How Resubly Reads a Contract with AI

The Short Version

You drop a contract PDF into Resubly. An AI model reads it and pulls out the fields that matter for a renewal: the vendor, the cost of the current period, the start date, the renewal date, the notice period, and the payment terms. For those fields the model also returns a confidence rating, and Resubly keeps that rating with the extraction. Nothing enters your live renewal list until you look at what it found and confirm it. That review step is not optional and it is not a nag. It is the design.

This post is the honest walkthrough of how that works, including where it is strong, where it is weak, and what happens when it gets something wrong. If you handle contracts for a finance or ops team, you should know exactly what a tool does with them before you upload the first one.

Step One: You Upload, We Extract

When a PDF lands in the drop zone, Resubly sends the document to a large language model (currently Google Gemini, reached through an AI gateway) with a strict instruction: return only the renewal-relevant fields as structured data, plus a confidence rating for each field it scores. The model does not get to freewheel. If its answer cannot be parsed as JSON, the extraction fails, and a failed extraction marks the document as failed for manual entry rather than guessing. Values that come back missing or malformed are simply left out, and you fill them in during review.

The fields it targets are the ones a renewal decision actually turns on: vendor name, the cost of the current period, currency, start date, renewal date, notice-period days, and the vendor domain. On an order form that covers several years, that cost is the first period's amount, never the grand total across all periods. Not the whole contract. Resubly is not trying to be a legal review tool. It is trying to answer one question: when does this renew, and by when do you have to act.

Step Two: Confidence Ratings Are Kept, Not Thrown Away

Resubly asks the model to include a confidence rating with each field it scores, and it stores those ratings alongside the extracted data instead of throwing them away after the parse.

Those ratings are not decoration. They are saved with the extraction as metadata, so the record of how sure the model said it was survives next to the values themselves.

Step Three: Nothing Is Trusted Until You Confirm

An extracted contract does not appear in your renewal list. It appears in a review state. The fields the AI pulled sit in a review form for you to work through. You correct anything wrong, fill anything missing, and confirm. Only then does it become a tracked subscription with a real renewal date and a real notice deadline.

This matters because the cost of a silent extraction error in this domain is not a typo. It is a renewal date that is thirty days off, which means a notice window that closes before you knew it opened. A tool that auto-imports contracts and trusts its own output is optimising for a demo, not for the one job that has to be right. Resubly would rather cost you two minutes of review than let a wrong date sit unexamined until the invoice arrives.

What It Is Good At, and What It Is Not

AI extraction is genuinely strong on clean, text-based contracts with labelled terms. It runs the same careful pass on the fortieth contract as on the first. It does not get bored.

Documents with no usable text layer, and contracts where the renewal terms live in an addendum, are exactly the ones where the first pass can come back with fields missing or empty. That is precisely why the confidence score and the human confirmation exist. The model is a fast first pass, not a final authority. You are the final authority, and the tool is built to keep it that way.

What Happens to Your Document

The document is sent to the AI model for extraction and the field values come back. The uploaded file is stored against your organisation and the subscription it belongs to, and it is deleted when you delete the contract or close your account. The security page lists every provider that touches your data and the full retention and deletion lifecycle.

Why We Built It This Way

The whole point of Resubly is to turn renewals from surprises into decisions. An extraction pipeline that quietly guesses would reintroduce exactly the surprise we are trying to remove. So the design rule is simple: the AI does the reading, nothing becomes a tracked renewal until a person has looked at it, and a person signs off before a date can ever trigger an alert. Fast where it can be, careful where it must be.

If you want to see the wider system the extraction feeds into, read Inside the Renewal Machine, or start dropping in your own contracts and watch the review step do its job.

Sources: Resubly product architecture (contract ingestion pipeline, per-field confidence scoring, and review-before-commit flow), 2026.