Why “AI for contracts” fails without your clean contract data
Everyone is racing to point a language model at the pile of contracts in their drive. We believe that the winning approach to this opportunity is to first turn the pile into structured contract data.
There is a version of AI contract review that demos beautifully, but actually fails. You upload a contract, ask a question, and get a fluent answer in seconds. It feels like magic the first time. Then you run it across 200 agreements, in 5 different formats, signed over 5 years. The magic falls apart. The model misses a renewal buried in an amendment. It confidently summarizes a clause that was superseded two versions ago. It cannot tell you which of your contracts share the same risky indemnity language, because it has never seen them as a set.
The problem is not the model. The problem is that you handed it unstructured text and asked it to behave like structured data. Most teams discover this gap only after they have bought something. It is worth understanding before you do.
What “unstructured” actually costs you
A contract is not really a document. It is a bespoke set of obligations, dates, parties, amounts, and conditions wrapped in thousands of phrases and numbers. When you treat it as prose, you can read it. Lawyers understand the nuances in it. Hardly anyone can operate on it. You cannot sort every agreement by renewal date, flag every auto-renew clause across the portfolio, or answer “what is our total committed spend next year” without opening each file and rereading it.
Large language models are extraordinary. They are not, on their own, a system of record. Ask a model to read one contract, and it shines. Ask it to be the single source of truth for what your company is bound to, and you are asking a reader to do the job of a database.
Structure is what turns reading into knowing
Structured data is simply the contract’s important facts pulled into consistent data fields: counterparty, effective date, term, renewal type, notice window, value, governing law, and key obligations. Once those facts live in fields rather than paragraphs, three things become possible that were impossible before.
First, you see the whole set at once. Every renewal in the next ninety days. Every contract has a particular clause. Every agreement over a certain value. The portfolio stops being a folder and becomes a view.
Second, the AI gets dramatically more reliable. A model reasoning over clean fields makes fewer mistakes than a model re-reading raw text every time, because the hard part, extraction, happened once and got checked, instead of being guessed at on every search or question.
Third, nothing falls through. Renewals, notice deadlines, and escalations stop depending on whether someone remembered to read the right clause. The date exists as data, so it can trigger an alert.
Why so many tools skip this step
Extraction is the unglamorous part. It is faster to ship a chat box over a document store and call it AI contract review. It demos well, and it is cheap to build. But it pushes the hard problem onto the customer, who discovers months later that the tool can answer questions about a contract they already found, and is no help at all with the contracts they have forgotten they have.
The forgotten contracts are the whole point. The risk in a portfolio is almost never in the agreement you are actively looking at. It is in the one you are not. That is the blind spot, and it is exactly the thing that costs you.
The risk in a portfolio is never in the contract you’re looking at. It’s in the one you’re not.
What good looks like
A contract intelligence system should do the boring work first. Read every agreement, extract the facts into a consistent structure, let a human confirm anything ambiguous, and only then put an AI layer on top for questions and analysis. The order matters. Structure first, intelligence second. Reverse it and you get a clever assistant sitting on top of a mess.
At Librari we built it in that order on purpose, and we kept it where the contracts already live, inside Google Drive, so the structure forms around your existing files instead of asking you to migrate into yet another system. The goal is not a smarter search box. It is to make everything you have signed legible, so decisions get made on what is true rather than on what someone happened to remember. You connect once, nothing moves, and you know what you have by the next day.
The takeaway
If you are evaluating AI for contracts, ask one question early: does this turn my contracts into structured data, or does it just read them on demand? The first is a system you can run a business on. The second is a demo. They look identical for the first ten minutes and nothing alike by month two.