How to let an AI agent prepare credit applications in your origination system

An origination agent should build applications, never decide them. In the origination system it may read, create the application record, populate fields from documents, attach evidence and submit to the underwriting queue, mostly at depth 3. Decision fields, pricing, limits and overrides stay outside its role. Every figure it enters carries its source.

What does application preparation involve, and who owns it?

Preparing a credit application means turning a borrower's request and documents into a complete, consistent file an underwriter can assess. Income, outgoings, existing debts, purpose, security and identity all have to be captured in the right fields, with evidence attached. Loan origination covers the whole path from enquiry to funding; preparation is the stretch before underwriting.

The lending operations or sales support team usually owns preparation. Underwriting owns the decision. Credit risk owns the policy and the system's decision rules. That split is the one to mirror in the agent's grants. How to wire the connection is covered in connecting an agent to your loan origination system.

Which permissions do you grant, step by step?

1. Agent, depth 1: reads the request. Application form, uploaded documents, existing relationship data.
2. Agent, depth 3: creates the application record. A new record in draft status, linked to the customer.
3. Agent, depth 3: populates fields from evidence. Income from payslips or accounts, commitments from statements, security details from valuations. Each field is stamped with the document and page it came from.
4. Agent, depth 1: runs consistency checks. Income on the form against income in the documents, declared debts against statement debits, dates against expiry.
5. Agent, depth 2: drafts the credit memo. A summary in the bank's format, for the underwriter to rely on or rewrite.
6. Agent, depth 3: submits to the underwriting queue. Only when the bank's completeness rules pass. Incomplete files go back to the preparer with a list.
7. Person: decides. The underwriter approves, declines, prices, conditions or refers. The agent has no write access to these fields.

Where exactly is the line on the decision?

Write access is granted field by field. The agent's role can write to applicant, income, commitment, security and document fields while the record is in draft. It cannot write to decision, pricing, limit, exception or override fields at any status. Once the file is submitted, its own fields lock too; later changes come through a person.

There is a less obvious line. The agent must not withdraw or park an application because it looks weak. Quietly dropping an application is a decline by another name, and a decline is a credit decision. If the file is incomplete, it goes back to a person with the gap list. The wider reasoning is set out in AI agents and credit decisions: where the line sits.

Pulling a credit bureau report is its own capability. It needs recorded consent and a stated purpose, and it should be granted separately at depth 3, not bundled into "prepare".

What is recorded?

• Field-level provenance. Every populated field shows its document, page and provenance class. A payslip the applicant uploaded is declared; a bureau return is external; your own account data is sealed.
• Every write as a verdict. The field, old value, new value, the bounds checked and the bound that decided.
• Refused writes. Any attempt to touch a decision field is refused and filed. A pattern of refusals is a sign the brief is unclear.
• Submission record. The completeness rules that passed, the time and the queue.

For consumer credit, this record also supports the obligations that follow the decision. Creditworthiness assessment of natural persons is listed as high-risk in Annex III of the EU AI Act, and in the US, ECOA and Regulation B require specific reasons when credit is denied. Both are easier to meet when you can show what the underwriter saw and where each figure came from.

What should you measure?

• Time from complete documents to underwriter-ready.
• Files sent back by underwriters, with reasons. The best single signal of preparation quality.
• Fields corrected by underwriters. Track by field type; income fields usually tell you most.
• Mismatches the agent raised that underwriters judged material.
• Applications waiting in draft beyond the service standard. A sign files are stalling rather than being sent back.

Where does it usually go wrong?

Filling gaps with guesses. A blurred payslip, and the agent estimates. It must leave the field empty and flag it. An estimated income field looks identical to a verified one in most origination systems.

One service account, all permissions. Integrations often use a single account with full rights. If the agent's credential can write decision fields, the field-level grant is only a convention. Use a role that cannot.

Memo drift. Over time a drafted memo can start to read like a recommendation to approve. Keep the memo descriptive, and use the bank's template, such as the credit memo template for agent-prepared applications.

Stale evidence. Documents that were current at upload may have expired by submission. The completeness check should test dates at submission time.

Banking-wide limits on agents are set out in AI agents in banking.