Domain Atlas / Lending & credit collections AI
Credit Acceptance Corporation's net-collections score
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings comes from a published baseline, not this deployment's own record. Evidence base: 12 published baseline.
Credit Acceptance Corporation is a publicly traded indirect subprime vehicle lender: it does not lend across its own counter but buys retail instalment contracts from dealerships enrolled in its program, which pay a monthly fee for access to its Credit Approval Processing System and its servicing. The pleaded scale is that approximately 1.9 million consumers obtained loans through the operator and its affiliated dealers between 2 November 2015 and 30 April 2021, that the network exceeded 12,000 affiliated dealerships, and that consumers obtained more than $4.9 billion in operator-financed loans in 2020 alone, with New York among its top five state markets. The operator's own quarterly report for the period ended 30 June 2026 gives the current shape: 11,004 active dealers in the quarter, a record and up 3.3 per cent year over year, with 1,456 new dealers enrolled; 84,615 consumer-loan unit assignments worth $1.0 billion in the quarter; an $8.0 billion average loan portfolio balance; $90.6 million of dealer holdback paid in the first half of 2026; and a $599 monthly per-dealer program fee. The pleaded borrower population had a median credit-bureau score of 546 and a gross annual income of approximately $35,000.[2]
What happened
A person with damaged credit needs a car. They walk into a used-car dealership that is enrolled in a lender's program, pays that lender $599 a month for access to its origination software — the Credit Approval Processing System, or CAPS, named as such in the lender's own quarterly report — and has been trained by that lender on which products to sell them. The salesperson opens the software and starts building a deal: this vehicle, this disclosed price, this term, this down payment, this trade-in, these products financed into the amount owed. Two numbers come out of the software as the deal is assembled. Only one of them is shown to anybody in the room.
The first number is the Score. Both parties in the litigation describe it the same way, and it can be stated plainly. The operator scores the proposed transaction — the applicant's credit-bureau attributes, the application data, the deal structure and the vehicle — and returns a figure from 0 to 100: its estimate of the percentage of total amounts owed that it expects to collect over the life of the contract. What the estimate counts is the part that makes this case unlike its neighbours. It counts the scheduled payments. It also counts late fees, the proceeds of selling the car at auction after repossessing it, what collection efforts recover afterwards, the deficiency judgment, and wage garnishment. The Score prices the payment the operator makes to the DEALERSHIP. It does not decide whether financing is offered. It does not set the borrower's interest rate. The borrower is never shown it. That is not a plaintiff's characterisation: the defendant's own dismissal brief says the Score "does not affect the terms of a Contract" and describes it as "an internal scoring metric that Credit Acceptance uses to calculate the CAC Payment to the Dealer."
The second number is the projected profit for the dealership on this deal, for this applicant, against this vehicle. The plaintiffs allege it is displayed live and recomputed as the deal changes, including as products are added, and they pleaded one screen: $1,605.31 of projected dealership profit with neither a vehicle service contract nor guaranteed asset protection attached, and $4,051.82 with both. A swing of about $2,446, on one screen, for one applicant. The same software instance then generates the borrower's contract and regulated disclosure form.
Now the allegations, and from here every sentence is one. On 4 January 2023 the Consumer Financial Protection Bureau and the People of the State of New York, by Attorney General Letitia James, jointly filed a 59-page complaint in the Southern District of New York pleading seven causes of action. Its structural allegation is in paragraph 3: the operator "has created a complex algorithm to predict how much it will collect from consumers over the life of a loan — not just from consumers' monthly payments, but also from potential collection efforts, repossessions, auctions, and deficiency judgments if the consumer defaults," and "does not use its score to determine whether to offer a loan to the consumer, whether the consumer can afford the loan, or what interest rate to offer." The complaint's own conclusion from that: the lending model "is indifferent as to a consumers' ability to repay loans in full."
