PrymeIntelligence

Compliance·Playbook

How to prepare customer exit decisions with an AI agent

Prepare customer exit decisions with an AI agent that assembles the case, checks fairness and drafts notices, while a named person decides each exit.

Exiting a customer is hard to reverse and can cut someone off from banking, so every exit should be decided by a named person on a complete file. The agent builds that file: reasons, evidence, history, vulnerability, notice period and consequences. It checks the case against your exit policy and drafts the notice. The decision-maker decides.

What is an exit decision, and who owns it?

Firms end customer relationships for financial crime risk, persistent breach of terms, fraud, conduct towards staff or commercial reasons. In many markets this is regulated: notice periods, explanations, treatment of vulnerable customers and, for basic accounts, specific restrictions. Supervisors look closely at exits for fairness and consistency, and at whether firms are exiting whole groups of customers rather than assessing risk case by case.

The decision usually sits with a senior person under a delegated authority: the MLRO or a delegate for financial crime exits, a senior manager for other reasons. Front-line teams and investigators recommend. Which compliance decisions must stay with a person lists exits among them, for good reason.

What does the agent prepare, and what does a person approve?

  1. Person, recommender: proposes the exit with the reason, from an investigation, review or operational team.
  2. Agent, depth 1: assembles the file. The recommendation, the evidence behind it, the customer's products and balances, tenure, complaint and contact history, prior reviews and any vulnerability markers.
  3. Agent, depth 1: checks the policy. Is the reason one your exit policy allows? Is the evidence what the policy requires for that reason? Are there product-specific restrictions or required notice periods?
  4. Agent, depth 1: checks consistency. Similar cases in the past and how they were decided, so the decision-maker can see whether this one is in line.
  5. Agent, depth 1: identifies consequences. What the customer loses access to, direct debits and incoming payments that will fail, balances to return, and any lending to be repaid.
  6. Agent, depth 2: drafts the decision paper and the customer notice. The notice gives the notice period, what happens next and what the customer can do, using approved wording. Where the reason cannot be disclosed, such as a financial crime concern, the notice uses approved neutral wording.
  7. Person, decision-maker: decides exit, an alternative such as restricting a product, or no action.
  8. Person, operations: sends the notice and carries out the exit on the decided timetable. The agent may schedule tasks; closure itself is a held action.
Step The agent does Depth Stays with a person
Recommend Nothing None Proposing the exit
File Gathers reasons, history, vulnerability 1 None
Policy check Tests reason and evidence 1 Interpreting edge cases
Consistency Finds similar past cases 1 Judging comparability
Consequences Lists impacts on the customer 1 None
Drafts Paper and notice 2 Approving wording
Decide Nothing None Named decision-maker
Execute Schedules tasks, closure held 2 Operations releases

Which controls apply, and what evidence does each leave?

Delegated authority checked. The decision-maker must hold authority for this kind of exit. Evidence: the authority matrix reference on the decision.

Vulnerability surfaced first. The file opens with any vulnerability markers. A customer in financial difficulty, with health problems or dependent on the account for benefits needs a different conversation and possibly a different outcome. Evidence: the vulnerability section, and the decision-maker's note on it.

Consistency shown. The decision-maker sees how similar cases were decided. Evidence: the comparison and, where this decision differs, the reason.

Tipping-off guarded. For financial crime exits, the notice must not reveal suspicion. Evidence: approved wording used, and legal or MLRO review where the wording departs from it.

Irreversibility respected. Account closure, return of funds and the end of the relationship are hard to undo; the agent does not hold that capability. The reasoning mirrors irreversibility in HR decisions, where dismissal is treated the same way.

What should you measure?

  • Exits by reason, over time and by customer group, to spot patterns that look like blanket exiting.
  • Decisions that differed from the recommendation, and why.
  • Complaints and ombudsman referrals about exits, and their outcomes.
  • Exits reversed after decision.
  • Time from recommendation to decision.

Where does it usually go wrong?

Thin evidence. A recommendation based on a single alert or an old adverse media hit. The policy check should test the evidence against the reason, not just the reason.

Consequences ignored. A customer whose wages and rent go through the account is seriously harmed by a short-notice exit. The consequences section exists to make this visible.

Group exits dressed as individual decisions. Many customers of one type exited in a short period, each "individually". Supervisors look for this. Measure exits by group.

Notices that confuse. Customers do not know what to do next. The notice should say plainly when the account closes, how to get their money, how to move their regular payments, and how to complain.

Supervisory scrutiny unprepared for. Exit decisions are a common topic in supervisory visits; preparing for a supervisory visit covers getting the file ready in advance.

The stance on compliance agents overall is in AI compliance agents.

Where Pryme fits

In Pryme, the agent preparing exits holds depth 1 on the customer file and depth 2 for the paper and notice. Closing an account is a hard-to-undo action that it does not hold, so any attempt is refused, and the verdict names the bound that decided. Every exit file, decision and notice is filed in the append-only record.

Questions people ask

Can the agent recommend an exit itself?

It can identify customers who may meet exit criteria, for a person to review. The recommendation and the decision should both come from people.

Should the agent check whether a less drastic option exists?

Yes. The decision paper should list alternatives the policy allows, such as restricting one product, so the decision-maker has a real choice.

What if the customer complains about the exit?

The complaint follows your normal process, with the exit file available to the handler. A complaint that reaches the ombudsman will rely on that file.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.