Contract review
against your playbook.

Every legal team has a set of positions: what it will accept on limitation of liability, which governing laws are fine, where it needs a carve-out, what triggers escalation to the general counsel. In most organisations this exists partly in a template, partly in precedent, and substantially in the heads of two or three people.

Extracting that into an explicit, versioned playbook is most of the work in a contract-review deployment, and it is work with value independent of any AI system — teams commonly discover during it that their stated position and their actual practice have diverged. The model then compares an incoming contract against it and reports differences.

Framed this way, the task becomes tractable and evaluable. “Is this contract risky?” is unanswerable and unmeasurable. “Does clause 12 depart from position 7, and how?” has a right answer that a lawyer can confirm or reject, which means you can build an evaluation set and know whether the thing works.

Recall over precision, deliberately. A missed deviation is a risk that reaches signature; an extra flagged one costs a lawyer ten seconds. We tune conservatively and say so, because the alternative failure is invisible until it is expensive.

Every finding is anchored. Clause, page, quoted text. A finding a reviewer cannot immediately verify will be ignored within a fortnight, and once a tool is ignored it is dead regardless of its accuracy.

The evaluation set is built from real contracts your team has already reviewed — with the deviations they found. That gives a genuine baseline, and it usually reveals disagreement between reviewers that is itself worth knowing about before you automate anything.

Confidentiality is structural. Contracts under negotiation are among the most sensitive documents an organisation holds. Self-hosted deployment means they are not transmitted outside your perimeter, which removes an objection that otherwise recurs at every renewal.