Production AI loan decisioning in six weeks: 90-second approvals for standard applications

Industry

Lending

Market

USA

Client

Consumer lender in the US

Sponsor

CTO

What we built

AI loan decisioning

The situation

The client, a US lender, had a decisioning process that depended on people for cases that did not need them. Standard applications, the ones that clearly met policy, waited in the same underwriting queue as the complex ones. Applicants waited, some went elsewhere, and underwriters spent their time confirming the obvious.

The CTO had already decided to fix it with AI and was evaluating agencies. They looked at four. Most arrived with decks.

The discovery call

We arrived with a working prototype. Before the call, we asked for a sample of the lender's historical application data and built a first version of the decisioning flow against it. In the discovery call, the CTO watched actual code running on their own data, not a slide describing what it might do. That is the point of our three-day discovery sprint: the argument about whether it can work is settled before anyone signs anything long term.

What we built

A decisioning system that handles standard applications end to end and sends everything else to underwriters with the work already done.

Extraction. Application data and supporting documents are read and structured automatically: income, employment, existing obligations, identity details, whatever the lender's policy requires.

Policy first. The lender's credit policy is encoded as explicit rules: eligibility criteria, hard declines, limits and the conditions that must send a file to a human. Policy is the lender's, written down and versioned, not buried in a model.

A risk model inside the policy. Within what policy allows, a model assesses risk on the applicant's data. Every decision records the factors behind it.

Reasons for every decline. US lending law requires specific reasons when an application is declined. Every adverse decision produces the principal reasons in plain terms, drawn from the factors the system actually used, ready for the adverse action notice.

Clear routing. Applications that fully meet policy and fall within the approved risk band are decided automatically. Anything borderline, incomplete, unusual or outside the band goes to an underwriter with the extracted data, the rule results and the model's view already assembled.

An audit trail. Every decision stores the inputs, the policy version, the rules that fired, the model output and the outcome, so compliance and model validation can reconstruct any decision.

How it went live

Discovery to production took six weeks. The system ran alongside underwriters first, with its decisions compared to theirs, before it began deciding standard cases on its own.

Results

The lender now approves standard applications in 90 seconds, with zero manual review for standard cases. Underwriters work the files that genuinely need judgement, with the groundwork already done.

What made the difference

Showing, not pitching. A working prototype on the client's data in the first meeting set the tone for the whole engagement.

Policy as code, model inside it. Keeping the lender's rules explicit is what made the system explainable to compliance and trusted by underwriters.

Designing for the decline, not just the approval. Adverse action reasons and a full audit trail are what make automated decisioning viable in a regulated lending business.