Can an AI agent write off a loan?

No. Writing off a loan is a financial and credit decision taken under delegated authority, with effects on the bank's accounts, its provisions and, often, the customer. An AI agent can find loans that meet the write-off policy, check every criterion, calculate the amounts and prepare the recommendation. A named person with the right authority level approves each one.

What does a write-off actually do?

A write-off removes a loan, or part of one, from the bank's books because recovery is no longer expected. The loss is recognised, provisions are used, and the account usually moves to a recovery or closed status. Banks set write-off policies with criteria, such as time since default, recovery attempts made, and the value of any security, and with authority levels: larger amounts need more senior approval.

A write-off is not the same as forgiving the debt. In many cases the customer still legally owes the money, and recovery may continue or the debt may be sold. Customers and staff often confuse the two. An agent that tells a customer "your loan has been written off" without explaining this can create a false belief that the debt is gone, with harm when collection resumes.

What conditions must be met before a write-off is even proposed?

• the loan meets the policy's criteria for age, status and recovery attempts;
• any security has been realised or assessed as having no recoverable value;
• the balance is reconciled, including interest and fees that should stop accruing;
• there is no open complaint, dispute or legal action that a write-off would affect;
• the case is checked for fraud indicators, because a loan obtained by fraud is handled differently;
• the right approver for the amount is identified from the authority table.

The agent checks each of these and records the evidence. The full process is set out in preparing loan write-off recommendations with an AI agent.

What may the agent do, and what stays with a person?

Where a suspicion of financial crime has been reported, the customer letter needs particular care, because careless wording could amount to tipping off.

An example: a batch with one exception

Take a lender's quarterly write-off review. The agent finds 120 unsecured personal loans that meet the policy: over a year in default, three recovery attempts recorded, no security. Total balance: 360,000.

It checks each. 112 are clean. Five have open complaints logged since the last recovery attempt; the agent removes them from the batch. Two have recent payments, small but regular, which suggests recovery is not hopeless; it flags them for a person to decide. One has a fraud indicator: the applicant's identity was later reported stolen. The agent sends that to the fraud team, because it may not be the named borrower's debt at all.

The 112 go to the credit risk manager whose authority covers the batch total. The manager approves, the agent posts the write-offs, and drafts the customer letters explaining that recovery may continue. The record shows the batch, the exclusions and reasons, the approval and the postings.

What controls make it safe?

• The authority table as a bound. The agent routes each write-off to the approver the table requires and cannot route lower.
• No approval capability. The agent posts write-offs only against a recorded approval.
• Exclusion reasons recorded. Every loan removed from a batch carries its reason.
• Clear customer wording. Letters are drafted from approved templates that explain the debt may still be owed.
• A written policy for the agent. What the agent may do in write-offs is set out like any other control; see how to write an AI agent policy.

Writing off trade debts in a company's accounts is a related but different task; see whether an AI agent can write off a bad debt. At the other end of the loan's life sits loan approval, which is also a person's decision. The general limits are in AI agents in banking: what they may and may not do.