Contract data extraction, on a model your firm owns
Your contracts hold the answers your team keeps re-finding by hand: renewal dates, notice periods, governing law, who signed what and when. The work is repetitive, it doesn't scale, and it goes wrong quietly. A missed auto-renewal is discovered months late.
AI is obviously suited to this. The obstacle is not capability. It's that the fastest way to apply it means sending client documents to somebody else's server, which is a confidentiality question your firm has to answer, in writing, to people who will not accept "the vendor says it's fine."
So build the model into your own building instead.
What it does
We fine-tune an open-weight model on your contracts, so it learns your paper: your templates, your clause vocabulary, the way your firm actually writes and marks up an agreement. The result extracts structured data (dates, parties, terms, obligations) from documents it has never seen before.
It runs on hardware you control. No API key, no per-document fee, no outbound connection. It can run fully air-gapped, with the network cable physically unplugged, and still work.
The accuracy question, answered with a number
Most vendors answer this with a demo. Here is ours, on CUAD: a public benchmark of 510 real commercial contracts annotated under the supervision of practising lawyers. We trained on 396 and tested on 99 the model had never seen, three separate times.
| Run | Off-the-shelf model | Fine-tuned on the contracts | Gain |
|---|---|---|---|
| 1 | 62.2% | 85.4% | +23.2 |
| 2 | 62.2% | 84.6% | +22.5 |
| 3 | 62.2% | 83.5% | +21.3 |
A general-purpose model reads these contracts correctly about 62% of the time. Trained on the contracts themselves, 83–85%. The same result three times, so it isn't luck.
What that number does and doesn't cover. It measures short, value-like fields: document name, agreement date, effective date. Locating a full clause span (a governing-law or notice-period paragraph) is a genuinely different task, needs a different metric, and we don't claim a number for it yet.
CUAD is public SEC-filed contracts, not yours. Your paper has different formats and house style. What the benchmark establishes is that the method works and roughly how much data it needs. The assessment measures your corpus.
That table is one model. We have since run 48 measured experiments across six open-weight models, three runs each, and every one improved. The best finished at 86.0% on the same short-field extraction task. All of the numbers are published, including the ones that went against what we predicted.
And the part worth more than the result: along the way the same pipeline produced a 72-point gain, a 45-point gain and an 88-point gain that were all false. We caught them, threw them out, and published exactly what each one actually was. If a vendor has never had to retract a number, ask how hard they have looked.
Why "we host it securely" isn't the same answer
A vendor hosting your documents securely is still a third party holding your documents. That may be acceptable to your firm, and for many it is, but it is a disclosure decision, and it has to be made, documented, and defended.
A model running inside your own network isn't a smaller version of that decision. It removes the decision. There is no transmission, no third-party sub-processor, no retention policy to audit, and nothing to renegotiate when the vendor changes their terms or their pricing.
The question to put to any AI vendor: "Can this run with the network cable unplugged, and can you prove it?" If the answer is anything but a clear yes, your documents are leaving the building.
What your firm ends up owning
- The model weights and adapters, not licensed, not rented. They keep working whether or not we do.
- The structured dataset built from your contracts, versioned and documented. It outlives any single model and is the foundation for whatever you build next.
- The evaluation harness, so you can re-measure accuracy yourself, later, without taking anyone's word for it.
- The pipeline and documentation to rebuild or retrain as your paper evolves.
We build on permissively licensed open-weight models (Apache-2.0 and similar) specifically so that ownership isn't complicated by a third party's terms.
How it starts
Not with a purchase order. It starts with a free readiness scorecard you run yourself, on your own machine. It opens no network connections and we never see your documents. It tells you what shape your contract data is in and whether a model could learn from it.
If the answer is promising, the next step is a $5,000 fixed-price assessment that measures what's actually achievable on your corpus, credited toward the build if you proceed within 90 days. If the answer is "your data isn't ready yet," we tell you that, and you keep the report.
How much of your data do you need? Less than most firms assume. In our measurements, 48 training examples changed nothing measurable; roughly 400 produced the 22-point gain above. Consistency matters more than volume: whether similar contracts were handled similarly. Most firms have far more than 400 contracts sitting in a document management system already.
Who this fits
- Law firms with a large contract book and a confidentiality obligation that rules out cloud AI.
- In-house legal and contract management teams tracking obligations, renewals and notice periods across hundreds or thousands of agreements.
- Regulated organisations in healthcare, finance, defense and government, where the document set is sensitive enough that where it's processed is a compliance question, not a preference.
It fits less well if you have very few contracts, if they're wildly inconsistent with no house style, or if what you actually need is search rather than extraction. We'll tell you that at the assessment rather than after the build.