Why Auto Lending Abandonment Spikes at the Identity Step

A buyer who has already picked the car will sit through a credit pull, an employer, three years of address history. Then the flow asks for a photograph of a licence and a photograph of a face, and a share of those buyers stop. The step takes ninety seconds when it works.

Identity verification drop-off in an auto application concentrates at document and face capture rather than at consent or reluctance. Government test programmes measuring commercial systems on real volunteers between 2024 and 2026 found genuine documents wrongly rejected across almost the whole scale, and genuine applicants falsely rejected between one in ten and one in two, depending only on which system was in the flow.

What does identity verification drop-off look like where it has been measured?

No lender publishes a step-level identity drop-off curve, so the measurements come from government services and test programmes instead, and they show spread rather than a single number. The US Government Accountability Office reported in October 2024, in GAO-25-106640, that the Small Business Administration observed a 30 to 40 percent failure rate during account creation on Login.gov.

One widely repeated figure came from a vendor rather than an evaluator. The Government Accountability Office recorded in June 2025, in GAO-25-107273, that ID.me had put its authentication rate at about 70 percent against about 40 percent under the previous system, in May 2022 testimony by the company whose product it describes. That report also found supplier-reported pass-rate data was not being independently validated. The wider shape this step sits inside is in the abandonment curve step by step.

Why does capture fail before matching does?

Failure to acquire a usable image, rather than failure to compare two faces, is the dominant error mode in tested systems. The Department of Homeland Security Science and Technology Directorate's 2018 Biometric Technology Rally report, published October 2018 with 363 subjects across eleven systems, found acquisition failures exceeded matching errors by six-fold or more in six of those eleven.

The pattern held seven years later in the remote case. In the Remote Identity Validation Rally results published by the Maryland Test Facility in January 2026, all sixteen selfie-to-document systems extracted a face from the selfie with a zero percent failure rate. Extracting the face from the identity document failed for between none and 37.47 percent of attempts. The face was never the hard part. The card was.

Ranges measured by US government test programmes on commercial remote identity verification systems, showing the distance between best and worst performer in each test rather than an industry average.

What was measured	Best system	Worst system	Source and date
Failure to extract a face from the identity document	0%	37.47%	Maryland Test Facility, Remote Identity Validation Rally, selfie-match-to-document track, January 2026
Genuine documents wrongly rejected	0.60%	97.30%	Maryland Test Facility, Remote Identity Validation Rally, document validation track, 2026
Genuine selfie failing to match a genuine document	0.08%	70.84%	Maryland Test Facility, Remote Identity Validation Rally, January 2026
Genuine users wrongly rejected by passive presentation attack detection	0.3%	57.7%	Maryland Test Facility, Remote Identity Validation Rally, 645 volunteers, February 2026
End-to-end false rejection of genuine applicants	10.5%	53.1%	Fatima and colleagues, Clarkson University, UNC Charlotte and GSA, September 2024, 3,991 participants

What do the device, the light and the applicant's skin have to do with it?

Image quality drives false rejection, and image quality is not evenly distributed across a customer base. The National Institute of Standards and Technology states in FRTE Part 8, NIST IR 8429, published 12 July 2022, that false negatives depend strongly on image quality and that poor photography of a face can induce a demographic effect.

The same NIST report is specific about the mechanism. It attributes false negative inequities substantially to poor photography of certain groups, including under-exposure of dark-skinned individuals, and gives the remedy as correcting the capture process with better cameras, imaging environments and human factors.

Measurement backs the mechanism. Cook and colleagues, in IEEE Transactions on Biometrics, Behavior, and Identity Science in February 2019, working from the same 363-subject DHS dataset, found lower skin reflectance associated with roughly a 10 percent reduction in similarity scores against historic reference photos and a 20 percent effect on transaction time against a 6.2 second baseline. Darker-skinned participants were held at the sensor longer and scored lower when they left it.

Now put that on a forecourt at seven in the evening. The Pew Research Center reported on 8 January 2026, from fieldwork run February to June 2025, that 16 percent of US adults are smartphone-only internet users, rising to 34 percent in households under $30,000. The oldest phone and the worst light sit in the subprime queue. An applicant who fails three times through no fault of their own has not made a choice about the loan. The design made it for them.

Does the choice of accepted document change the drop-off?

The accepted document changes the physics of the check. The European Union Agency for Cybersecurity, in its March 2021 report on remote identity proofing, describes the split. A passport carries a chip readable over NFC, returning a facial image and integrity checks from the chip. A document with no chip leaves the check resting on photographs of a card.

North American driving licences fall on the second side of that line. The AAMVA DL/ID Card Design Standard, current edition June 2025, gives jurisdictions a common design, and each issues on its own cycle, so a template library carries many live versions per issuer. A licence a system has not seen is a licence it cannot read.

The bind for vehicle finance is that the licence is the document every applicant already carries and the one with the least machine-readable structure. Accept only passports and the check gets cleaner while the queue gets shorter. Nobody buys a truck that way.

How do applicants feel about being asked?

