When your team uses AI, but not in a way you can rely on

What it looks like#

A client workshop finishes at eleven. By two, three people have each asked an AI tool to write the follow-up. One version promises a date nobody agreed to. One leaves out half of what was discussed. The third is fine, but nobody can say which notes it was written from.

Nobody did anything wrong, exactly. Each of them used the tool the way they’d worked out for themselves.

Signs it’s happening:

• The same request gets different results depending on who asks.
• A prompt that worked well lives in one person’s chat history.
• Checking a draft takes as long as writing it, because nobody knows what it was based on.
• Client or staff details go into tools nobody has looked at properly.
• AI output sits in chat windows instead of with the work.

Why it happens#

AI usually arrives as a personal tool, not as a step in how the work gets done. A step in a process has six things that ad-hoc use doesn’t:

• A trigger. Everyone knows when it runs, and for what.
• Defined inputs. What goes in, where it comes from, and what stays out.
• Instructions written once. Not retyped from memory each time.
• A named check. Someone reads the result against the inputs before it’s used.
• A home for the result. It lands with the work, with the inputs beside it.
• A way to improve. Corrections are recorded, so the same mistake isn’t fixed by hand forever.

Without those, quality depends on who’s asking and how busy they are, and the time AI saves on writing gets spent on checking.

How to set it up#

1. Pick one task that repeats and can be checked

   Good first tasks have their inputs already written down and a person who’ll read the result: meeting notes into a follow-up email and action list, an enquiry into a first reply, notes into a report section, a long document into a one-page summary for a decision.

   Leave for later anything that decides, approves, prices or goes to a client without being read.

2. Write down the inputs

   List what goes in and where it comes from: the meeting notes, the agreed scope, the last email. Then list what stays out: personal details the task doesn’t need, anything confidential, anything your client agreements or privacy obligations rule out.

   A template for the inputs, like meeting notes with the same headings every time, often does more for the result than longer instructions.

3. Write the instructions once

   A short brief: the task, the format, what to do when information is missing, and what it must never do. ‘If the notes don’t say, write [CHECK]’ is worth more than any clever phrasing. Give the brief a version number and keep it where the team can find it.

4. Test it on past examples

   Run the brief on five or so past examples where you know what a good result looks like. Adjust it until the drafts need only light edits. Keep those examples: they’re how you’ll test the next change.

5. Name who checks, and against what

   The checker reads the result against the inputs, not against their memory of the meeting. Their name and the date go with the result. For anything leaving the business, the check happens before it’s sent.

6. Decide where the result lives

   In the client file, the CRM or the job record, with the inputs beside it, so anyone can see what a draft was based on. Not in someone’s chat history.

7. Keep a corrections log, and review it monthly

   Each time the checker fixes something, note it in a line. When the same correction comes up twice, change the brief. After a month, decide whether to keep the step, change it or stop.

Where AI fits, and where it doesn’t#

Before client or staff information goes into any AI tool, find out what kind of account it is, where the data is kept, whether it’s used to train models, and who can see the history. Use a business account with settings someone has chosen, not a personal one.

You are drafting a follow-up email after a client meeting.

Use only the meeting notes and the agreed scope below.
Do not add commitments, dates, prices or advice that are
not in them. If something is unclear or missing, write
[CHECK: what is missing] instead of guessing.

Write:
- one line thanking them for their time
- what was agreed, as a short list
- who is doing what, and by when, as a list
- open questions, as a list

Keep it under 200 words, in plain Australian English.

After the email, list anything from the notes you left
out, and why.

Meeting notes:
[paste here]

Agreed scope:
[paste here]

Worked example: one follow-up email, done the same way#

Start this week#

1. Ask the team which tasks they already use AI for. You’ll probably find more than you expected.
2. Pick one that repeats and has written inputs.
3. Fill in the task card: inputs, instructions, checker, and where the result goes.
4. Test the instructions on five past examples.
5. Start a corrections log, and book a review for a month’s time.

Template: AI task card#

When a task card stops being enough#

1. Several people run the same task every dayA shared tool or form that holds the approved instructions, so nobody pastes their own version.
2. The inputs already live in your systemsA connection that pulls the notes or records in, so nothing is copied into a chat by hand.
3. Results need to land with the workDrafts saved straight into the client file or CRM, waiting for the checker, with the inputs attached.
4. You need to know it’s still workingA log of inputs, drafts and corrections, and a regular sample check, especially after the tool changes.

What this won’t do#

• It won’t make AI right without checking. It makes checking possible.
• It won’t fix a task nobody has defined. If the inputs are a mess, start there. Guides 01 and 02 cover capture.
• It isn’t legal or privacy advice. What you can put into which tool depends on your obligations and agreements.