Then the number the case is built on. For more than 39 per cent of loans nationwide, and about 25 per cent of New York loans, the plaintiffs alleged the algorithm projected that the operator would not collect even the AMOUNT FINANCED. Keep that separate from the portfolio figure, which is a different quantity: a pleaded average forecast of 64 nationwide and 66 in New York, meaning 64 to 66 cents per dollar of TOTAL AMOUNTS OWED, interest included. The first figure is about principal and it is the load-bearing one. For New York loans originated between 2015 and 2020 whose projected collections fell below the amount financed, the plaintiffs alleged nearly 70 per cent were sixty or more days past due, repossessed, or sold at auction.
The alleged economics explain why a projection like that was not a reason to stop. The plaintiffs pleaded that the operator paid the dealership on average about 72 per cent of projected net collections, and that the nationwide average dealer payment ran about 22 per cent BELOW the amount financed — so recovering roughly 78 per cent of principal was enough to exceed the cash actually at risk, a gap they put at nearly $2,500 per New York loan. The dealership's back-end share, the earnout, was pleaded as rarely paid: under 12 per cent of new loans nationwide sat in dealer pools receiving any, and total earnout came to under 2 per cent of the value New York dealers received. And the ordinary instrument was pleaded as absent by design. The interest rate does not vary with the borrower's risk — New York contracts disclosed 22.99 or 23.99 per cent against a 25 per cent state criminal usury cap — so the risk premium was alleged to have been relocated into the amount financed, where it is invisible to comparison shopping and accrues interest. The plaintiffs' recomputation of that, treating dealership compensation as a proxy for the true cash price and the gap as a concealed finance charge, produced a median New York rate of about 34 per cent with nearly 90 per cent of loans above the cap. The defendant called the proxy "invented ... for this litigation" and irreconcilable with the Truth in Lending Act's defined terms. No court resolved it.
Read the exemplar in the pleading, because it is the whole file in one paragraph. A February 2016 loan scored 60.1. Projected net collections: $7,994, against approximately $13,300 the borrower was obligated to pay. The dealership was paid about $5,614. The vehicle was a seven-year-old sedan with more than 82,000 miles, priced at $8,195. It was repossessed twice, at $700 in repossession charges, and auctioned for a net $772. The borrower was sued for a $7,550 deficiency. Lifetime net collections exceeded $8,400 — more than $2,800 above what the dealership had been paid. The borrower is anonymized in the complaint and stays anonymized here.
What was supposed to stop a deal like that? The plaintiffs pleaded the operator's stated compliance parameters, and pleaded what they do not reach. Proof of income, plus a monthly payment at or under 25 per cent of gross monthly income, which a manager may approve up to 30 per cent — computed with no recurring debt obligations, no housing cost, no food, healthcare or childcare cost, no debt-to-income ratio, no residual-income calculation, and no adjustment for the number of dependents in the household. A vehicle price at or under 115 per cent of the highest published book value for that make and model, applied without inspecting the individual car; against which the plaintiffs measured a median disclosed selling price about 77 per cent over wholesale book value and slightly more than 50 per cent over reported dealer cost, and compared a dealer-compensation-implied markup of about 17 per cent to the roughly 12 per cent industry norm they cite. A list of pre-approved products, of which 90 per cent of loans carried at least one: an average vehicle service contract at $1,545 retail adding an average $2,243 with interest, average guaranteed asset protection at about $782 adding an average $1,615, approximately $250 million of add-on revenue nationwide in 2020, and a flat dealership commission of about $385 per service contract.
The collections leg is pleaded in the same quantified register, and it is the leg the forecast had already counted. Within days of a missed payment the operator would, as a matter of policy, disable the vehicle through a GPS starter-interrupt device — a practice the complaint states was used on New York-financed vehicles until late 2018. Contracts ran 60 to 72 months; vehicles repossessed and resold averaged under two years from origination to auction, and a quarter reached auction inside the first year. A majority of the roughly 1.9 million contracts became delinquent at some point. Nationwide the operator repossessed more than a quarter of financed vehicles; in New York about 44 per cent, and 21 per cent of repossessed New York vehicles were repossessed more than once. Auction proceeds satisfied on average 29 per cent of the remaining amount owed, leaving an average deficiency near $8,500. More than 138,000 accounts were referred to collection attorneys nationwide. Judgments were entered against more than one in six New York borrowers whose loans reached maturity by May 2021, with more than 7,000 default judgments obtained in New York across 2017 and 2018 — in a window when 40 of the thousands of borrowers sued had a lawyer. And, excluding support personnel, the operator employed nearly twice as many people in servicing and collections as in origination.