Reluctance to hand over identity documents is real, measurable, and smaller than the design problem. The Office of the Privacy Commissioner of Canada's 2022-23 Survey of Canadians on Privacy-Related Issues, fielded to 1,500 residents in late 2022, found 65 percent would not be comfortable having their face scanned to verify their age online.

The same Canadian survey found 74 percent had refused at some point to provide personal information over privacy concerns. The UK Department for Science, Innovation and Technology's Digital Identity Sectoral Analysis Report 2026, published 8 July 2026 from a November 2025 survey of 5,658 UK residents, found 41 percent still preferred physical documents for in-person age checks.

Reluctance sets a ceiling on how many applicants start the step. It does not explain a spread from 10.5 to 53.1 percent among people who agreed and tried.

Where does this collide with fraud control?

Every setting that reduces false rejection of genuine applicants loosens something on the fraud side, and one number prices the trade. In the Maryland Test Facility's February 2026 results, passive systems wrongly flagged 0.3 to 57.7 percent of genuine presentations. Fraud at the desk sits in point-of-sale fraud, the ladder of checks in liveness checks explained.

Which fixes are ranked highest for the effort?

Ranked by effort against expected recovery, four changes come before rebuilding anything, and all four sit on the capture side of the step, because capture is where the measured failures concentrate. The fifth is an engineering project, last in effort and first in effect.

1. Show a worked example before asking for the capture. NIST IR 8171, a contactless fingerprint usability study published March 2017 by Furman, Stanton, Theofanos and colleagues, measured unaided capture success at 10, 3 and 53 percent across three devices. After a short instructional video those devices reached 73, 45 and 78 percent.
2. Give live feedback and a visible ready state. The same NIST study named a missing ready indicator and no confirmation of capture as the gaps.
3. Measure the system on the applicant population that actually walks in rather than on a published benchmark. Independent validation of supplier-reported pass rates was the gap the Government Accountability Office named in June 2025.
4. Keep a second path open for applicants the automated path rejects, and offer it up front. Login.gov's pilot proofed 48,505 people between January and March 2024 and finished 1,138 in person.
5. Re-engineer capture conditions with better cameras and controlled lighting. Highest effort, and NIST's own first recommendation.

How would recovery be measured?

Recovery shows up in attempts before conversion. Attempts per applicant, time in step, and the share reaching a second and a third try are all more sensitive than a pass rate, which moves last and hides which direction the flow moved.

Three splits carry most of the information: device operating system and age, time of day, and document type presented. False rejection has to be counted apart from abandonment, because a fast rejection and a system that never answers produce the same exit and need opposite fixes.

What we couldn't verify

No published step-level identity verification drop-off exists for any lending funnel, in any country, from any independent publisher. Figures that circulate come from identity verification vendors reporting their own funnels, with sample, denominator and measurement window undisclosed. None are used above, and the one vendor-originated number here is labelled as such in the sentence.

Nothing above is from a vehicle finance application. Government services and laboratory rallies are the closest analogues, and those populations differ from a Saturday forecourt in motivation, urgency and what failure costs them.

The systems in the Maryland Test Facility rallies are anonymised, so a buyer can read the spread and cannot act on it. The General Services Administration's equity study is public only as a September 2024 preprint, with no final peer-reviewed version located as of August 2026.

Nothing above is Canadian except the attitude survey. Canada has no equivalent of NIST's evaluation programme or the Maryland Test Facility, and nobody publishes verification performance results on Canadian documents or Canadian faces. A lender buying a check on a Canadian licence has no independent number to buy against.

No published data exists on how many attempts an applicant makes before leaving, the one measurement that would settle whether identity drop-off is refusal or exhaustion. Our read is exhaustion, and we cannot prove it.

Common questions

What causes identity verification drop-off in a loan application?
Capture failure more than refusal. The Maryland Test Facility's Remote Identity Validation Rally results, published January 2026, found all sixteen tested systems read a face from the selfie with zero failures, while reading it off the identity document failed for up to 37.47 percent of attempts.

How much of the drop-off is the applicant's fault?
Less than most funnel reviews assume. Fatima and colleagues, in a September 2024 study of 3,991 participants run with the US General Services Administration, measured false rejection of genuine applicants at 10.5 to 53.1 percent across the five commercial services tested.

Do identity checks fail more for some applicants than others?
Yes, and NIST attributes the pattern to photography rather than to the applicant. NIST IR 8429, published 12 July 2022, states that false negatives depend strongly on image quality and that under-exposure of dark-skinned individuals substantially causes false negative inequities.

Is a passport a better document to accept than a driving licence?
Mechanically, yes. The European Union Agency for Cybersecurity's March 2021 remote identity proofing report describes passports as chip-carrying documents readable over NFC, while a licence with no chip leaves the check dependent on photographs of a card. Narrow acceptance shrinks the pool.

Is there Canadian data on identity verification completion?
No. No Canadian body publishes verification performance evaluations comparable to NIST's or the Maryland Test Facility's. The nearest Canadian measurement is attitudinal: the Office of the Privacy Commissioner of Canada found in its 2022-23 survey of 1,500 residents that 65 percent were uncomfortable with an online face scan.

---

Carousel builds the intake layer that sits between an applicant and a lending decision, capture step included. See the auto intake suite