Short answer
AI underwriting means software that reads borrower documents, extracts and calculates the numbers, checks the file against credit policy and drafts the write-up. In consumer lending it can also return the approve or decline decision itself. In commercial lending it prepares the analysis and the memo, and a human committee still decides.
The phrase covers two different products that get sold with the same words. One of them replaces a human judgment on a consumer application. The other does the preparation work in a commercial file and hands a credit officer something to argue with. Knowing which one a vendor is selling is most of the evaluation, and it is the thing demos are least clear about.
The five steps in a loan file, and which ones software finishes
Walk a commercial file from intake to committee and there are five distinguishable jobs. Software today finishes the first three reliably, assists on the fourth, and does not touch the fifth.
The economics follow from that split. The first three steps are where most of the calendar time in a small credit shop goes, and they are also the steps where being wrong is cheap and checkable. The credit decision is where being wrong is expensive, which is exactly why nobody serious is selling it in commercial lending.
- Collect and identify. Sort the borrower's dump of files, work out what each one is, and flag what is missing
- Extract. Pull the numbers off the statements, returns and schedules into structured fields
- Spread and calculate. Build the spread, calculate the ratios, and trace every figure back to the page it came from
- Check and draft. Test the deal against credit policy, surface the exceptions, and draft the memo narrative
- Decide. Weigh the character, the collateral and the exceptions, and take the risk
Consumer and commercial are different products
In consumer lending the file is small, the data is standardised and the volume is high, which is exactly the shape a statistical model wants. Vendors here build a model on the lender's own portfolio and return the decision with adverse-action reasons attached, and the mature ones treat fair-lending testing as part of building the model rather than as a report afterwards.
In commercial lending the file is large, the entity structure is idiosyncratic and the volume is low. There is no useful population to train a decision model on when a bank does 200 commercial deals a year across a dozen industries. So the work software does is analytical: reading an 80-page tax return, tracing a K-1 through to the guarantor, building the global cash flow, and writing the first draft of the memo.
A commercial lender evaluating consumer decisioning vendors will hear a confident story about automated approvals and then discover the product has no idea what to do with a partnership return.
| Consumer AI underwriting | Commercial AI underwriting | |
|---|---|---|
| What it produces | An approve, decline or counter decision | A spread, a policy check and a memo draft |
| Trained on | The lender's own historical portfolio | Document structure rather than credit outcomes |
| Who decides | The model, within policy | The credit officer and committee |
| Main compliance question | Fair lending and model validation | Traceability of every figure to its source |
| Typical deployment | Weeks to months on the existing LOS | Weeks to months on the existing LOS |
Traceability is the feature that decides whether it survives loan review
An AI-produced spread is only useful if someone can defend it. That means every calculated figure has to point back at the document and the page it came from, so a reviewer can check the derivation rather than take the number on trust.
This is the question to ask in every demo: show me a ratio, then show me the exact line on the exact document it was built from. Products designed around traceability answer it in two clicks. Products that treat the AI output as the deliverable answer it with a description of their accuracy rate, which is not the same thing.
The same logic applies to policy checks. A flag that says a deal is outside policy is worth little without the specific policy clause and the specific figure that triggered it.
What it does not do
It does not replace the underwriter, and vendors that imply otherwise are describing a product that will be switched off within two quarters. What it removes is transcription, reconciliation and first-draft writing, which is most of the hours and almost none of the judgment.
It also does not remove the review obligation. Anything that scores, decides or drafts is a model your risk function has to document and periodically validate, and that internal cost is real. Pick vendors that produce the documentation as a by-product of how they work, because the alternative is your compliance officer writing it from scratch.
Frequently asked questions
Can AI approve a commercial loan?
No vendor covered on this site sells that, and you should be suspicious of one that does. Commercial volumes are too low and files too idiosyncratic to train a decision model on. What AI does in a commercial file is the analysis and the memo draft, with the decision staying at committee.
Do we have to replace our loan origination system?
Not necessarily. Several products here are designed to sit on top of the LOS an institution already runs, connecting through APIs, which turns a multi-quarter migration into a deployment measured in weeks.
How accurate is AI extraction on a real loan file?
Vendors quote 98% to 99.5%, which holds on the document types their models were trained on. The number that matters to your operation is what percentage of files route to a human and how long each of those reviews takes.
What does an examiner ask about AI underwriting?
What the model is, how it was tested, how you monitor it, and how a decision or a figure can be explained after the fact. Products that generate fair-lending or traceability artifacts as part of running answer most of that before the question is asked.