The operator denied all of it. It moved to dismiss the complaint in its entirety, argued the plaintiffs were "using this action to sidestep the legislative process and impose sweeping regulatory reform", noted that as an indirect lender it has no contact with the consumer until after the dealership's contract is executed and assigned, and said the plaintiffs had abandoned any standalone claim over non-disclosure of the Score. Through its Chief Legal Officer it said publicly that the case "never should have been brought."
Then the case ended without ending. Stayed in August 2023 pending the Supreme Court's decision on the Bureau's funding; the stay lifted in July 2024; the revised motion to dismiss fully briefed by 29 October 2024. On 24 April 2025 the Bureau filed a consented motion under Rule 21 to be dropped as a plaintiff and to withdraw its counsel's appearances. The defendant consented, New York did not object, and the court granted it five days later. The Bureau's supporting memorandum argues only that Rule 21 gives the court broad discretion, that dropping it causes no prejudice, and that at this stage it could have dismissed by notice were it the sole plaintiff. It offers no substantive reason and makes no representation about the merits of the claims it was abandoning. Contemporaneous trade reporting placed the withdrawal within a wider 2025 pattern of agency enforcement dismissals; that is press attribution, not an agency statement, and no motive beyond the filed record is asserted here.
New York continued alone. In September 2025 the operator offered $45.0 million to settle this action jointly with the parallel multistate investigation; in January 2026 it reached preliminary alignment on material terms at a potential $75.5 million across both. The case was reassigned to Judge Jesse M. Furman on 28-29 January 2026, and on 6 February 2026 the court, commending the parties on their efforts to resolve the case, deferred ruling on the fully briefed motion to dismiss and TERMINATED it, saying it would restore the motion if settlement failed. It never had to. On 5 June 2026 the court, advised that all claims had been settled in principle, ordered the action dismissed and discontinued without costs and without prejudice to a right to reopen within sixty days, mooted the pending motions, and noted that it would not retain jurisdiction to enforce a settlement agreement unless the agreement were made part of the public record. It was not. On 23 July 2026 the parties reported agreement on all terms and obtained a forty-five day extension so this resolution could coincide with the multistate one. As of the docket's last update there is no executed settlement, no consent judgment and no public statement of terms — and the $75.5 million is the operator's own disclosure of a POTENTIAL payment across two matters, not a New York settlement amount and not final.
Two things keep this file from resting on an untested pleading. The first is that one jurisdiction's version of these theories already landed. On 1 September 2021 the Massachusetts Attorney General announced a $27.2 million settlement in Suffolk Superior Court — the largest of its kind, that office said — resolving allegations that the operator made high-interest subprime loans it knew or should have known many borrowers would be unable to repay, imposed hidden finance charges violating the state's 21 per cent usury cap, engaged in unlawful collection practices, and failed to tell investors that higher-risk loans were placed into securitisation pools. Relief for more than 3,000 borrowers: cash, debt forgiveness, credit-bureau deletion of related negative marks, and required changes to loan-handling practices, which were not enumerated publicly. No admission; the operator said the suit had been vigorously contested and that it looked forward to continuing to serve customers in the Commonwealth.
The second is that the operator publishes the forecast's own accuracy, every quarter, by assignment year. Its report for the period ended 30 June 2026 sets each year's current forecast collection percentage against the percentage forecast when the loans were assigned: 2020 running 4.7 points ABOVE its initial forecast, 2022 running 8.2 points below, 2025 and 2026 inside a tenth of a point. The company states it reviews that variance monthly and periodically adjusts the statistical pricing model, and adds the sentence that makes this case what it is: "Since all known, significant credit quality indicators have already been factored into our forecasts and pricing, we are not able to use any specific credit quality indicators to predict or explain variances in actual performance from our initial expectations." Accurate forecasting of loan performance is the first of the company's three published critical success factors.
