A decade ago, most lenders looking at document automation reached the same conclusion: Optical character recognition (OCR) was a commodity.
The models were available and the engineering talent was already in the building. Paying a vendor for something the team could stand up internally was hard to justify.
The projects launched with confidence, but very few delivered on their promise.
The ones that did rarely reached the accuracy needed to automate a funding workflow, much less allow for lights-out funding, and the cost of maintaining them kept climbing long after the initial build was finished.
That history matters now, because the same conversation is happening again with generative AI.
The tooling is new, the reasoning is identical and the outcome is trending the same way.
The build case always sounds right
The logic behind building in-house is rarely wrong on its face. A lender has proprietary document flows, specific state requirements and a team that understands the business better than any outside vendor.
General-purpose AI models are more capable than the OCR engines of 10 years ago, and they are easy to call through an API. A working prototype can be assembled in a matter of weeks.
The prototype is where the trouble starts.
A demo is not production
A model that extracts a VIN correctly in a demo is not a model that extracts a VIN correctly across every title application, odometer statement and skewed fax in a production deal jacket. A VIN read that misses a single character is a wrong VIN.
A model that handles a clean pay stub is not a model that handles a pay stub split across a page by unrelated Social Security information, or an address broken across multiple lines and scattered through the OCR text.
The demo clears the first 10 documents; production sends the next 10,000.
The hidden costs are where the budget goes
The visible cost of a build is the engineering time to produce the first working version. That number is almost never the real number.
Auto lending documents carry problems that do not show up until a model meets them at scale. Text encased in boxes on title applications fragments under standard recognition. Retail installment sales contracts and buyers’ orders run past the token limits of common extraction models, forcing pre-processing that lowers automation.
Itemization sections follow no fixed schema, so a model must understand that a sales tax figure sits inside a cash price figure rather than beside it. Signatures must be detected, assigned to the right party and evaluated against whether that section even required a signature.
Each of these is a separate engineering problem, and each one must be solved, monitored and kept working.
The build is never finished
New ancillary products, new state forms and revised disclosures arrive constantly. A model trained on last year’s forms quietly degrades against this year’s paperwork.
Catching that degradation requires benchmarking against manually annotated data on an ongoing basis. The work does not end at launch; it becomes a standing obligation.
License charges add up faster than the model improves
Generative AI introduces a cost the old OCR builds did not have. Every document runs against a general-purpose model, which carries a per-call charge, and deal jackets are long. A monthly invoice becomes a line item that grows with volume rather than shrinking with efficiency.
Ultimately, spend can rise while automation stays flat. A lender can pour real money into model calls, engineering hours and annotation, and still be reviewing the same share of documents by hand because the hard cases were never solved.
Why the OCR history repeats
The lenders that tried to build their own OCR were not short on talent or intent. They underestimated how much domain knowledge was buried in the documents.
Reading the characters on a page is the easy part. Knowing that positions four through eight of a VIN are predefined for a specific vehicle model, that a form number cut off at the bottom of a page can still be inferred from the administrator, that a checkbox for assigned without recourse changes how a contract has to be verified, is the part that took years and millions of documents to learn.
Generative AI does not remove that requirement. It raises the ceiling on what is possible and leaves the domain problem exactly where it was.
A model that can reason about text still does not know what a valid auto lending document is supposed to contain unless someone has taught it, benchmarked it, and kept it current against tens of millions of deal jacket documents and the annotations that make them usable.
What buy actually buys
The case for buying is not that a lender cannot build. It is that the parts that make document automation work are the parts that take the longest to build and never stop needing attention:
- Auto-specific models trained on the documents that come through a dealer;
- Standardization that turns a scattered address into a usable one;
- A knowledge base that corrects an OCR read against what a valid entry has to look like;
- Service level commitments on the verifications needed to board a loan; and
- Monitoring that surfaces concept drift before it costs a funding decision.
A lender that buys these capabilities gets industry results without funding the research and development to reach them, and without carrying the maintenance that follows.
A lender that builds takes all of it on, indefinitely, and finds out too late how much of it there was.
The question was never whether a team can build. It is whether a year of engineering spend and a permanent maintenance burden is the best use of the budget, when the alternative already works across the documents a lender sees every day.
Everyone said they would build their own OCR. The ones that are automating today mostly did not.
Jessica Gonzalez is the vice president of customer success at Informed.IQ and has more than 15 years’ experience in the financial services industry, including tenures at Santander Consumer USA and Visa.
Content sponsored by Informed.IQ
Auto Finance Summit, the premier industry event for auto lending and leasing, returns October 5-7 at Caesars Palace Las Vegas. To learn more about the 2026 event and register for early-bird pricing through Aug. 21, visit www.AutoFinance.live/AFS.
