Reading the AUS Findings: Where an AI Underwriting Agent Adds Value, Where SR 11-7 Draws the Line, and Why the Model That Touches Credit Is Governed Differently

The AUS Answers One Question and Leaves the Rest

A mortgage underwriter runs the loan through Desktop Underwriter or Loan Product Advisor and gets a recommendation: Approve/Eligible, Refer, or something in between, with a list of findings and conditions attached. That recommendation answers one question, whether the loan as represented meets the investor's eligibility and the automated risk assessment. It does not clear the conditions, does not verify that the file actually supports the data the AUS scored, and does not catch what the AUS was never shown. The recommendation is the start of the underwrite, not the end of it, and everything between the recommendation and clear-to-close is where the underwriter's time goes.

We build the agent that runs across mortgage origination on lender platforms, and the AUS findings are the most instructive place to be precise about what an AI underwriting agent should and should not do. The agent adds value in the space the AUS leaves open, reconciling findings against the file and clearing conditions. The moment the agent's output influences the credit decision, it becomes a model under SR 11-7, and it has to be governed as one. Getting that distinction right is the difference between an agent that speeds underwriting and an agent that is an ungoverned risk in the credit path.

What the Findings Actually Contain

A Desktop Underwriter findings report is a structured document with a recommendation, the data the recommendation was based on, and a list of verification messages and conditions. Each finding is a specific instruction: verify this income with these documents, document this asset, resolve this credit item, confirm this property fact. The findings are generated from the data the lender entered, which means a finding is only as sound as the data behind it. An Approve/Eligible built on an income figure the file cannot support is an Approve/Eligible that will not survive the reverification, and the underwriter is the one who has to catch that the data and the documents disagree.

The agent reads the findings as structured instructions and maps each one to the file. For a finding that says verify employment and income with a particular document set, the agent checks whether those documents are in the file, whether the income they support matches the income the AUS scored, and whether the calculation follows the applicable guideline. This is the same income calculation discipline applied to the AUS output: the finding says verify, the agent verifies against the actual documents, and it flags the case where the AUS scored a number the file does not support. A finding that is satisfied on paper but contradicted by the documents is exactly the defect a pre-funding quality-control review catches later, and catching it while the loan is still open is cheaper by an order of magnitude.

Clearing Conditions Is a Reconciliation, Not a Checkbox

The conditions attached to a recommendation are the file's to-do list, and clearing them is where loans stall. A condition is cleared when the document that satisfies it is in the file and actually satisfies it, which is a reconciliation, not a checkbox. A condition that says provide two months of bank statements is not cleared by two months of bank statements that show an unsourced large deposit, because the deposit raises a new condition. Clearing conditions well means understanding what each condition is for, so that the document either satisfies it or generates the follow-up the file now needs.

The agent works the condition list by matching each condition to the evidence and testing whether the evidence closes it. When the document satisfies the condition cleanly, the agent marks it cleared with the document referenced, so the underwriter sees what cleared it rather than a green check. When the document raises a new question, the agent surfaces the follow-up rather than clearing the condition and moving on, because a condition cleared on a document that opens a new issue is a defect deferred, not resolved. This connects directly to the conditions and incompleteness work in the loan origination system, where the same reconciliation determines both cycle time and the fair-lending record of what was requested and when.

The Line the Agent Does Not Cross

The AUS returns the recommendation. The underwriter owns the decision. The agent sits between them, and the value of the agent depends on keeping that position honest.

The agent produces inputs and verifications. It does not issue the credit decision, does not override the AUS recommendation, and does not tell the borrower they are approved. When the file supports the AUS recommendation, the agent has assembled the evidence that it does. When the file contradicts the recommendation, or when the loan is a Refer that needs manual underwriting judgment, the agent lays out what it found and hands the decision to the underwriter. Manual underwriting in particular is a judgment about compensating factors and the whole borrower picture that the agent informs but does not make, because the decision to extend credit on a loan the AUS would not approve is a credit judgment the lender is accountable for under fair lending and the ability-to-repay rule.

