Upstart's own published data shows its underwriting model approves 44.28% more borrowers than a traditional credit model, at 36% lower APRs, according to Upstart's own reporting. Zest AI reports it now has over 600 active models in production, per Zest AI's own site.
The advantage isn't raw computing power. It's how much manual work assembling the risk picture used to take.
Pulling bureau data, bank transaction history, employment records, and alternative payment history into one file used to mean a human underwriter working through each source one lookup at a time, before they could even start applying judgment to what they found.
This guide covers agentic AI in underwriting specifically, part of the broader picture in Tecla's Agentic AI in Finance and Banking guide.
Here we go deeper: how the risk picture actually gets assembled, where alternative data helps and where it doesn't, and what a lender still has to get right on fair lending and explainability.
What Is Agentic AI in Underwriting?
Most loan origination systems already pull a credit score and check it against an approval matrix. What agentic AI changes is everything that happens before that score gets used.
It gathers bureau data, bank transaction history, and supporting documents into one file, checks them against each other, and flags what doesn't add up before a decision ever gets made.
A static scorecard treats every applicant with the same credit score identically. An agentic system can tell the difference between a 680 built on five years of steady income and a 680 that just cleared a recent collections account.
That's because it's actually reading the transaction history behind the number, not just the number itself.
The result isn't a system that approves more people by lowering the bar. It's a system that catches the applicants a static score alone would have missed, and can explain its reasoning when asked why.
From Credit Scorecards to Agentic Risk Assessment
Underwriting automated in stages, the same as everywhere else in finance. First came static scorecards: a fixed set of variables and weights applied the same way to every applicant.
Scorecards are transparent and easy to audit, which is exactly why regulators are comfortable with them. They're also blunt.
A scorecard built on traditional bureau data has nothing useful to say about a thin file applicant with six months of on time rent payments and no credit card history at all.
Agentic underwriting doesn't replace the scorecard. It decides which additional data is worth pulling for a specific applicant, checks whether that data actually supports a different conclusion, and produces the reasoning trail a scorecard alone never had to produce.
Most production underwriting in 2026 runs both. A scorecard or rules engine handles applicants who clear every threshold cleanly.
Agentic review kicks in where a static score and the underlying data start to disagree, and a senior underwriter signs off on anything the system can't resolve with a documented reason.
The Underwriting Workflow, Step by Step
The workflow below is the coarse version. The sections after it go deeper into the parts that determine whether this holds up under regulatory scrutiny: alternative data, explainability, and fair lending testing.
Alternative Data and Its Limits
Alternative data, rent payments, utility bills, cash flow patterns, exists specifically to help applicants who don't have enough traditional credit history for a scorecard to read with any confidence.
It's most useful exactly where a scorecard is weakest: thin file and new to credit applicants who may be creditworthy but invisible to a model trained only on bureau data.
The limit isn't the data itself. It's that any new variable, alternative or traditional, needs the same disparate impact testing before it goes into a live model, not just at launch.
A variable can be predictive of default and still correlate closely enough with a protected characteristic to produce an outcome regulators will act on.
Explainability and Adverse Action Notices
The Equal Credit Opportunity Act and Regulation B require lenders to give applicants a specific, accurate reason for any adverse action, and the CFPB has been explicit that a lender can't point to a complex or uninterpretable model as an excuse for not providing one.
That requirement shapes the entire architecture. A model has to produce interpretable factors as part of its output, not just a score, because a declined applicant is legally entitled to know why, regardless of how the decision was generated.
Model Risk, Fair Lending, and Disparate Impact
In July 2025, the Massachusetts Attorney General reached a $2.5 million settlement with a student loan lender over its underwriting model's use of a college's average default rate as a scoring variable.
The AG alleged that variable produced disparate outcomes in approval rates and terms for Black and Hispanic applicants. A separate rule that automatically denied applicants without a green card drew the same allegation on national origin grounds.
Neither variable mentioned race or immigration status directly. Both still produced the disparity regulators pursued, because the underlying testing to check for it hadn't been done.
That's the actual lesson: the risk isn't a model that discriminates on purpose. It's a proxy variable nobody tested before it went live.
The regulatory framework itself keeps moving too. SR 26-2 replaced SR 11-7 in April 2026 and explicitly excludes generative and agentic AI from its formal scope.
The fair lending testing obligation under ECOA doesn't wait for that framework to catch up, though. It applies regardless of what generated the model.
The Human Underwriter's Role
The underwriter's job shifts from re-keying data pulled from five separate sources to reviewing what the system already assembled, and deciding the files it couldn't resolve on its own.
Complex commercial credits, relationship-based lending decisions, and genuine judgment calls still belong with a person who can weigh context a model wasn't built to hold.
Implementation: Guardrails Specific to Underwriting
What makes agentic underwriting dangerous isn't a model that gets caught being wrong. It's one that stays quietly, defensibly wrong across thousands of files before a pattern shows up in the data.
The guardrails below are built around that specific failure mode.
Rolling This Out: What to Expect
Pick a single loan product or a single new data source for the first rollout. Trying to convert the whole underwriting function in one pass makes it impossible to tell what actually changed.
Run the system in shadow mode alongside the existing process first, comparing its decisions against what a human underwriter actually did, before letting it decide anything on its own.
Document the bias testing methodology before go live, not after. The Massachusetts settlement above turned in large part on the fact that the disparate impact testing hadn't happened at all, not on the specific variable itself.
A tested variable that shows a disparity is a design problem to fix. An untested variable is a liability with no defense.
Vendor and internal model documentation matters as much as the model's accuracy. A senior underwriter and compliance officer should both be able to explain, in plain language, what the model actually weighs and why, before it touches a live decision.
The Team Behind Production Agentic AI
Getting agentic underwriting right takes more than a capable model. It takes a team that has actually built the testing and documentation discipline around one, not just the scoring logic itself.
Tecla's Agentic AI services design, build, and operate this workflow directly, the same data assembly, scoring, and fair lending documentation above, running in your stack with the evals and guardrails production requires.
Or bring the expertise in-house: AI engineers who've worked on live finance systems, past the demo stage.
Tecla runs a network of senior engineers across the US and Latin America, built over more than a decade, with a top 3% acceptance rate and first candidates in 3 to 5 business days.
FAQ
