AI underwriting with human gates: where models help and where people decide
AI underwriting with human gates: where models help and where people decide
Credit underwriting is mostly reading. Bank statements, tax returns, contracts, invoices, title documents, loan tapes. An analyst spends days pulling numbers out of documents, checking them against each other, and deciding whether a borrower fits the credit box.
AI is very good at the reading. It's not yet trustworthy enough to make the decision on its own. The useful design puts AI where it's strong and puts people at specific gates where judgment and accountability are needed.
Where AI helps
Extraction. Pulling structured data out of unstructured documents: account balances from statements, payment terms from contracts, line items from invoices. Work that takes an analyst hours takes a model seconds.
Reconciliation. Comparing data across sources. Does the revenue on the tax return match the deposits in the bank statements? Does the loan tape match the servicer's records? Models can check every field, not a sample.
Anomaly detection. Flagging what doesn't look right: a statement with inconsistent fonts, a payment pattern that changes abruptly, an invoice from a customer who doesn't appear anywhere else.
Rule testing. Checking each application against every criterion in the credit box and listing which pass and which fail.
Monitoring. Re-running all of the above every time new data arrives, instead of once at origination.
Where people decide
Exceptions. Any application that fails a rule, or that a model flags as unusual, goes to a person. The model's job is to explain what it found. The person decides.
Judgment calls. Some credit questions don't have a data answer. Is this management team credible? Is this industry about to change? Is this explanation for a gap in payments reasonable?
Final approval. A named person signs off on every funding decision. That's where accountability sits, and it's what a regulator, an investor, or a court will look for.
Model oversight. People review what the models are getting wrong, how often, and whether the rules still make sense.
Why the gates matter
AI models fail in particular ways. They can read a document wrong with full confidence. They can be fooled by documents designed to fool them. They can drift as the data they see changes. And they can't explain their reasoning the way a person can be asked to.
None of that makes them useless. It makes them a first pass, not a final answer. A model that checks 100% of fields and flags 3% for review catches far more than an analyst who samples 10% of fields by hand. But the 3% still needs a person.
There's also a regulatory reason. Lending decisions in most jurisdictions have to be explainable to the borrower and defensible to the regulator. "The model said no" isn't an explanation.
A practical structure
- Documents come in. AI extracts and structures the data.
- AI reconciles the data against primary sources and against itself.
- AI tests every rule in the credit box and flags anomalies.
- Clean applications go to an analyst for confirmation. Flagged ones go to an analyst for investigation.
- A credit officer approves or declines, with the full record attached.
- After funding, AI monitors continuously. Any change outside set limits goes back to a person.
The result is faster than all-human underwriting and safer than all-machine underwriting. Every decision has a data trail and a named owner.