How to check income and affordability evidence with an AI agent

The agent reads each payslip, statement and tax document, extracts the figures, reconciles them against each other and against account data, and flags anything that does not match. It prepares an affordability workpaper with every figure marked as verified or declared. An underwriter reviews the flags and makes the affordability decision.

What does checking affordability evidence involve, and who owns it?

An affordability assessment asks whether a borrower can repay without hardship. Evidence checking is the part that turns a customer's documents into figures the assessment can trust: income, its stability, and the commitments already drawing on it.

This is slow, detailed work, and agents are good at the detail. They are not the right owner of the decision, a point our guide to AI agents in banking makes for credit generally. Creditworthiness assessment of natural persons is listed as high-risk in Annex III of the EU AI Act, and lenders in most markets must be able to show how each affordability conclusion was reached. The head of credit owns the policy. Underwriters own each decision. For whether an agent should ever make the call, see whether an AI agent can check affordability for a loan.

What does the agent prepare?

1. Agent, depth 1: classifies each document. Payslip, bank statement, tax return, employer letter, accounts. Evidence: document list with type and date range.
2. Agent, depth 1: stamps each source. An uploaded file is declared. Account data retrieved directly from the customer's bank with their consent is signed. Evidence: provenance class per figure.
3. Agent, depth 1: extracts income. Gross and net pay, pay dates, employer, regular and irregular elements. Evidence: extracted values with a pointer to the page and line.
4. Agent, depth 1: reconciles. Payslip net pay should match a credit on the statement on the pay date from a payer with the employer's name. The method for statements is covered in whether an agent can verify income from bank statements. Evidence: a match table.
5. Agent, depth 1: finds commitments. Loan repayments, card payments, rent, childcare, payments to other lenders and deferred payment providers visible on statements. Evidence: commitment list with sources.
6. Agent, depth 1: flags anomalies. Figures that do not reconcile, round-sum credits, a payer name that differs from the employer, document properties that suggest editing, gaps in the date range. Evidence: flag list, each with its reason.
7. Agent, depth 2: prepares the workpaper. Income, commitments, the affordability calculation under your policy, the flags, and what is verified versus declared. Evidence: the workpaper.
8. Person, underwriter: reviews flags and decides. Accept, request more evidence, or decline on affordability grounds. Evidence: decision and reasons, linked to the workpaper.

The full list of what an assessment should cover is in the affordability assessment checklist.

Which controls matter, and what evidence do they leave?

Every figure points to its source. An underwriter, auditor or complaint handler can click from the workpaper to the line on the payslip. Evidence: references in the workpaper.

Verified and declared kept apart. A figure the customer typed into the form is declared until a document or account data supports it. Evidence: class per figure.

Flags cannot be cleared by the agent. Only the underwriter clears a flag, with a reason. Evidence: flag status and clearing note.

Policy as configuration. The calculation uses your current policy parameters, versioned, not the agent's idea of reasonable spending. Evidence: policy version on the workpaper.

No decision language. The agent's workpaper says what the evidence shows, not whether to lend. Evidence: workpaper template without an approve field for the agent.

What should you measure?

• Time from documents received to workpaper ready.
• Flags per application, and the share the underwriter upheld, counted as a ratio of upheld to raised.
• Fraud found after approval in cases where the agent raised no flag. Each one is a lesson for the extraction or flag rules.
• Requests for more evidence, and whether they were needed.
• Underwriter changes to extracted figures. Frequent corrections point to an extraction problem.

Where does it usually go wrong?

Irregular income is averaged. Overtime, commission and seasonal work are smoothed into a stable-looking monthly figure. The workpaper should show the pattern and let the underwriter apply policy.

Commitments are missed. Small recurring payments to deferred payment providers or family members do not look like debt. List every regular outflow and let a person decide which count.

Declared becomes verified by repetition. The customer's stated income appears on the form, in the workpaper and in the memo, and by the third mention it reads as fact. Keep the class visible.

Document tampering slips through. Edited statements are common. The agent should flag what it can see, and the process should include a direct data route where possible.

Proxies creep in. Postcode, spending categories or names can stand in for protected characteristics. Keep the agent's work to the evidence the policy names.

For the separate step of building a customer's full budget, see how to run income and expenditure assessments with an AI agent.