Reducto has launched Deep Extract, a new agent‑in‑the‑loop extraction engine that self‑verifies until accuracy hits the target. The tool tackles long, multi‑page invoices and financial statements by breaking them into sub‑tasks, re‑extracting missing fields, and confirming totals before delivering results. Early beta runs reached 99–100% field accuracy on documents up to 2,500 pages.

Customers have long complained that single‑pass models collapse on documents with hundreds of rows, often skipping entries or consolidating data. Reducto counters this by defining correctness in the system prompt—for an invoice, for example, the agent must verify that all line items sum to the stated total. The result is a self‑correcting cycle that mimics human audit work.

Beta testing saw the system process more than 28 million fields across documents reaching 2,500 pages, outperforming expert human labelers on accuracy while cutting review time. When citations are enabled, the output includes bounding boxes for every extracted value, enabling audit trails and traceability in high‑stakes workflows.

Developers can activate Deep Extract by setting deep_extract: true in the Extract endpoint and adding optional verification criteria to the prompt. For enterprise teams processing high‑stakes documents, the company offers direct support to assess fit. The addition delivers faster, cheaper, and more consistent extraction at scale, tightening compliance and reducing manual labor.