The counteroffer layer of the Underwriting Engine
Auto-Structure searches the permitted structure space on every salvageable application, validates each candidate against your lender-owned policy, and returns ranked counteroffers in seconds. It is exactly as aggressive as the rules you give it, and no more.
Built for subprime and near-prime auto finance companies.
Hundreds searched · top 3 returned
The Underwriting Engine beneath it is in production with Tracir Financial Services, a multi-state auto lender
Auto-Structure is now entering design-partner deployment: two lenders for 2026
Your deterministic policy validates every candidate. Never a black box.
Where the money leaks
Most salvageable files do not die because the borrower was unfundable. They die because the right structure never surfaced while the customer was still in the building.
The bureau pull was paid. The dealer relationship was paid. Then the file died in pending, and the competitor who found a permissible structure first captured both the yield and the dealer.
Dealers shop every application and want to close the customer while they are still at the desk. The first lender back with a workable structure usually writes the deal. A slow yes performs like a no.
Restructuring skill lives in one or two heads, and those heads go home at six. Nobody systematically searches every salvageable file for the structure that would pass. Auto-Structure runs that search on every file, every time.
The mechanism
For each salvageable file, the system permutes the levers you allow, has your LOS simulate the numbers, validates every candidate against your deterministic policy, and returns the survivors, ranked and explained. The search is exhaustive so your people do not have to be.
The deal as submitted, read from the system of record.
What the customer can add, carry, and sign for.
Cash down, term, payment, and amount financed, permuted.
Payment, taxes, fees, and LTV computed by the calculator of record.
Every candidate passes or fails your deterministic rules.
Survivors ordered by fit, economics, and friction, with the why.
The dealer picks a structure; a changed file re-enters the loop.
The accepted structure lands in the system of record. Never around it.
Your LOS calculates and persists every number of record.
Underbot searches, ranks, and explains. It never decides.
Your deterministic policy is the only thing that says yes.
Where the AI sits, exactly
It interprets the application, picks productive search strategies, and explains the ranked options in plain lender English.
Deterministic rules validate every candidate. No counter leaves the building without passing your policy.
It stays the authoritative calculator and the system of record for every number on the contract.
The AI never determines eligibility, computes a number of record, or authors the official reason for a credit decision.
Cash down, term, payment, amount financed. A finite grid the machine can walk exhaustively and re-validate against policy, candidate by candidate.
A different VIN means a different price, book value, and mileage. That is not a slider value; it is flagged as an opportunity and evaluated on real inventory data.
A co-borrower brings consent and a fresh bureau file. Auto-Structure can suggest one where the pattern fits. It never invents one into the numbers.
See the search run
This sample file is stuck: fundable borrower, failing structure. Tell the system what the customer can actually do, then let it search. Every number below is computed live from the disclosed demo policy, not staged.
Set the two constraints, then search. The machine does the rest.
Vehicle swap and co-borrower are deliberately absent. Each one creates a new application state: real inventory, real consent, a real bureau file. In production they are evaluated on real data, never simulated with sliders.
That grid was 42 candidates on two levers. Production searches finer grids across more levers, on every salvageable application, at any hour, and every counter still has to pass your policy.
Candor section
Auto-Structure is
Auto-Structure is not
Transparent arithmetic
The same math, line by line, ready to be challenged:
Illustrative and deliberately conservative. "Expected contribution per funded deal" is your number, not ours, and 3% assumes the machine only rescues a sliver of what dies today. The audit runs this arithmetic on your book, with your figures, and shows every assumption.
The first step is a number, not a contract
We encode your credit policy in the Underwriting Engine, replay about twelve months of your declined and died-in-pending applications, and hand you the readout: the policy-eligible structures that appear to have been missed, with estimated dollars attached, dealer by dealer, every assumption documented.
In scope
Read before you buy
Led by the engineer who built end-to-end automated underwriting at a national subprime auto lender, 2021 to 2023, before wiring that playbook into Underbot.
From readout to automation, one gate at a time
Historical and read-only. Your policy encoded, your last 12 months replayed, the missed-structure readout with dollars and assumptions attached.
The engine counters every live salvageable file silently, logged next to what your desk actually did. Agreement and misses get measured, not argued.
Your underwriters send machine-ranked counters with one click. Same policy, same trail, human finger on the button.
The gate before automation: LOS write-back verified, data quality proven, expirations and re-entry handled, lanes signed off by you.
Counters fire without a human touch, only in the lanes you authorized. The gray zone stays human. Every counter still passes your policy.
Each rung is a keep-or-walk decision with its own evidence. Historical analysis finds possibilities; shadow mode proves them against live dealer behavior before anything is automated.
Straight answers
Automation runs only in the lanes you explicitly authorize, and those lanes hold the files where judgment adds nothing but latency. Your veterans stop re-deriving obvious counters and spend their day on the files where judgment actually earns its keep.
A counter only exists if your deterministic policy passes it. Auto-Structure changes how many permissible structures get found and how fast they reach the dealer. It changes nothing about what is permissible unless you change the policy yourself.
Constraints come from the dealer conversation, and every counter states its ask explicitly: this structure works with $1,000 more down, this one with none. The dealer and customer decide what is real. The system never books an assumption as a fact.
A vehicle swap or a co-borrower creates a new application state: real inventory, real consent, a real bureau file. Auto-Structure surfaces the opportunity and, once real data exists, evaluates it through the same policy gate as everything else.
The LOS is the authoritative calculator and the system of record. Auto-Structure requests simulations and reads results; it never maintains a competing set of numbers that could drift from the ones on the contract.
Accepted structures post back through the LOS's supported interfaces, with idempotency and a full audit trail. The integration is scoped per LOS during design-partner onboarding, and if an application changes mid-flight, counters expire and the changed file re-enters the loop.
Any data-integrity doubt on a file disqualifies it from automated handling. A counter is a promise to a dealer, and the system refuses to make promises on numbers it cannot trust.
Every decision and counter traces to deterministic, factor-level reasons, and the AI never authors the official reason for a credit decision. That supports your compliance and audit readiness. It does not replace it: your lending team and counsel own policy, notices, and workflows.
An unfunded application has no repayment outcome, so the audit never claims one. It estimates expected contribution from funded deals with comparable profiles and structures, labels every assumption, and shows confidence alongside the number.
The Underwriting Engine answers 'would we fund the submitted deal?' and shows its work. Auto-Structure runs that engine hundreds of times per file to answer 'what fundable deal should we offer instead?'. It is built on the engine and sold as its layer, not as a replacement.
See it on your own book
Request an audit fit review. We confirm your available LOS data, policy scope, and historical volume before proposing the audit, so the first thing you commit to is a conversation about whether the case exists on your book.
Want the evaluation layer first? Start with the Underwriting Engine, the scorecard Auto-Structure is built on.