← Back to Projects

Underwriting Without the Credit Bureau: Small Models, Then Agents

How industry-specific models and a multi-agent review replaced a dependency on bureau credit scores for B2B buyer risk, and where each approach belongs.

Case study, 2026. AI product lead at a B2B financing platform.

Illustration: three groups of blue bar charts, one per industry model, connected to a central decision point that links to a dark hub with four orange nodes, representing the review agent and its four sub-agents.
0.65AUC, up from 0.50 with the bureau score
20–50%more delinquencies flagged
Optionalbureau score, once alternative data was in
4research sub-agents in the review layer
Role
AI product lead. Designed the risk models and the agent architecture, working with finance and risk leadership.
Context
A B2B financing platform deciding whether business buyers can repay.
Approach
Industry-specific, explainable models first. A multi-agent review only where context changes what the data means.
Note
Company, customer and data details are generalized. Figures are the ones I'm able to share.

00

The problem

The platform used a major credit bureau’s score to decide which business buyers could repay. For this market it was unreliable. On our own repayment data it performed little better than chance (an AUC of about 0.50), so it approved buyers who later went delinquent and gave little signal on the ones who wouldn’t.

A general-purpose score is built for every borrower at once. Our buyers were businesses in a handful of specific industries, and we had something the bureau didn’t: our own record of how those buyers actually paid.

The betA smaller model that knows the industry will predict repayment better than a large general one that doesn't.

01

One model per industry

I trained custom machine-learning models on the platform’s payment history and buyer delinquencies, then investigated both internal and external signals to find what predicted a buyer’s ability to repay.

The key finding: the signals that matter differ by industry. A signal that strongly predicts repayment in one industry can be noise in another. One blended model averages those differences away, so I built industry-specific models instead.

  • Deterministic and weighted. Each model is a weighted, explainable scoring model, so every score can be traced to the signals behind it. That matters for underwriters, for customers and for anyone auditing a decline.
  • Internal plus alternative data. Payment behavior on the platform, combined with external and alternative data sources, in place of a single bureau number.

02

Testing the bureau score

The obvious question was whether to keep the bureau score as one input among many. I tested the models with and without it.

With the internal and alternative data in place, adding the bureau score barely changed the results. The models performed just as well without it. That turned the bureau from a dependency into an optional input, with the cost and reliance that came with it.

03

A model that re-ranks itself

The models were backtested against historical outcomes before going live. Once live, they re-rank automatically based on each quarter’s repayment performance, so the weights follow what is actually happening rather than what happened when the model was trained.

Results against the bureau-based approach:

  • AUC rose from 0.50 to 0.65.
  • 20–50% more delinquencies were flagged, including clear declines the previous approach had missed.

04

Where deterministic models stop

Some risk doesn’t fit a weighted score, because the same data can mean different things depending on the business. A lawsuit, a funding round or a fast hiring curve can be a good sign for one company and a warning for another, depending on its size, industry and stage. Conditions also keep changing, so a fixed rule written today can be wrong next quarter.

These are the cases a human underwriter handles with a manual review: read everything relevant, weigh it against the company’s situation, and make a judgment. That is the part I gave to agents.

05

The underwriting review agent

I designed a multi-agent system that acts as a manual underwriting review:

  • Four sub-agents with research tools, coordinated to investigate a buyer from different angles.
  • A credit-review skill that encodes how an underwriter works through a file: what to look for, and how to weigh it for a company of a given size, industry and stage.
  • Research, not fixed retrieval. Instead of pulling a predefined set of fields, the agents follow leads one to four degrees out from the buyer, because which information is relevant depends on what they find.
  • Signals with sources. The output surfaces the signals that drove the recommendation, so an underwriter can check the reasoning rather than trust a number.
Underwriting architecture. Buyer data (payment history, delinquencies, alternative data, optional bureau score) feeds two paths: industry models for clear cases, which are weighted, explainable and backtested, and a review agent for when context matters, with four research agents, a credit-review skill and leads followed one to four degrees out. Both produce a decision with reasons: score and rank, approve or decline, evidence and sources. Quarterly repayment outcomes update the model weights.

The agent layer doesn’t replace the models. The models handle volume and the clear cases consistently and cheaply. The agents handle the cases where context decides.

06

A decline the score missed

One case showed the difference clearly. A well-funded, fast-growing, venture-backed company applied, and the scoring model rated it a good risk.

The review agent flagged it for decline. It had read the company’s open legal cases and their current status, compared that exposure with the company’s growth and funding, and concluded that the litigation risk was too high, even for a company of that size and growth rate.

About a month later, the company publicly filed for bankruptcy.

The scoring model wasn’t wrong about the data it saw. The data just didn’t capture what the legal filings meant for this particular company. That kind of judgment needs context, and that is where agents earn their place.

07

What I took from it

Agents are not a universal answerDeterministic models were better for scoring at volume: consistent, explainable and cheap. Agents belonged only where context changes the meaning of the data.
Small, precise models beat one big oneAI tooling makes it practical to build and maintain many specialized models, which used to be too tedious. Each one predicts better for its use case than a single general model.
Test whether you need the incumbentRunning the models with and without the bureau score showed it added almost nothing once better data was in. That one test turned a dependency into an option.
Credit scoring is moving past the bureausAlternative data plus first-party repayment history can outperform traditional scores in specific markets, and agents can cover the judgment calls those scores never handled.