Reducto launches Deep Extract to automate verification-heavy data extraction

Reducto today unveiled Deep Extract, a new "agent-in-the-loop" approach it says closes the gap between model output and real-world accuracy. The system runs extraction, then verifies and corrects its own results in a loop until they meet a quality threshold — effectively offloading what many teams still do manually. The result, the company argues, is a way to handle long, messy documents without a human staring at a screen for hours.

What Deep Extract does

Think of it as a checker that checks itself. Instead of a single pass that often shortcuts on thousands of repetitive rows, Deep Extract spawns sub-agents to break documents into bite-sized tasks, extracts fields, verifies them against the source, and re-extracts anything that fails the test. You can tell it what "correct" means — for example, "ensure all invoice line items sum to the stated total" — or it will infer a reasonable rule. With the citations flag enabled, outputs include granular bounding boxes so every value can be traced back to its exact spot in the original file. Handy for audits. Handy for people who hate surprises.

Why it matters

Long, complex documents have been a sore spot for extraction pipelines. Totals that don’t reconcile, dropped line items, endless human-in-the-loop checks — these are real pain points for finance and compliance teams. Deep Extract targets that emotional crux: the relief of no longer having to babysit extraction for hours. It rides the broader wave of long-horizon agent architectures — the same trend powering multi-step automation — and applies it to a problem that’s annoyingly sticky in enterprises.

Beta claims and caveats

It has been reported that Deep Extract has already processed over 28 million fields in documents up to 2,500 pages during its production beta, and that it achieves 99–100% field accuracy on the documents that matter most, allegedly out-performing expert human labelers on some tasks. Those are big claims; customers will want to validate them against their own data and edge cases. Still, if the numbers hold, this could shave days off workflows that used to need human reviewers — and that's something people will notice in their next close cycle.

Sources: reducto.ai, Hacker News