PrymeIntelligence

Banking and fintech·Playbook

How to process personal loan applications with an AI agent

Process personal loan applications with an AI agent that verifies income and builds the affordability file while an engine or underwriter decides.

Decision-ready, not decided: that is the agent's job on a personal loan. It checks identity and fraud signals, runs the bureau search with recorded consent, verifies income and outgoings from evidence, and sets out affordability using your policy. Your approved decision engine or an underwriter decides. Declines and their reasons, and any release of funds, stay with people.

What does processing a personal loan involve, and who owns it?

A personal loan application asks for an amount over a term for a purpose. The lender must decide whether the applicant is who they say, whether the application is genuine, whether the loan is affordable without causing hardship, and whether it fits the lender's appetite. Creditworthiness assessment and credit scoring of natural persons are listed as high-risk uses in Annex III of the EU AI Act, and many markets have their own conduct rules on affordability.

Lending operations owns the flow. Credit risk owns the scorecard, the decision rules and the affordability policy. Underwriters handle referrals. Many lenders already decide most applications through an approved decision engine. The agent does not replace that engine or override it. It prepares what goes in and handles what comes out to a person.

What are the steps?

  1. Agent, depth 1: reads the application. Amount, term, purpose, declared income and outgoings, employment.
  2. Agent, depth 1: checks identity and fraud signals. For existing customers, it reuses verified identity where policy allows. For new applicants, it runs the checks used in processing personal account opening applications.
  3. Agent, depth 3: runs the bureau search. Only with the applicant's consent recorded and the purpose stated. This is its own certified capability.
  4. Agent, depth 1: verifies income and outgoings. From payslips, bank statements or account data the applicant has shared. Each figure carries its source and provenance class.
  5. Agent, depth 1: applies the affordability policy. It calculates disposable income using the lender's rules and flags any figure it could not verify.
  6. Agent, depth 1: passes the file to the decision engine or, for referrals, to an underwriter, with a summary drafted at depth 2.
  7. Person: decides referrals and every decline. An underwriter decides referrals. Any decline, including one proposed by the engine, is confirmed by a named person and its reasons checked.
  8. Agent, depth 3: prepares the agreement for the applicant to sign once the loan is approved.
  9. Person: releases funds. Disbursement is depth 4 work.
Step The agent does Depth Stays with a person
Read Reads application 1 None
Identity and fraud Runs or reuses checks 1 Fraud referral
Bureau Searches with consent 3 Consent disputes
Income and outgoings Verifies with sources 1 Alternative evidence
Affordability Applies policy 1 Policy exceptions
Hand-off Passes file and summary 1, 2 Referral decisions
Decline Drafts reasons 2 Confirming every decline
Agreement Prepares documents 3 Changed terms
Funds Nothing None Release

Why do declines stay with a person?

A decline affects someone's access to credit and may be reported to bureaux as a search. Where a decision with legal or similarly significant effects is made solely by automated means, GDPR Article 22 applies. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used. A person confirming each decline, and checking that the reasons given are the real ones, keeps the lender on safe ground in both places.

It is also where errors hide. A misread payslip or a miscategorised outgoing can turn an affordable loan into a decline. The person who confirms the decline should see the figures that drove it.

Which controls apply, and what evidence do they leave?

Consent before the bureau search. Evidence: the consent record and the purpose, filed with the search.

Provenance on income. Income verified from your own account data is sealed. A payslip the applicant uploads is declared. The affordability workings show which class each figure has.

Lawful basis per source. The record shows the basis for each data source, following what lawful basis an agent needs to process personal data.

No approval capability. The agent has no write access to decision fields. Whether an agent can approve a loan is answered by its role: it cannot.

Decline reasons checked. The draft reasons, the person who confirmed them and any edits are filed.

What should you measure?

  • Time from application to decision-ready.
  • Referrals where the underwriter changed an agent figure. Especially income.
  • Declines overturned on appeal or review.
  • Early arrears on approved loans, by whether the file was agent-prepared.
  • Bureau searches without consent on file. Should be none.

Where does it usually go wrong?

Upselling. An agent that suggests a larger loan "since you qualify" is serving the lender, not the borrower. The agent quotes the amount requested and nothing more.

Missed hardship signals. Bank data may show gambling spend, payday borrowing or missed bills. The agent flags these for the underwriter. It does not decline on them, and it does not ignore them.

Proxies for protected characteristics. Some variables track age, ethnicity or disability. The agent uses the lender's approved factors only; it does not add its own.

Joint applications flattened. Two incomes are combined, but only one applicant's commitments are counted. The agent should verify and list each applicant's figures separately before showing any total, and both applicants' consent must be on file before either bureau search.

Purpose taken at face value. "Debt consolidation" may mean borrowing to repay a lender in arrears. The agent should say what the statements show.

The credit memo template for agent-prepared applications suits referred files that need a fuller write-up. General bounds are in AI agents in banking.

Questions people ask

Can the agent tell an applicant they have been approved?

Once the decision engine or underwriter has approved the loan, yes, using an approved message. It cannot predict an outcome before then.

What if the applicant disagrees with a figure the agent used?

The agent records the disagreement and any new evidence and passes it to an underwriter. It does not change verified figures on the applicant's say-so.

Does the agent need a separate certificate for secured loans?

Yes. Secured lending adds valuation, title and security checks, which are a different capability with different cases.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.