what a reasonable senior manager would do. With a human team, sampling a few files a month might be enough. With an agent, a reasonable person would want to know its limits are enforced by the system, not just written in a policy, and would want exception reports rather than samples.
Autonomy granted by configuration is also hard to defend. If an agent was switched to send letters without a person because someone ticked a setting, the senior manager will struggle to show a reasonable basis. If it was moved up only after tested evidence on the firm's own cases, the basis is on file.
Which steps should a senior manager be able to show?
The table lists the steps a reasonable senior manager would take for agents in their area, the control behind each, and the evidence it leaves.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Know what agents operate in your area | Agent register filtered by senior manager | Register extract reviewed and signed at a set interval |
| Approve each agent's authority on evidence | Depth raised only after certification on the firm's own cases | Certification results and a signed approval for each step up |
| Understand the limits that apply | Five bounds checked at every action: role scope, certificate, sources, firm limits, platform guardrails | Bound definitions and the verdict record per action |
| Receive regular management information | Reports on volumes, held items, refusals, overrides and complaints | Report packs and the senior manager's recorded responses |
| Act on warning signs | Thresholds that trigger review, depth reduction or suspension | Exception log with decisions and dates |
| Keep competent people in the loop | Approvers trained on each agent's failure modes | Training records and approval quality reviews |
| Be able to stop it | A tested way to pause an agent or drop it to depth 1 | Results of stop tests and real pauses |
| Hand over properly | Agent responsibilities included in role handovers | Handover note listing agents, open issues and limits |
Notice that nearly every row produces a document created at the time. That is deliberate. Reasonable steps are assessed after something has gone wrong, and a reconstructed story is weaker than a record made when the decision was taken.
What will a regulator ask a senior manager?
The questions are personal and specific.
- When did you first learn this agent could send customer communications without approval?
- What information did you receive about it in the three reports before the breach, and what did you do with it?
- On what evidence did you approve raising its authority?
- Who told you the approvers were reviewing its work properly, and how did you check?
- When the complaint volume rose, what did you change?
- Could you have stopped the agent, and how long would it have taken?
A senior manager who can answer each one with a dated document is in a strong position. One who answers "the technology team handled that" is not, because delegation of a task does not delegate the responsibility.
What does the standard not cover?
It does not set a required control list. The table above is a reasonable approach, not a regulatory checklist, and a smaller firm may meet the standard with less. It does not apply to the agent itself, nor to the vendor that supplies it.
It also does not protect a senior manager who took reasonable steps from all criticism; it addresses personal enforcement. The firm can still be in breach. And it sits alongside, not instead of, model risk expectations such as PRA SS1/23 for banks and data protection duties under UK GDPR. For the wider frame, see our guide to agent governance as a bound.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is approving every agent output a reasonable step?
Not on its own. If approvers click through drafts without reading them, the approval is a formality. A reasonable senior manager would check approval quality, for example by tracking how often approvers change drafts and by sampling approved items.
Does a vendor's assurance report count as a reasonable step?
It helps, but it describes the vendor's controls, not how the agent behaves on your cases with your limits. Reasonable steps include testing the agent on your own operation and reviewing the results yourself.
How often should a senior manager review their agents?
Often enough to catch drift before it harms customers. A sensible pattern is a monthly report on exceptions with a deeper review whenever an agent's certificate is renewed or its authority changes.