Large enterprise

For an organisation with a data estate, a vendor relationship, and a board asking what the strategy is.

1. Pick problems with a measurable unit

   Choose work with a countable output — claims processed, tickets resolved, documents reviewed, calls handled. Anything without a unit produces a pilot that ends in enthusiasm rather than a number, and enthusiasm does not survive a budget cycle.

   What good looks like: A baseline measured before anything is deployed.

2. Assume the model layer is a commodity and design for replacement

   Model quality converges and prices fall. The durable advantage is in your data, your evaluation, and the workflow around the model — not in which provider you chose. Build so the model can be swapped in a fortnight, and treat any architecture that makes switching hard as a cost.

   What good looks like: A second provider tested against the same eval set.

3. Fix the data access problem first

   Most enterprise AI projects fail on retrieval and permissions rather than on the model. If the system can reach documents a given user should not see, that is a breach waiting to be found; if it cannot reach the right ones, it is useless. Permission-aware retrieval is the hard engineering and it is where the time goes.

   What good looks like: A retrieval layer that honours existing access control, tested adversarially.

4. Cost it at full adoption

   Model the consumption cost at realistic usage across the whole workforce, including reasoning models that consume far more tokens per answer, and including the human review time. A pilot that looks free at fifty users can be a serious line item at ten thousand.

   What good looks like: A cost per completed task, not a cost per token.

5. Decide what happens to the time saved

   If a process gets faster, say in advance whether that means fewer people, more throughput, or better quality. Organisations that leave this unstated get quiet resistance from the people asked to implement it, and they deserve to.

   What good looks like: An answer the affected team has heard, in plain words, before rollout.

6. Put one accountable executive on it

   Distributed ownership produces twelve pilots and no capability. The pattern that works is one accountable owner, a small central team that builds the shared plumbing, and business units that build on it.

   What good looks like: A name, and a budget that person controls.

Reviewed 2026-09-20 · All playbooks