How to prepare loan write-off recommendations with an AI agent

A write-off recommendation is evidence work, and the decision belongs to people. The agent finds loans that meet the write-off policy, assembles the recovery history, tests for open complaints, fraud markers and uncalled guarantees, and drafts a recommendation with a reason code. A person with delegated authority approves, and finance posts the write-off.

What is a write-off recommendation, and who owns it?

A write-off removes a loan, or part of one, from the balance sheet because the bank no longer expects to recover it. It does not always mean the debt is forgiven: many banks keep pursuing written-off balances or sell them. The recommendation is the document that shows the policy is met, recovery efforts have run their course, and nothing should stop the write-off.

Recoveries or collections usually prepare recommendations. Approval follows a delegated authority matrix, with larger amounts going higher. Finance posts the entry against existing provisions. Whether an agent can write off a loan itself is a clear no; the work here is the preparation underneath.

What are the steps?

1. Agent, depth 1: identifies candidates. Loans meeting the write-off policy: time in default, recovery exhausted, insolvency concluded, deceased borrower with no estate.
2. Agent, depth 1: assembles the recovery history. Contacts made, plans offered and broken, legal action, security realised, insolvency dividends received, settlement offers made.
3. Agent, depth 1: tests for blockers. Open complaint, dispute, vulnerability marker, fraud marker, guarantor not yet pursued, collateral not yet realised, a pending sale to a debt purchaser.
4. Agent, depth 1: states the amounts. Principal, interest and fees from the loan system, and the provision already held, from finance's records.
5. Agent, depth 2: drafts the recommendation. One per loan, or one batch with each loan listed and any exceptions on top, each with a reason code from the policy.
6. Person: approves. The authority holder for that amount band approves, rejects or asks for more work.
7. Person, finance: posts the write-off. The entry is depth 4 work: it changes the ledger and the customer's position. No agent holds it here.
8. Agent, depth 1: records what happens next. Whether recovery continues, the balance is sold, or it is closed, and that credit reporting matches.

Which controls apply, and what evidence do they leave?

Blocker tests on file. Each test is filed with its result and the data it read. An approver can see that the complaint check ran and what it found.

Authority matrix in the bounds. The agent routes each recommendation to the right authority holder by amount. Routing errors are refused, not left to the approver to notice.

Batches show exceptions first. A batch of many small balances is common, and approvers tend to sign the total. The draft puts any loan with a partial blocker or unusual history at the top.

Amounts tied to finance's records. The balance in the loan system and the provision held by finance should explain each other. Where they do not, for example because interest kept accruing after the loan defaulted, the draft states the difference and finance resolves it before approval. Evidence: both figures and the resolution on the filed recommendation.

Separation from posting. The agent's role has no ledger write access. Any attempt is refused and filed, with the role scope named as the deciding bound.

Fraud markers stop the path. A loan with a fraud marker is not a credit loss until a person has decided it is. Under the FATF Recommendations, suspicious activity has to be reported whatever happens to the balance, and writing off a loan does not close that question.

What should you measure?

• Recommendations returned by approvers, with reasons.
• Blockers found after approval. Each one means a test missed something; add the case to the certification set.
• Recoveries on written-off balances. Large recoveries may mean write-off came too early.
• Time from eligibility to decision. Loans sitting eligible but unreviewed distort provisions.
• Loans written off as credit loss later reclassified as fraud.

Where does it usually go wrong?

Guarantees left uncalled. The borrower's company is insolvent, the loan is written off, and the director's personal guarantee is never pursued. The blocker list must include every guarantee and every piece of security on the facility.

Write-off read as forgiveness. A letter that says "we have written off your debt" tells the customer they owe nothing. If recovery continues, communications must say so. Agents drafting customer letters need the exact status.

Small balances waved through. Batches of low-value loans include vulnerable customers who were never offered a plan. The vulnerability check runs on every loan, whatever the amount.

Settlement offers and write-offs out of step. Collections offers a discounted settlement while recoveries prepares a write-off. The agent should check for open offers and settlement quotes issued before drafting.

Recovery history with gaps. If the history shows no contact for many months, approvers should ask whether recovery was in fact exhausted. The agent states the gaps; it does not explain them away.

For non-loan receivables, the parallel question is whether an agent can write off a bad debt. The general limits are in AI agents in banking.