Any decision that a named person answers for, that has legal or similarly significant effect on a customer, or that accepts a risk on the firm's behalf. That covers filing a suspicious activity report, exiting a customer, deciding whether to notify a regulator, approving policy, granting waivers and signing attestations. An agent can prepare every one of them.
What puts a decision on the person's side of the line?
Three tests, and a decision that meets any one of them stays human.
Personal accountability. Some decisions are assigned by law or regulation to an individual: the money laundering reporting officer who decides whether to report, or, in the UK, a senior manager under the Senior Managers and Certification Regime. You cannot delegate that accountability to software, so you should not delegate the decision.
Effect on a person. Exiting a customer or refusing service changes someone's position. Under GDPR, decisions based solely on automated processing with legal or similarly significant effects are restricted, and fairness expectations from conduct regulators point the same way.
Judgement under uncertainty. Accepting a residual risk, granting an exception or interpreting an ambiguous rule means choosing what the firm is willing to live with. That choice belongs to people who carry the consequences.
Which decisions does that cover?
| Decision | What the agent prepares | Why it stays with a person |
|---|---|---|
| Filing a suspicious activity report | The case file and a draft narrative | Named accountability; the report cannot be recalled |
| Exiting a customer | The evidence pack and options | Significant effect on the customer |
| Notifying a regulator of a breach | The timeline, facts and draft notice | Judgement on materiality, firm-level consequences |
| Deciding a data breach must be reported | The facts and the assessment questions | Legal judgement under time pressure |
| Approving or changing policy | The redline and impact summary | Sets what every agent and person must follow |
| Granting a waiver or exception | The request and precedents | Accepts a risk on the firm's behalf |
| Signing an attestation | The evidence and a draft | A personal statement of fact |
| Overriding a control flag | The reasons for and against | Removes a safeguard |
The customer exit process is set out in how to prepare customer exit decisions with an AI agent, and the breach question in whether an AI agent can decide a data breach must be reported. Finance has a parallel list in which finance tasks should always stay with a person.
What can an agent decide on its own?
Routine outcomes against clear rules, where a mistake is caught and fixed cheaply. Closing an obvious screening false positive under a documented rule, chasing a missing document, scheduling a periodic review, or marking a training record complete can all run at depth 3 once the agent is certified for them on your own cases. The general position is in the guide to AI compliance agents.
The line moves with evidence, not with confidence. A capability crosses to the agent's side when your own cases show it performs reliably and the outcome is easy to reverse, never because a decision feels routine.
What does this look like in practice?
Take a consumer lender's compliance team. Its agent reviews complaints for signs of a systemic issue and finds forty cases over a quarter where the same fee was applied after a payment holiday ended. It drafts a summary, links each case, sets out the policy wording and notes which accounts it could not check because records were archived.
The head of compliance decides whether it is a breach, whether it must be reported, and what remediation customers receive. The agent then prepares the remediation file and the draft notice. Every step of the agent's work is filed; every decision carries a person's name.
What does the person need from the agent to decide well?
Two things are easy to miss. First, a record of what the agent looked for and did not find, because a clean summary with silent gaps invites a wrong decision; see why compliance agents must record what they did not find. Second, the exact policy version the agent applied, which is only possible if policy libraries are versioned before agents use them.
Questions people ask
Can a person approve an agent's decision without reading it?
They can, and that is the risk. If approvals are rubber-stamped, the decision has moved to the agent in practice. Measure review time and sample approved decisions.
Does this list differ by country?
The named roles and legal tests differ, but the three tests travel well. Map each decision to the person your regulator would hold accountable for it.
What if the person disagrees with the agent?
The person's decision stands and is filed with their reason. Patterns of disagreement are useful evidence for retesting the agent.