Keeping this line is not caution for its own sake. It is what makes the agent's governance tractable, because an agent that produces documented, reviewable inputs is a component a human decision-maker can check, and an agent that makes the decision is a decision-maker that has to be governed as one. We chose the first architecture deliberately, and the choice is what the model governance below is built on.

When the Agent Becomes a Model

Here is the part that gets skipped. The moment an AI agent's output influences a credit decision, even as an input the underwriter relies on, it is a model in the sense the regulators mean, and it falls under model risk management. SR 11-7, the interagency guidance on model risk management issued by the Federal Reserve and the OCC, defines a model as a quantitative method that processes input data into estimates and applies to the full model lifecycle: development, implementation, use, and validation. An agent that reads a file, calculates income, and tells an underwriter the loan supports the recommendation is processing inputs into an estimate the underwriter relies on, and treating it as ordinary software rather than a model is how an ungoverned model ends up in the credit path.

SR 11-7 asks for three things that shape how we build. It asks for a development record: what the model does, its assumptions, its limitations, and the data it was built on. It asks for effective validation: independent testing that the model performs as intended, including outcome analysis against actual results and testing at the boundaries where it is likely to fail. And it asks for governance: ownership, controls, change management, and ongoing monitoring, so the model does not drift out of the envelope it was validated in. We treat the underwriting agent as a model under this guidance, which means it has documented intended use and limitations, it is validated independently of the team that built it, and its performance is monitored against the outcomes that matter, defects found in downstream quality control and reversals by underwriters.

The NIST AI Risk Management Framework sits alongside SR 11-7 and pushes on the parts specific to AI: the govern, map, measure, and manage functions that address the ways an AI system can fail that a traditional statistical model does not, including the behavior of a language model reading a document. We use both, SR 11-7 for the model-risk backbone that a bank examiner expects and NIST AI RMF for the AI-specific failure modes, because an agent reading a tax return can be wrong in ways a scorecard cannot, and the governance has to cover those.

Testing the Agent at the Boundary

Validation that only measures average accuracy misses the point, because a model in the credit path fails at the edges, not at the average. The cases that matter are the self-employed borrower with a complex return, the income structure the guideline treats specially, the document that is slightly off from the standard form, the deposit that looks sourced but is not. Those are where an agent's reading diverges from a careful underwriter's, and those are where a defect gets funded.

We validate the underwriting agent against a held-out set of files that were underwritten by experienced humans, and we measure agreement on the hard cases specifically, not the easy ones. Where the agent and the human disagree, the disagreement is the signal: either the agent is wrong and the case teaches us where its reading breaks, or the human missed something and the case teaches us where the agent adds coverage, and both outcomes go back into the development record. We also monitor the deployed agent against downstream quality control, because a defect that a pre-funding or post-close review catches on a loan the agent touched is direct outcome data on where the model's envelope has a hole. An illustrative target we hold internally is that the agent's condition-clearing decisions agree with independent re-review at a high rate on the routine files and route the hard files to humans rather than deciding them, because the failure mode we most want to avoid is confident clearing of a condition that a re-review would reopen.

The Honest Read

The AUS recommendation answers eligibility and hands the underwriter a list of findings and conditions that the recommendation does not clear. That space, reconciling findings against the file and clearing conditions against the evidence, is where an AI underwriting agent adds real value, and it is also where the agent has to stay on the input side of the credit decision. The underwriter owns the decision. The agent assembles and verifies what the decision rests on.

The part that separates a serious underwriting agent from a risky one is governance. An agent whose output influences a credit decision is a model under SR 11-7, and it has to have a development record, independent validation, boundary testing, and ongoing monitoring against downstream outcomes, with the NIST AI RMF covering the AI-specific failure modes on top. At Sei, we build the underwriting agent as a governed model that produces reviewable inputs, tested where it is likely to fail and monitored where it is deployed, because an ungoverned model in the credit path is not a productivity gain, it is an examination finding waiting to happen. The AUS answers one question. The agent does the reconciliation the answer leaves open. The underwriter makes the call, and the governance makes the agent something the underwriter and the examiner can both trust.