Thresh reads filled forms, contracts, and scanned packets and returns the names, dates, amounts, checkboxes, and tables that matter, each one linked to the exact spot on the page it came from. No re-typing, no page-flipping.
Accuracy measured on 37 filled documents (five complete real-estate transaction packets plus scanned and handwritten paperwork) with verified ground truth, July 2026. How we measure.
Pick a sample from the benchmark corpus. Everything you see is a real extraction the production API already ran, replayed in your browser: click a field and it lights up the exact spot on the page it came from. Nothing is uploaded and no model is called.

11 FIELDS EXTRACTED · RUN c9600a50 ON THE PRODUCTION API
The workspace adds the whole Whitfield deal and the synthesis replay.
Pick the saved template and drag in a file or a whole packet. PDFs, scans, and handwriting all work.
Names, dates, amounts, checkboxes, and tables come back as organized data. Every value is linked to the page it came from, and anything uncertain is flagged instead of guessed.
Review mode shows each value beside the highlighted spot on the original document. Accept or fix in one click, and results flow into the system your team already uses.
Real documents travel in packets: a contract, its riders, disclosures, notices. Group them into a deal and Thresh synthesizes one deal object across all of them: parties, financials, key dates, a role for every document, and the riders the contract references cross-checked against what's actually in the packet.
The real product against the live API: sign in, create a deal, upload a packet, synthesize. Recorded from the production console; the synthesis wait is trimmed.
Your documents don't live in Thresh; they live in your CRM, your ERP, your claims system. Signed webhooks deliver every extraction to your endpoint the moment it completes, so results land where your team already works. Here's our Salesforce worked example: a contract attached to an Opportunity comes back as native field records, from roughly 150 lines of reference code. The Follow Up Boss example and the rest are on the integrations page.
Recorded against a real Salesforce Developer org and the live API: register the webhook, attach the document, and the extracted fields arrive as records with page and confidence. Not a managed package; it's reference code an integration team adapts in an afternoon. Ask for the reference integration.
Thresh is an API underneath: define a schema, upload a document, get conforming JSON back with signed webhooks. The quickstart, SDKs, and the timed zero-to-integration run live on the developer page.
Transaction packets into verified data: contracts, riders, disclosures, and inspection notices, with deal grouping and cross-document discrepancy checks.
Read moreCompensation, wRVU terms, tail coverage, and restrictive covenants pulled from employment agreements, each value linked to its source clause.
Read moreWhy text-first extraction loses tables, checkboxes, and provenance, and what schema-first extraction does differently. Honest about when OCR still wins.
Read moreEvery plan includes fields, tables, Q&A, deal synthesis, webhooks, and the console. Deal synthesis counts one page per document in the deal (minimum 3). Failed extractions are never billed. Thresh is currently invite-only: request an invite and we'll set you up. Design partnerships are free.
Row-level security enforced by the database itself, not application code.
TLS everywhere; documents encrypted in storage.
Extraction runs inside our cloud boundary; pages are not retained for training.
Every upload, extraction, key event, and webhook delivery is logged.
Architecture built from day one toward SOC 2 and HIPAA readiness (not yet audited).
We're taking on a small number of design partners. Send ten of your real documents, the ones your current process gets wrong, and within a week you'll get back verified results, a per-document accuracy report, and a working integration. Free during the partnership; you help shape the roadmap.
support@usethresh.com