How to let an AI agent set up payment plans in your collections system

Plan records are agent work only inside the policy you have already approved. At depth 3 it creates the plan, sets the schedule and pauses collection contact; any concession on interest or fees, any plan outside the limits, and any decision to treat a plan as broken is held for a person. Each grant can be revoked.

What does setting up a plan involve, and who owns it?

Agreeing a plan is a conversation with the customer. Setting it up is the system work that follows: a plan record with amounts and dates, a change to the account's collection status, a pause on letters and calls, and often a payment mandate. If the system work is wrong, the customer who kept their side of the agreement still gets an arrears letter the next morning.

Collections operations owns the process. Credit risk owns the plan policy: the shortest and longest terms, the minimum payment, when an income and expenditure assessment is required. The conversation side is covered in agreeing repayment plans within set limits. This piece covers what the agent may write in the collections system, and the connection itself is covered in connecting an agent to your collections system.

Which capabilities do you grant?

In order:

1. Agent, depth 1: reads the account and any flags for vulnerability, dispute or legal status. A dispute or legal flag stops the flow and routes to a person.
2. Agent, depth 1: checks the agreed terms against policy. Term, instalment, start date, and whether an income and expenditure assessment is on file and current enough.
3. Agent, depth 3: creates the plan record with the agreed schedule.
4. Agent, depth 3: changes the collection status so the contact strategy pauses arrears letters and calls for as long as the plan holds.
5. Agent, depth 3: records the payment mandate where the customer has given the instruction through an authenticated channel. Taking a payment there and then is depth 4 and a separate grant.
6. Agent, depth 2: drafts concessions. Freezing interest or waiving fees changes what the customer owes. A person grants it.
7. Person: decides a plan is broken. The agent drafts the assessment when a payment is missed; a person decides what follows.

Why is a broken plan held for a person?

Breaking a plan restarts collection activity and can lead to default notices, credit file reporting and, later, legal action. Several of those steps cannot be undone, which is why irreversibility in collections sits at the centre of the design.

A missed instalment has many causes: a bank holiday, a changed pay date, a mandate that failed for a technical reason, a customer whose circumstances have worsened. The agent can gather these facts and say which apply. Whether the plan is broken, renegotiated or extended is a judgement that needs a person who can look at the whole account.

How do you revoke access, and what is held meanwhile?

• Remove the create-plan grant. The agent can still read and draft; people create plans from its drafts.
• Revoke the credential. The collections system role used by the agent is disabled; nothing it held is lost, because drafts sit as held verdicts in the record.
• Let the certificate lapse or fail. After 90 days without renewal, or the same day as a run that clears no depth, the agent drops to depth 1; a short run keeps only the highest depth it cleared.

Plans already created stay in force. Revocation stops new writes; it does not unwind agreements made with customers.

What does the audit trail show?

Each plan creation is filed with the agreed terms, the policy rule that allowed them, the reference to the customer's agreement (call recording, chat transcript or signed form), the income and expenditure reference if required, and the verdict. Status changes record the contact strategy the account moved to, so you can show that letters stopped when they should. The collections contact strategy matrix is the reference for which status suppresses which contact.

Held drafts record who granted or rejected the concession and why. Refused actions, such as an attempt to create a plan on a disputed account, record the bound that decided.

What should you measure?

• Time from agreement to plan live in the system. Any gap is a window for a wrong arrears letter.
• Arrears contact sent after a plan was agreed. The clearest failure signal; it should be zero, and each case is reviewed.
• Plans held for policy exceptions, and their outcomes.
• Plans kept to term versus broken, split by whether the agent or a person set them up.
• Mandates that fail on first collection.

Where does it usually go wrong?

Status not changed. The plan is created but the account stays in the arrears queue, and the contact strategy carries on. Make plan creation and status change one action, filed together.

Policy in a document, not in the bounds. If the maximum term lives in a manual and not in the agent's bounds, the agent cannot be refused for exceeding it. Write the limits where the minimum chain can test them.

Mandate confusion. Recording a customer's direct debit instruction is different from taking a payment. Whether an agent can set up a direct debit or standing order depends on the authority the customer gave; keep it a separate grant.

Vulnerability flags overwritten. A plan status must never clear a vulnerability marker. Treat those flags as read-only to the agent.

General limits for bank agents are in AI agents in banking.