Reducto turns your worst documents into data.

Here is a fact that the current AI boom would prefer you not think about too hard: a large language model is only as good as the data you feed it, and an enormous amount of the world's most valuable data is trapped inside documents that computers find nearly impossible to read. Not text files. Documents - the scanned insurance claim, the 200-page hedge fund prospectus, the compliance questionnaire with the checkbox that someone filled in by hand, slightly outside the box. Reducto's entire business is that gap, and the gap turns out to be worth quite a lot of money.

Reducto was founded in 2023 by Adit Abraham and Raunak Chowdhuri, who met at MIT. They named the company after the Reducto Charm from Harry Potter, a spell whose job is to blast solid objects into pieces. The object being blasted, in this case, is the document, and the pieces are clean, structured, machine-readable data. This is the kind of name that sounds like a joke until the company raises $108 million, at which point it sounds like foresight.

The technical claim is more interesting than the usual "we use AI" gesture. Traditional optical character recognition - OCR, the technology that has been reading text off images for decades - is fast and cheap and falls apart the moment a document contains a complicated table, a rotated scan, or a layout that doesn't match its assumptions. Reducto's approach runs a single page through multiple vision models and pairs classic computer vision with newer vision-language models, the kind that can look at a page and reason about what it's seeing rather than just transcribe pixels.

The part that engineers tend to point to is what Reducto calls Agentic OCR. In plain terms: after the system parses a document, it reads its own output back, looks for mistakes, and fixes them in a second and third pass. It is, roughly, a proofreader that never gets bored. This matters because the difference between a document-AI system that works in a demo and one that works in production is almost entirely a question of what happens on the messy 4% of pages, and the messy 4% is where money and lawsuits live.

"It's probably the only AI product that has actually worked for us." - an engineering leader at a top-five global hedge fund

The customer list is the tell. Vanta, the compliance-automation company, ran Reducto head-to-head against Amazon's own AWS Textract and switched. Harvey, the legal-AI platform used across more than 1,300 organizations, uses it for document understanding. Scale AI, Newfront, Medallion, Rogo, JLL, Toast, Mercor, and a Fortune 10 company are on the roster, along with, per the company, a global top-five hedge fund. When a category leader replaces a hyperscaler's built-in tool with a two-year-old startup, that is worth more than any benchmark.

What can you actually do with it? Reducto exposes the work as a developer API with a handful of verbs. Parse reads a document with its layout and meaning intact. Extract pulls structured fields out according to a schema you define. Split takes a 500-page blob that is really forty documents stapled together and separates them. Classify routes each one to the right workflow. And Edit - which the company bills as the industry's first unstructured document editing API - writes back into the document, filling blanks, tables, and checkboxes without a template. Strung together, those verbs are an entire back office.

The numbers Reducto reports are specific in the way that suggests they came from real deployments rather than a pitch deck: 99.24% extraction accuracy in healthcare, audits running 16 times faster in insurance, monthly processing volume up sixfold in the six months after its Series A. The company says it has parsed billions of pages. None of this is glamorous. All of it is the plumbing that has to work before any of the flashier AI on top of it means anything.

Investors have noticed the plumbing thesis. Andreessen Horowitz led the $75 million Series B in October 2025; Benchmark led the $24.5 million Series A six months earlier; First Round Capital, BoxGroup, and Y Combinator have been along since the seed. The bet is straightforward and not obviously wrong: not every valuable AI company will be a model lab. Some will be the picks and shovels, and the pick here is a spell that breaks documents apart.