Keywordsmanaging a team that uses aiBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searchesmanaging a team that uses aiai review standards for managershow managers should use chatgptai productivity without more pagessetting quality bars for ai at workhow to use managing a team that uses ai
The job is the standard, not the tool
People will use AI whether you publish a policy or not. The managerial question is what you will accept as finished work. If the standard is “a document exists,” models will fill the week with fluent pages. If the standard is “a named person verified the claims and will own the outcome,” models become a drafting layer and review becomes the craft you are actually managing.
This is not a technology speech. It is the same work managers already do around quality, utilisation and development, now happening in smaller increments every week. Recovered hours appear on Thursday, not in the annual headcount round. If you refuse to talk about them, individuals will still decide how much effort to put into prompts and how much of the gain to reveal.
You do not need to be the best prompt engineer on the team. You do need to recognise a verified artefact, a plausible one, and a paste that should never have happened.
Say what good looks like
Write the bar down. For a weekly pack: every number in a source extract, every risk with an owner, every decision binary or multiple-choice. For a client email: no invented dates, no apology unless you have accepted fault, a human read before send. For analysis: facts, interpretations and hypotheses in separate lists. The model can be asked to meet that bar. The manager is the one who notices when the team has quietly dropped it because the draft looked senior.
| If you reward | You will get | Better reward |
|---|---|---|
| Speed of first draft | Volume, then a review bottleneck | Time to a decision-ready artefact, including review |
| Length and polish | More slides than anyone can compare | A short, sourced recommendation |
| Private heroics | One person who “is good with ChatGPT” | A shared prompt and checklist a new joiner can find |
| Hours saved alone | Saved hours filled with extra drafting | Hours saved and errors caught, plus a use for the surplus |
Review is a designed step
Junior staff who once produced slow, cautious drafts may now produce fluent ones that look senior. The skill required of the reviewer has gone up, not down. Fluency conceals gaps that a clumsy draft used to advertise. Schedule review as a first-class step. Teach people to ask the model for citations, inferred claims, and a short list of likely errors before a human starts reading. Do not cut review in proportion to the drop in typing time. That is how you ship faster mistakes.
One-to-ones should include a look at how the person is using AI, not as surveillance, as craft. Which prompt earned its place. What the model missed. What they added. Whether any material crossed a classification line. People hide tools when the only message from above is fear. They hide quality problems when the only message is “use AI more.”
Do not judge someone’s performance from a model’s summary of their tickets, emails or calendar. That is a characterisation generated from incomplete traces. It is not evidence.
Agree what happens to recovered time
When a pack that took six hours now takes three, someone captures the surplus. If you immediately load three more packs onto the same person, you capture volume and they capture fatigue. If the team uses the hours for the two accounts that actually move the forecast, both can capture value. Say the split out loud: how much returns as capacity, how much is reinvested in quality, how much remains professional discretion. Ambiguity here is how unofficial slowdowns and unofficial overwork both appear.
What to do this week
Pick one repeating artefact your team produces. Write the quality bar on a single page. Ask two people to run the next cycle with AI and to bring the prompt, the checklist, and a log of review time. Decide the surplus rule before the second cycle. If you cannot name the human who is accountable when the output is wrong, you are not managing AI at work. You are managing a way of producing text without an owner.