Government

For an agency deciding what to do, in what order, with a workforce that is already using these tools whether or not anyone has approved it.

1. Find out what is already happening

   Before any strategy, establish actual current use — anonymously, without sanction, and quickly. Every agency that has done this has found more use than it expected, concentrated in drafting, summarising and translating. The answer changes what you build, and it is the only baseline you will get.

   What good looks like: A number you believe, and three or four use cases named by the people doing them.

2. Give people a sanctioned tool before you write the policy

   A policy without an approved tool is a policy that gets routed around. Provide something adequate and governed inside weeks, even if it is narrow, then write the rules for the thing people are actually using. The reverse order produces a document nobody reads and behaviour nobody sees.

   What good looks like: Access in under a week, and a one-page statement of what may and may not go into it.

3. Draw the line at decisions affecting rights

   Separate assistance from determination and hold it. Drafting, summarising, searching, translating and first-pass triage are assistance. Anything that affects an entitlement, a liability or a right is a determination, and a determination needs a human who is accountable, who has the material, and who could have decided differently. Administrative law does not have an exception for software.

   What good looks like: A written list of what is assisted and what is never automated, approved at executive level.

4. Build an eval set from your own cases

   A hundred real cases with the answers a competent officer would give, run whenever anything changes. This is the artefact that turns a pilot into a decision and a procurement into something defensible. It takes a domain expert about a week.

   What good looks like: A score you can quote in a brief, and a rerun that takes an afternoon.

5. Settle records and privacy before scale, not after

   Where do prompts and outputs live, for how long, and are they public records. Who can see them. What happens to them under an RTI request. Whether the vendor trains on them. These answers are cheap to establish early and expensive to retrofit once there are two years of material.

   What good looks like: A retention position and a vendor clause on training, both in writing.

6. Train for judgement, not for prompts

   Prompt technique is a fortnight of learning that dates within a year. What lasts is knowing when the tool is the wrong instrument, how to check an output efficiently, and what failure looks like in your own domain. Build the capability around the judgement, because that is what transfers.

   What good looks like: Staff who can say why they did not use it for a particular task.

7. Publish something

   Agencies that say publicly what they are using, where, and with what safeguards get better questions, fewer rumours, and a defensible record. Silence reads as concealment even where there is nothing to conceal.

   What good looks like: A page on the website that a journalist could read without finding a surprise.

Reviewed 2026-09-20 · All playbooks