Use case · Corporate legal

Policy and playbook assistant

Answers from your own policies, with the citation attached.

The problem

The policy exists. It is forty pages long, it was updated in March, and the version people quote is the one somebody pasted into a deck in 2023.

How we would run it

  1. STAGE 1

    Evidence

    A written picture of the current state, with its gaps named.

    Collect the policies, the real questions people ask about them, and the answers currently given. The gap between the three is the problem.

  2. STAGE 2

    Baseline

    A baseline measurement both sides agree on.

    Test retrieval against those real questions before building anything. If the source material cannot answer them, no model will.

  3. STAGE 3

    Design

    A design your architecture and risk leads have approved.

    Decide the answerable scope, the citation format, and what the assistant must refuse. Refusals are designed, not left to chance.

  4. STAGE 4

    Ship

    Working software or a live process, in production.

    Release to one team with citations on every answer and an obvious route to a human when the answer is not there.

  5. STAGE 5

    Prove

    A result measured against the baseline, and an instrument you keep.

    Report answer quality on the held-out question set, refusal behaviour, and how often people follow the citation.

What you get

  • Curated policy set with ownership and review dates
  • Assistant that answers with citations to the source paragraph
  • Refusal and escalation behaviour, specified
  • Held-out question set with approved answers, for re-testing

What it needs from you

  • Current, approved policy documents — and a named owner for them
  • Real questions from the last quarter
  • Somewhere to put it that people already use

Where the human stays

The assistant answers only from approved policy text and cites it. Where it cannot, it says so and routes to a named person rather than reasoning its way to an answer.

How we would know it worked

  • Answer quality on the held-out question set
  • Share of answers where the citation supports the answer
  • Repeat questions reaching lawyers

Where it starts

Starts as a sprint

Prompt & Workflow Starter Kit

Ten tested prompts for the work your team actually repeats.

Sprint details →

Grows into

AI Build — pilot to production

One use case, built and put into production, with evidence it works.

Program details →
Scope this as a sprint →
Start here

Start with one sprint

Two to three weeks, one approver, a deliverable you keep. Tell us the problem and we'll come back within one business day with a scope, a date and a fee.

Scope a sprint →