The wrong question
Somebody above you has asked what your team is doing with AI. The instinct is to find something impressive: a dashboard, a chatbot, a demo that makes people nod in a meeting. Resist it. Impressive is a terrible selection criterion, because the impressive use cases are almost always the ones where being wrong is expensive.
The better question is smaller and far more useful. Which pieces of work does your team do over and over, where the first version does not have to be right, and where a human is going to look at it anyway before it leaves the building?
That is the whole audit. Everything else is detail.
In plain English
- Pilot:
- A small trial with one team or one task before rolling anything out widely.
- Use case:
- A specific job you point AI at. Not a tool, not a department. A task with a start and an end.
- Human in the loop:
- A named person who reviews the output before it counts as done.
- Hallucination:
- When an AI states something false with complete confidence. Names, figures and dates are the usual casualties.
The three tests
A task is a genuine candidate if it passes all three.
High volume. It happens weekly or daily, not twice a year. Saving four minutes on something that occurs once a quarter saves you sixteen minutes a year, which is not a project, it is a rounding error.
Low judgment. The work is mostly assembly, formatting, summarising or first-drafting. If the value sits in someone deciding a difficult thing, AI cannot do the valuable part.
Tolerant of a draft. Being roughly right on the first pass is genuinely useful, because the next step is a person editing it. If the first version must be correct, you have not saved anyone anything.
Most teams find their best candidates are dull: meeting notes turned into actions, long documents turned into summaries, first drafts of recurring updates, messy notes tidied into a consistent format, a rough answer turned into a polite reply, one piece of content reshaped into three formats.
If a use case sounds boring when you describe it out loud, that is a good sign. Boring work is repetitive work, and repetitive work is where the hours actually are.
The ones to leave alone
Three categories are worth ruling out on day one, before anyone gets attached to an idea.
Anything needing accountability. Performance ratings, disciplinary notes, hiring decisions, anything that ends up in a legal file. Somebody has to own those with their own name and their own reasoning, and a partly generated document makes that ownership blurry in exactly the wrong situation.
Sensitive people decisions. Redundancy selection, pay, promotion, medical or wellbeing matters. Even where the output would be fine, the message it sends to your team is corrosive, and you will not recover the trust cheaply.
Where being wrong is expensive. Numbers going to a client, regulatory text, safety instructions, anything published under the company name without a second reader. AI is confidently wrong often enough that the cost of one bad output can eat a year of saved minutes.
โ Weak prompt
Prompt
How can my team use AI?
Output
Your team could use AI for automation, data analysis, customer service, content creation, forecasting and knowledge management.
A list of categories, not a decision. You cannot pilot the word automation on Monday morning.
โ Good prompt
Prompt
Here are the six recurring tasks my team of eight does every week, with rough hours for each. For each one, tell me whether it is a good AI candidate using three tests: high volume, low judgment, and tolerant of an imperfect first draft. Say clearly which ones fail and why. Do not suggest new tasks.
Output
Weekly client update: strong candidate. High volume, mostly assembly, always reviewed before sending. Incident write-ups: weak. Low volume, and the judgment is the entire point of the document.
Your actual work, your actual tests, and a verdict per item. This is something you can act on before lunch.
I manage a team of [8] people who do [what your team does].
Here are the recurring tasks we do and roughly how long each takes per week:
[task 1, hours]
[task 2, hours]
[task 3, hours]
Judge each task against three tests:
- High volume (weekly or daily, not occasional)
- Low judgment (mostly assembly, formatting, summarising or first-drafting)
- Tolerant of a rough first draft that a person will edit
Give a verdict of strong, weak or avoid for each, with one sentence of reasoning.
Do not invent tasks I did not list.
Checkpoint
A good candidate is high volume, low judgment and tolerant of a draft. Anything carrying accountability, touching sensitive people decisions, or expensive when wrong stays off the list.
Running the audit without a workshop
You do not need an offsite. Ask three people on your team what they did last week that they would happily never do again. You will have your shortlist in twenty minutes, and it will beat anything a strategy deck produces, because they actually do the work.
Below are answers from my team to the question: what did you do last week
that you would happily never do again?
[paste the answers]
Group these into no more than five themes. For each theme, estimate how many
hours a week it consumes across the team based only on what they said.
Flag any theme where the frustration is about a broken process rather than
about the work being repetitive, because AI will not fix those.
That last instruction matters. A good chunk of what annoys your team is not repetitive drudgery, it is a process that should not exist. Point AI at a broken process and you get a faster broken process.
Do not let a vendor demo choose your use case. The demo is built to look good, not to match your work. Start from the tasks your team named, then look for a tool, in that order.
Write down three, pick two
End the audit with three candidates on one page, each described in a single sentence naming the task, who does it, and how often. Then pick two. Not five. Two candidates that get proper attention beat five that get a fortnight of enthusiasm and then quietly stop.
Here are the two AI use cases I am considering for my team:
- [task, who does it, how often]
- [task, who does it, how often]
For each one, argue the case against it. Specifically:
- What could go wrong if the output is wrong and nobody notices?
- Who would be accountable for that output?
- Is any confidential or personal information involved?
Be blunt. If either use case is a bad idea, say so plainly.
Asking an AI assistant to argue against your own idea is one of the cheapest sanity checks available. Claude, ChatGPT and Gemini will all do it if you ask directly, and the objections you cannot answer are the ones worth taking seriously.
๐ Quiz
Question 1 of 4Which set of characteristics makes a task a genuine AI candidate?