How the AI works
Why a model at all
A rules engine can find the word “renew”. It cannot find an auto-renewal written as “shall continue for successive periods of equal length unless”, and it cannot tell you whether a ninety day notice window is aggressive for this kind of agreement, because that judgement lives in the thousands of agreements the model has read and in a corpus of standard terms, not in a regular expression.
The second half is harder. Explaining a liability cap to somebody who runs a dental practice, in four sentences, with the consequence in dollars, is a writing problem. That is the part customers pay for and it is the part only a language model does.
So the model is the engine. Everything else in this product is plumbing around it: getting the text out of a PDF, keeping a long agreement in one pass, and refusing to show you a quote that is not in your document.
The four steps
The whole agreement, in one pass where it fits, with a system prompt that names the clause families that actually cost small businesses money: auto-renewal, termination, liability, indemnity, payment, scope, intellectual property, exclusivity, assignment and forum. The model is told to quote verbatim and to leave out anything it cannot quote exactly.
For every flagged clause the model states the norm for that agreement type with a number in it, and says where this clause sits against it. A clause that is genuinely normal is rated low and said to be normal, because a reviewer that flags everything is a reviewer nobody reads twice.
Two to four sentences, naming the consequence, using the numbers from the contract. No Latin, no shall, no hedging. The test we hold it to: could the reader raise this on a phone call without rereading it.
The ask in one sentence, then the replacement wording, pitched at what a smaller party can realistically win. This is a separate routed step because drafting and judging are different jobs and may not want the same model for long.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the review engine reads, so if it is wrong here it is wrong in production.
| Step | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Read the contract and flag the risky clauses | gpt-4.1-mini | openai | balanced | $1.60 |
| Draft the change to ask for | gpt-4.1-mini | openai | balanced | $1.60 |
| Write the overall risk read | gpt-4.1-mini | openai | balanced | $1.60 |
| Work out what kind of agreement this is | gpt-4.1-nano | openai | fast | $0.40 |
Above roughly 26,000 characters, about fifty pages, the read escalates to the frontier model rather than the balanced one, because that is where long-document accuracy starts to matter more than cost. Reviews are capped at 48,000 characters in one pass, and anything past that is reported rather than dropped silently.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a bad afternoon should not cost a customer their review, and it does not.
Every quote is located in your own document before it is shown, and the structure is validated against a schema. A quote that does not appear is dropped and counted. This is the only part of the pipeline the model does not get a vote on.
Candidates
Every model below sits behind the same interface. Adding a lab is one case in one file, which is the entire point of building it this way.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
In this deployment the OpenAI and Anthropic adapters are written and the OpenAI one is live. The rest are declared with their real model names and switch on when their key is present. No model is trained on customer contracts, by us or by our providers.