In a benchmark Reducto ran with MicroOne, frontier models beat specialized parsers on precision but lost on recall — they quietly left information out. Adit Abraham says that failure mode is especially dangerous because dropped content is nearly invisible in production review.
Adit says generic benchmarks are increasingly promotional, so 'state of the art' tells you little about your own data. He recommends testing on a representative distribution and using a thoughtful LLM-as-judge rubric. He notes one public benchmark's published leader got much worse than rivals once documents were tilted or blurred.
Adit argues labs are racing toward superintelligence, and a ChatGPT vision encoder has to handle selfies and image filters. Reducto instead narrows to document tasks it must be exceptional at, which he says is a big part of why a small team can stay ahead.
At scale Adit says cost and latency matter alongside accuracy, and a frontier model is far too expensive to apply per page. Reducto aims for frontier-level quality on documents at roughly a cent per page, which he says is orders of magnitude more aggressive on price.
For Reducto's Deep Extract, Adit says the harness is model-agnostic and gives the model tools to validate its own work — for example, a parent agent confirming that a large invoice's line items add up to the stated total. That self-checking is how it reaches near-perfect accuracy.
Adit notes Excel files can be arbitrarily large, with one giant sheet or many tabs scattered across the workbook, and models struggle to consume them even with large context windows. He says the fix is building a model-friendly representation of the spreadsheet rather than feeding the raw file.
Adit highlights that 16% of American tax filers still file by paper, part of a huge amount of manual data entry he expects to disappear first. He frames physical documents being typed in as rote work that humans shouldn't have to do, and the clearest near-term target for automation.
Adit cites a stat that in 80% of Medicare appeals the original prior-authorization denial was reversed, which he calls shocking and evidence of real inefficiency. He ties it to how healthcare and immigration processes bottleneck on needing accuracy high enough to justify trusting a human reviewer.
Adit points out that human 'ground truth' is often wrong — people miss content, drop rows, or fill in only part of a form. When an agent both extracts and reviews its own output against validation rules, it can push structured extraction past human accuracy.
Adit describes a healthcare customer where a pipeline read 'chest pain' as just 'pain,' causing a claim to be rejected when a doctor searched for a history of the condition. He notes such packets can run tens of thousands of pages, and even the best tools hit accuracy sometimes below 50%.
Adit compares codegen, where tools hit a million ARR with shaved heads and later passed a billion, to healthcare, legal, and finance, which still lag despite obvious value. He argues the gap is guardrails, citations, and infrastructure — not raw model intelligence — because mistakes in those fields carry real stakes.
