Nutrient

If your AI agents get even 10% of its data extraction wrong, can you really trust it at all? We discuss why reliability is so important when working with your documents.

Most AI tools handle 90% of documents fine. The last 10% is scanned pages, messy formats and tracked changes, and that’s where you stop trusting the answer. We sat down with Nick Winder to talk about NuDocs, the AI-native tool built for that last 10%. Try it on macOS: nudocs.ai

Fast answers, straight from the document. Watch the mailroom cat make a decision with Nutrient Data Extraction.

Every mailroom needs a good hire. This one sorts documents before you finish your coffee. 🐾📑 Nutrient’s classify tooling sorts incoming documents by type so the right document reaches the right team, without anyone sorting by hand.

What is the Nutrient Decision Model, and why would you ever use it over a large language model? Easy answer: cost and speed

Most questions you ask about a document have a short answer. Is this an invoice or a contract? Is this extracted value actually on the page? Where does one invoice end and the next begin? Yet the usual approach is to send the whole file to a general-purpose LLM and wait while it writes a reply, one token at a time. So we built something different: the Nutrient document decision model. It reads a document, both the page image and the text, and returns a choice plus a calibrated confidence score. No generated paragraphs, no waiting. What that gets you: → About 200 ms per answer → Scores calibrated to how often the model is actually right → Visual understanding that text-only approaches miss It’s already live in Nutrient DWS, powering our classify and split endpoints and checking extraction results against the source document. Nutrient AI engineer, Austin Nguyen, explains why we built it, how we trained it, and where it falls short: bit.ly/4jpIZyF Try the demo on Hugging Face: bit.ly/4i8tEjE

Story time: we gave our extraction a four-page California small claims form, and it came back with the currency as "USD," even though those three letters appear nowhere in the document. The form is a real court form filled in with test data, and we'd asked for six things: - The plaintiff - The defendant - Amount claimed - Currency - Date of the incident - Reason for the claim Five of them were right there on the page. The currency wasn't, but our setup mentioned that US court forms always use dollars, so the model did the sensible thing and filled in USD. It wasn't really wrong, but it was a guess rather than something read off the form, and in the extracted data it looked exactly like the five values that were. If you're automating claims, contracts or court filings, that's the kind of answer that slips through until someone has to explain where it came from. We caught it because every value comes back with a citation showing the page it came from and the exact spot on that page. The currency pointed to a different page from the amount it belonged to and scored noticeably lower, so a review step would send that one field to a person and let everything else through. And that's really what we want from AI on such documents. It doesn't have to be perfect, but it does need to alert a human when it is not exceptionally confident in a result. Remember, you are responsible for every document you submit - whether AI assisted or not, so you better be sure that it is correct. Don't play AI roulette. See the full walkthrough: bit.ly/4zwT8OJ

A picture of a signature on a PDF doesn't tell anyone whether the contract changed after it was signed. A digital signature does, because any later change to the signed content shows up when the file is checked. How to add one with Nutrient API: bit.ly/4s1GowY

An expense report can spend far more time waiting in inboxes than it ever spends being reviewed. It goes to a manager, then finance, then a VP if it's over $10,000, and every handoff is another inbox where it can sit over a weekend while the contract waits to be paid. Nutrient Workflow sends it to the right person the moment it's submitted, nudges reviewers automatically, moves it to a backup approver if someone doesn't act in time, and pushes the approved numbers into your finance system so nobody retypes them. We save you time, and time is money. Expense that here: bit.ly/4yKWayY

It's only a matter of time before a major incident gets blamed on a mistake made by an AI agent, and "the AI did it" is an excuse nobody is going to accept. Each of us is responsible for what our agents produce. That makes auditability more important than it has ever been, especially now that AI is drafting redlines, discharge summaries, and responses that people then sign off on. With Nutrient Document Authoring, each AI suggestion lands as a tracked change with a comment explaining it, so a human is always aware of what is happening, and can approve changes on a case-by-case basis. An instrumental tool in high leverage situations. Call it CYA for the age of AI agents. The AI won't be the one sitting in the incident review when something goes wrong, but you will, so bring the receipts. Learn more: bit.ly/3WeiFxK