Document Intelligence reads a contract and tells you what's in it. This one argues back. Paste the counterparty's clauses on the left and your negotiating playbook on the right. One structured-output call rates each clause against your positions and drafts three fallbacks for every problem term — the language you'd send if you had all the leverage, the language you'd actually settle on, and the line you don't cross. Plus what to trade for what, and the order to take them in.
Browser
└─→ POST /api/lab/chat
- system: redline prompt + schema
(versioned: contract-redline.v1)
- user: { clause_text, playbook_positions }
- temperature: 0.15 (very low — drafted legal
language must be stable and literal)
- max_tokens: 7600 (three drafted positions per
clause is the token-heavy part; 4096
truncated mid-object in testing)
← single response, parsed as JSON: {
document_context, overall_risk,
executive_summary, deal_breakers[],
clauses[ { original_text, risk,
playbook_status, issue,
business_impact,
positions{ ideal, acceptable,
walk_away },
negotiation_note } ],
trade_analysis[], recommended_sequence[]
}
← each clause renders as a card; the three
positions are a 3-col grid, colour-coded by
how much ground each one concedes
One call does four jobs that are usually four passes: segment
the text into discrete clauses, classify each against the
playbook (within / outside /
not_covered), rate the residual risk, and
draft replacement language at three levels of
concession. The third of those is what makes it a negotiator
rather than a reviewer — and not_covered is quietly
the most useful status, because it surfaces the terms your
playbook never anticipated.
First-pass contract review is the most heavily leveraged work in legal services. An associate or a junior in-house lawyer reads the counterparty's paper, marks what deviates from standard positions, and drafts alternatives. It is high-volume, highly patterned and almost entirely playbook-driven — which is exactly the shape of work where structured generation earns its place.
The output here is deliberately not a single "correct" redline. Real negotiation is a ladder: you open at your ideal, you expect to land near the middle, and you need to know in advance where your floor is so you don't discover it live on a call. Producing all three at once turns a review into a negotiating brief.
The trade analysis is the part clients actually pay for. Knowing that clause 11 is bad is cheap; knowing that conceding on audit frequency is what buys you a liability cap is the judgement. The model is not reliably good at that yet — but it is good at proposing candidate trades a lawyer can accept or discard in seconds.
Honest caveat: this drafts language; it does not advise. It has no view of the commercial relationship, the governing law's treatment of the terms it is rewriting, or the fifty pages of the agreement you did not paste — and a limitation-of-liability clause cannot be assessed in isolation from the indemnities and the insurance schedule. Treat every output as a first draft for a qualified lawyer to accept, edit or reject. The value is removing the blank page, not removing the lawyer.
Run the demo to see telemetry.