Work · Credit · Underwriting
$5,000,000 In loans funded automatically, inside the rules the board approved
A new lender needed to originate at scale without an underwriting team to match. We built the automated underwriting model that funded $5M, inside the framework its risk owners approved.
Our board approved a credit box, not an AI project. The model runs the box, so my credit team works the files that need judgment and calls the broker on deals that are close instead of declining them by template. Once a quarter we sit down with the decision data and tighten or loosen deliberately. That conversation didn't exist before. We were too busy processing.
$5M
Funded inside the board-approved credit box
3
Outcomes per file, no gray zone
100%
Of decisions replayable
01The challenge
A new lender needed volume without an underwriting team to match. Manual review capped throughput and burned senior attention on files that were never fundable. The risk appetite framework was non-negotiable.
02The system
Intake and extraction
Applications and bank statements go in, structured data comes out. OCR pulls the figures straight from the source documents. Scoring runs on parsed numbers, not stated ones.
Cash-flow scoring
Every file is scored against the framework’s thresholds: capacity, stability, concentration, exposure. Concentration caps check live portfolio exposure, and a file that would breach a limit cannot fund, whatever its score.
Three outcomes
Fund, decline with stated reasons, or route to a human with the file already assembled. Borderline files go to a review queue with the full analysis attached. The system never guesses on a close call.
Controls built in
Thresholds live in configuration the credit team edits directly, no developer in the loop. Every decision writes a replayable record: inputs, rule path, model version, outcome. Declines are logged the same way as approvals.
Change control
Every threshold change is versioned with who made it, when, and why. Versions lock once a decision references them, so any past decision replays against the exact setup that made it.
Reviewed each quarter
Once a quarter the credit team reviews decision patterns against the framework, using the system’s own data. Tightening or loosening is a deliberate, signed change. Nothing drifts on its own.
03The impact
$5M
$5M in loans went out with no manual underwriting on the approved files. That is origination volume without the review headcount it normally takes.
Every file
Any decision can be pulled up later and replayed step by step: inputs, rule path, model version, outcome, declines included. When the board or an auditor asks why a file went the way it did, the answer is on record, not in anyone’s memory.
Same policy
The credit policy the board approved is exactly the policy the system runs. The growth arrived with the risk profile the board already signed off on, not a quieter, looser one.
The build, in detail Expand
The policy, translated
The engagement began with the credit policy document, not with data. Every element of the approved framework maps to a control in the system, and the mapping itself was reviewed and signed by the credit owners before the first file ran.
| Policy element | How the system enforces it |
|---|---|
| Concentration limits | Hard caps checked on every file against live portfolio exposure. A file that would breach a limit cannot be funded, regardless of score. |
| Cash-flow floors | Computed from parsed bank statements, not stated figures. Thresholds live in configuration the credit team edits directly. |
| Exception handling | Files inside the gray band route to a human queue with the full analysis attached. The system never guesses on a borderline. |
| Change control | Every threshold change is versioned with who, when, and why. Decisions replay against the exact configuration that made them. |
Governance in practice
Model and configuration versions are immutable once a decision references them. Quarterly, the credit team reviews decision distributions against the framework and adjusts deliberately, a meeting that runs on the system's own data. The audit posture is simple: any decision, any time, replayable step by step.
I sign the quarterly threshold review the way I used to sign credit memos. Same accountability, except now I can see every file the change would have flipped before I approve it.
Next step
Have a book to build, or a queue that will not scale?
Sixty seconds to scope, then a 30-minute call. Bring the workflow, we bring the business case and the controls plan. If it does not hold up, we say so.
More closed work