Use this when: the team has moved past AI curiosity but not into shared standards. People are producing faster drafts. No one can clearly say when a model is appropriate, what evidence counts, who reviews the output, or when another human must be involved. Adoption metrics look fine. Quality is harder to read.
We work in the artefacts the team already owns: client memos, product decisions, briefs, policy drafts, incident reviews, board packs. Not in toy examples. The point is not prompt fluency. It is better work when models are in play, and a clearer signal upward about what AI is actually doing for the team.
Three patterns we keep finding after the AI day.
None of these are technology problems. They are conditions that the introductory event did not address, and that no platform will fix on its own.
People are using models, but no one trusts the answers
Teams are quietly relying on models to surface internal knowledge, but no one is checking whether the underlying sources are current, correct, or authorised. The first time a confidently-wrong answer reaches a client or a customer, trust collapses faster than it built.
The tools are deployed, but the workflow has not changed
Licences are paid. Training was delivered. People keep doing things the old way because the value of the new path is not obvious in their actual work, or because the trust is not yet there. Adoption metrics look fine. The work itself looks the same.
Things get faster, but the team learns nothing from it
Individual task completion gets quicker. Collective sensemaking gets thinner. There is no practice for capturing which AI-assisted decisions actually improved the work. So the team gets quicker at producing the same drafts and slower at improving them.
High AI usage is not the same as good AI use.
Most of the people on your team are using AI partly because everyone else is. Not from conviction, but from the quiet anxiety of looking like the one who is not. Usage rates rise. Confidence rises with them. Quality, judgement, and shared standards do not necessarily follow.
Part of this engagement is helping the team know honestly whether they are using AI well, not just whether they are using it often, and giving the leader something more useful to report upward than an adoption number.
From private hacks to shared standards.
- A simple team rule set for when AI is helpful, when it is risky, and when it is out of bounds, written for your real work, not lifted from a generic policy.
- Review standards for AI-assisted work, including source checks, named human reviewers, and the unresolved trade-off named explicitly before anything is sent.
- A visible handoff pattern between people and tools, so accountability does not quietly disappear into the model.
- A regular cadence for updating the team norms from real examples surfaced by the team itself, not by us.
- One feedback loop that captures which AI-assisted decisions actually improved the work: the small step toward measuring capability, not just usage.
Two examples from the kind of rules teams actually adopt.
Example rule: A model can produce a first-pass synthesis. Any recommendation that changes customer, employee, legal, or financial exposure must include named sources, a human reviewer, and the unresolved trade-off named before it is shared.
Example handoff: When AI-assisted work crosses a team boundary, the receiving team gets the prompt, the sources, and the reviewer's name, not just the polished draft.
The whole rule set is short on purpose. Five to ten lines, agreed by the team, visible in the artefacts they share, and reviewed monthly.