Connect the agent to the origination system so it can read applications, chase and classify documents, fill fields from evidence and draft the credit memo. Mark every field it writes as agent-filled with a source. Hold bureau searches until consent is on file, hold offers and decline notices for a person, and refuse pricing and decisions.
What does the connection involve, and who owns it?
Loan origination runs from application to offer: collecting information, verifying it, assessing creditworthiness, pricing and deciding. The origination system holds that workflow and its data. Agents can take most of the collecting and verifying off underwriters' desks. They should not take the decision. Creditworthiness assessment and credit scoring of natural persons are listed as high-risk in Annex III of the EU AI Act, and lenders in most markets must give specific reasons when they decline.
The head of lending operations owns the connection. Credit risk approves which fields the agent may write. Data protection approves access to bureau and personal data. The integration team builds and runs it. For the wider frame, see our guide to AI agents in banking.
Which capabilities should you grant, and at what depth?
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Read application | Reads form, documents, status | 1 | Nothing |
| Request documents | Sends checklist requests to the applicant | 3 | Requests outside the standard checklist |
| Classify and attach | Files documents against the right item | 3 | Nothing |
| Fill fields from evidence | Writes extracted values, marked agent-filled | 2 | Underwriter confirms before decision |
| Run bureau search | Prepares the search | held | Released only once consent is recorded |
| Draft credit memo | Assembles evidence and calculations | 2 | Underwriter edits and signs |
| Set price or terms | Not available | refused | Pricing owner or approved model |
| Decide | Not available | refused | Underwriter or credit committee |
| Issue offer or decline notice | Drafts the letter | 2, held | Named person sends |
Preparing applications in this system is covered step by step in letting an AI agent prepare credit applications in your origination system.
How do you connect it, step by step?
- Person, head of lending operations: approves the capability list. Evidence: approved list with depths.
- Person, credit risk: approves writable fields. Typically income, employer, address history, commitments extracted from documents. Never decision, grade, price or override fields. Evidence: field allow-list.
- Integration team: builds routes per capability and adds a field-level marker: every value the agent writes carries "agent-filled", its source document and page. Evidence: field lineage in the application.
- Integration team: gives the agent its own identity. Evidence: identity with grants matching the allow-list.
- Person, data protection: approves data access. Bureau data, identity documents, any special-category data in the file. Evidence: access approval and record of processing entry.
- Platform, on each call: checks bounds. Capability in scope, field on the allow-list, consent recorded before a bureau search, certificate in date. Evidence: verdict per call.
- Person, underwriter: confirms agent-filled fields before decision. Evidence: confirmation per field or per section.
- Integration team: wires revocation per capability and for the connection. A lapsed certificate drops the agent to read automatically. Evidence: revocation events.
Which controls matter, and what is recorded?
Field lineage. Anyone reviewing the file can tell which values a person entered, which the applicant declared, and which the agent extracted, and from where. Evidence: lineage on every field.
No decision fields. The agent cannot write grade, decision, price or override fields, even by mistake. Evidence: refused verdicts on attempts.
Consent before search. A credit search leaves a mark on the applicant's file that the lender cannot remove. Evidence: consent record linked to each search.
Pricing outside the agent. Price is set by your pricing policy or an approved model under model risk governance, not by an agent's reasoning. The question is answered directly in whether an AI agent can price a loan. Evidence: refused verdicts.
Affordability prepared, not decided. The agent assembles the evidence; the underwriter decides. See whether an AI agent can check affordability for a loan. Evidence: memo signed by the underwriter.
Applicant messages from templates. Document requests and status updates use approved wording, so the agent cannot hint at an outcome before a decision exists. Evidence: template version on each message.
Full log. Every read, write and draft with agent identity, field, source and verdict. Evidence: the record, retained for the life of the loan.
What should you measure?
- Time from application to complete file.
- Agent-filled fields changed by underwriters, by field. A field often corrected needs better extraction or a different source.
- Document requests sent and returned.
- Bureau searches held for consent, which shows where consent capture is failing upstream.
- Attempts to write disallowed fields. Should be rare and each reviewed.
Where does it usually go wrong?
Agent-filled values look like verified ones. Without lineage, an extracted income figure sits in the same field as a verified one and nobody can tell them apart.
The memo becomes the decision. An agent's memo that ends "recommend approve" is read as approval. Keep memos to evidence and calculations; let the underwriter write the recommendation.
The agent nudges the outcome. It requests extra documents only from some applicants, or fills fields more generously for some. Monitor request rates and corrections across applicant groups.
Bureau searches run on old consent. Consent for one application is reused for another.
Revocation is untested. A certificate lapses and the agent keeps writing fields because the integration ignores the platform's signal.
Questions people ask
Can the agent pre-fill the application from open banking data?
Yes, with the applicant's consent, and those values can be marked signed rather than declared because they come directly from the account provider. The underwriter still confirms them.
Should the agent see the applicant's credit file?
Only the parts the memo needs, and only after consent and a search the underwriter has released. Most capabilities can work from a summary.
Can the agent withdraw an application?
It can draft a withdrawal when the applicant asks. A person confirms it, because a withdrawn application may affect what the applicant is told and what is reported.