Yes, it can do the assessment work. An AI agent can verify income, categorise spending from bank statements, apply the lender's affordability method and show the result with every input. It must stop where judgement begins: interpreting irregular income, deciding which costs a borrower can cut, weighing conflicting evidence and making the lending decision stay with a person.
What does an affordability check involve?
An affordability check asks whether a borrower can repay without hardship, not just whether they are likely to repay. It compares reliable income with committed and essential spending, then tests whether the new repayment fits with room to spare. The concept is set out in what an affordability assessment is.
Most of the effort goes into evidence. Payslips, tax documents and bank statements have to be read, matched against each other and turned into a monthly picture. That is detailed, repetitive work, and the steps are covered in how to check income and affordability evidence with an AI agent. An agent can do it consistently and show its working.
What may the agent do with the applicant's data?
| Step | The agent | Where it stops |
|---|---|---|
| Verifying income | Matches payslips, tax records and bank credits | Deciding how much irregular income to count |
| Categorising spending | Sorts transactions into essential, committed and discretionary | Treating discretionary spending as removable without the borrower's input |
| Identifying commitments | Lists other credit, rent and regular obligations | Judging undisclosed debts once found |
| Applying the method | Runs the lender's affordability calculation and stress test | Changing the method or its buffers |
| Checking consistency | Flags differences between declared and observed figures | Deciding which figure is right |
| Summarising | Drafts the affordability summary with inputs shown | The lending decision |
The applicant's data is used for this assessment only. Transaction categorisation can reveal sensitive matters, such as medical payments or gambling. The agent applies the lender's policy on how those are treated; it does not draw its own conclusions about the applicant's character.
A worked example
Take a lender assessing a 10,000 car loan. The applicant declares a salary and shares three months of bank statements.
The agent finds the salary credits match the payslips. It categorises spending: rent, utilities, phone, insurance, groceries, and two existing credit agreements. It also finds monthly payments to a buy now pay later provider that the applicant did not declare. The agent adds them to commitments, recalculates, and shows both versions: with declared commitments only, the loan fits comfortably; with the undeclared payments, it fits only just within the lender's buffer.
The agent does not decide whether the undeclared payments are a concern or an oversight. It flags the difference, with the transactions, and holds the case. The underwriter asks the applicant, who explains the plans end next month, and approves with a lower amount.
What controls make it safe?
- The lender's method, not the agent's. The agent applies the approved calculation, buffers and stress test. It cannot change them.
- Every figure traceable. Each income and spending figure in the summary links to the documents and transactions behind it.
- Declared versus observed shown side by side. Differences are flagged, never silently resolved.
- A checklist for completeness. Every assessment covers the same items; the affordability assessment checklist gives a structure.
- Credit checks under policy. The agent runs only the searches the lender's policy requires; see whether an AI agent can run a credit check.
The EU AI Act lists creditworthiness assessment of natural persons as high-risk, which brings duties on data quality, logging and human oversight for lenders using AI in this step. The wider limits for agents are set out in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, each fact in an affordability summary carries its source class through the Pryme Context Protocol. A salary credit read from bank data is marked differently from an income figure the applicant declared, so the underwriter sees at a glance which numbers were verified and which were stated. The lending decision itself sits outside the agent's bounds and is held for a person.
Questions people ask
Can the agent decide that discretionary spending could be cut?
No. Whether a borrower can reduce spending is a question for the borrower and the underwriter. The agent can show the discretionary total; it should not assume it will disappear.
How should the agent handle self-employed income?
It can gather accounts, tax records and bank credits and show the trend. Deciding which years to rely on, and how to treat a recent drop, is a judgement for the underwriter.
Can the agent use open banking data instead of statements?
Yes, if the applicant consents and the lender's process allows it. The data is the same; the source is more reliable than an uploaded statement because it comes directly from the bank.