That series is a calibration record on the operator's own recovery — a running check of whether what it forecast matches the rate it actually collected. It is not an error rate, and nothing in this file converts it into one. But it is the reason the ordinary question — was the model any good? — does not get you anywhere here. On the record the company publishes, it was good. The question this file is about is what it was pointed at.
The sociotechnical reading
Read this deployment as a two-audience system and its shape becomes legible.
One audience is dense, interactive and fully instrumented. The dealership sits inside the software, assembles the deal, watches the projected profit move against each choice, and is trained by the lender on which products to attach. Every pathway on the lender's side is wired to that channel: the deal file it builds, the projection it is shown, the advance formula it prices against, the rating that feeds what it is paid.
The other audience is a single unidirectional form. The borrower receives the regulated paperwork the same software instance prints, and nothing else. There is no notice of the forecast, no explanation of it, no contest path and no adverse-action-style disclosure attaching to it, because the forecast is not an adverse action: it decides nothing about the applicant. That is the governed surface in this case — not a decision, but the distance between what one system computed and what its subject was told, manufactured inside one piece of software that speaks to both parties at once.
The objective function is displaced, and that is the mechanism. The model optimises expected recovery TO THE LENDER rather than expected repayment BY THE BORROWER. Both quantities can be estimated well at the same time, and they can move in opposite directions, so "the model performs" and "the borrower is fine" are simply different statements. This is why the usual diagnostic vocabulary fails here. There is no protected-class proxy and no fair-lending claim. There is no false-positive rate, because nothing is classified. There is no denied applicant, because nothing is denied. The accuracy IS the mechanism the plaintiffs complain of, which is a shape the rest of this catalogue does not carry.
The forecast is self-confirming by construction. What it predicts includes what the same organisation's collections arm will recover — auction proceeds, post-default recoveries, deficiency judgments, garnishment. The realised recoveries are then booked as the actual outcomes the model is recalibrated against. So a forecast of "we will collect 60 per cent of this" is partly a forecast about the company's own future behaviour, and the company controls that behaviour. The prediction and the intervention that fulfils it sit inside one organisation, with the monthly variance review closing the loop between them. The correction arm being staffed at nearly twice the intake arm is the organisational form of the same fact: recovery is not a failure mode here, it is a budgeted line.
The controls that exist are parameters rather than reviews. The payment-to-income bound, the price bound and the pre-approved product list run automatically inside the software and bound what a dealership may submit. They are real. They are also bounded by construction — the payment bound is approvable upward by a manager, the price bound is measured against a published book value for the model rather than against the car, and the household's other obligations sit outside the arithmetic entirely. In the Lab's vocabulary that is a bounded automated screen, partial by construction, and it is the honest shape for a control that exists and catches a fraction.
The watching cadence points inward. The one documented review runs monthly and reads the operator's own collection rate against its own forecast; it is published, it is funded, and it is the reason a calibration series exists at all. Nothing in the record re-evaluates the compliance parameters on any schedule. Outside the firm, a decade of subpoenas, a civil investigative demand, a notice-and-opportunity-to-respond letter and two notice-of-intent letters produced a pleading, then a federal withdrawal, then a motion terminated without a ruling, then a dismissal on a settlement in principle whose terms are not public. The external channel ran the whole way and adjudicated nothing.
What is absent is the ordinary instrument. Risk-based pricing to the borrower — the thing that would normally make a lender's estimate of a borrower's risk visible to that borrower as a price — is absent here by design, because the interest rate is flat near the state cap. The plaintiffs allege the risk premium was relocated into the principal instead, where it accrues interest and cannot be compared. Whatever a reader concludes about that allegation, the structural point survives it: this is a deployment where the model's estimate of a person reaches everyone in the transaction except that person.
And the people served are not in the model. No credit decision, price, advance, default, repossession, auction result, deficiency or judgment for any person is computed from the Lab network drawn from this file. The delinquency, repossession and judgment statistics above are recorded observations from a pleading that was never tested, and the operating and calibration figures are recorded observations from the operator's own filings. What the network models is the operator side: who computes, who is shown, who writes, who collects, and who was entitled to ask.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.