Document data extraction: Clairic, Deep Review and my road from 50% to 99.9% with no manual edits

In one week of October, three news items on document processing came out: Cloudflare showed Clairic, a model that pulls typed values out of text in a single pass, Google opened Deep Review for checking long documents, and OpenAI described how lawyers at Sullivan & Cromwell check contracts with agents that argue with each other. I've been building parsers for insurance estimates and inspection reports since 2023 and went from half the documents needing no manual edits to 99.9%, so I read this news with one thought: which of this would have shortened that road for me.

What came out

On October 19, Cloudflare introduced Clairic, a preview of an extraction model on Workers AI, with a promise to open the weights. The idea comes from the “System One” models TypeSafe introduced a month earlier: the model answers in a single pass, with no reasoning and no token generation, you pay only for input, and the answer usually comes back in under a second. The input is text and up to four images plus a list of fields with types: text, number, date, time, duration, list, estimate, forecast. The output is a value of the given type for each field with a confidence score, and null if the answer isn't in the document. The model can be wrong, Cloudflare says.

On October 14, Google Research opened Deep Review, a system for checking complex documents that grew out of the Paper Assistant Tool. Over a year it was run on more than 16,000 manuscripts for the STOC, ICML, NeurIPS and ICLR conferences, and now it accepts PDFs and text from any field, including legal and financial documents. The design: the document is cut into logical segments, each gets its own compute budget, then separate agents check the segments in parallel with access to the whole document.

On October 15, OpenAI described Agreement Analyzer at Sullivan & Cromwell. Partner John Savva started with a single-user prototype built on Codex. A lawyer uploads a draft contract, adds the firm's precedents and a playbook with the client's positions, the tool does several rounds of review and returns a memo with edits the lawyer accepts or rejects. There is a separate description of a scheme where some agents raise objections and others challenge them, so the result shows what held up under argument, what was dropped and what remains unverified. Savva insists that any claim about market practice must have a visible source.

What this means for a company where documents are retyped by hand

Take a car rental company. Every week, invoices from repair shops, inspection reports and insurance documents after accidents come in. The manager opens a PDF, a scan or a phone photo, finds the amount, the VIN and the date and retypes them into a spreadsheet. A mistake in one digit of the VIN surfaces a month later, when the insurer sends the document back.

All three news items are about one problem: how to get a value you can trust out of a document, and how to tell when you can't trust it. Clairic answers with a type, a confidence score and an honest null. Deep Review and Agreement Analyzer answer with a separate checking layer that argues with the first conclusion. The conclusion for the owner: you need the combination of extraction, verification and a path back to a person. Without the third element, the first two will quietly corrupt the database.

How I've walked this road since 2023

For a PDR company in the US I build parsers for insurance estimates, inspection reports and damage assessments from PDFs, scans and photos, and photo parsers that read VINs. Insurance documents run to 60 to 70 pages. The foundation is Python: OpenCV prepares the image and draws virtual lines on it, Tesseract reads the text, Camelot pulls tables out of PDFs, Pandas cleans and structures. The main work is done by my own logic, which knows how these documents are built.

The parsers have been in use since mid-2023, when about half the documents went through with no manual edits. In 2024 I switched to new recognition methods, and about 75% went through without edits. In 2025 I added new extracted fields and calculations, and the share grew to about 95%. In 2026 I connected AI validators, three different services, Anthropic, OpenAI and Gemini, each for its own task, and shipped the release version: about 99.9% of documents with no manual edits. A person does the final check. The parsers free up about 60 hours of manual work a week.

The typed answer with a confidence score and null in Clairic is exactly what I was missing in 2023, when half the documents went to manual editing with no “I'm not sure here” signal. The scheme at Sullivan & Cromwell, where some agents propose and others challenge, is in essence the same as my validators, only in my case the challengers are different services with different tasks. I haven't tried Clairic on my own estimates and won't promise anything on its behalf.

What I would do

1. Collect 30 to 50 real samples of each document type and measure how many minutes one takes by hand.
2. Write down the fields you need downstream and where they go: the CRM, the estimate, the report.
3. Where the structure is stable, extract with code; bring in a model for the remainder. Every value should have a confidence score, and every document a path back to a person.
4. Verify with a second tool, separate from the one that extracted. Log the discrepancies.
5. Count the share of documents without edits every month and expect gradual growth. It took me three years, and I won't promise numbers before looking at the samples.

Related reading

On where AI helps parse documents and how to check the result: the article AI in data processing: extraction and verification. Service areas: Python for documents and repeatable processing and AI in a process where the result can be checked. Cases: parsers for estimates, reports and VIN photos and technical partner to a PDR company in the US. On the same topic: AI decides, a person confirms, where to store 70-page scans and car photos and what happened at our company when the parsers freed up 60 hours a week. Earlier I wrote about how to keep the meaning of tables and charts when searching PDFs. If your documents are still being retyped into a spreadsheet, tell me which documents they are and where the data should go: Discuss your task.