The owner nobody agreed

This is a fictional example. Fernway Group is a made-up UK company, and everyone in it is invented. You can download the meeting notes used here and try it yourself.

Turning rough meeting notes into a clean list of actions is one of the jobs AI does best. It is also where one particular failure is most likely: the tool is asked for a tidy list, the notes aren't tidy, and it closes the gaps by guessing. The guesses look exactly like everything else on the list.

Before

Maya Roberts, Fernway's Office Manager, took notes at the weekly operations sync. Present were Priya Shah (chair), Tom Elliott, Maya and Dan Okafor. The notes were typical: half-sentences, question marks, a couple of things nobody picked up. Among them:

- Holiday cover for August, half the ops team off at some point. Need a
  rota. Nobody picked this up... someone should own it
- Tom mentioned a big renewal (Harlow account) due mid-Aug, wants sign-off
  on a small discount to keep them. Priya: bring the actual numbers to
  next week.
- AOB: summer social, provisionally 22/08, Maya to send a date-check round

Actions (some half-finished)
- Tom: bring Harlow renewal numbers next week
- ???: own the August holiday cover rota (UNASSIGNED)
- ???: book / confirm summer social venue for 22/08 (UNASSIGNED)

Maya's own notes were honest about the gaps. Two actions had no owner, and she marked them that way. So that the summary could be addressed to everyone, she pasted the team list (names and roles) above the notes.

The AI approach

Turn these meeting notes into a clear, professional summary with a table of actions, owners and deadlines that I can send to the team.

[the team list, then the full meeting notes]

What went wrong

The summary looked excellent. Here are three rows from the actions table:

None of those rows matches what was agreed.

• Owners nobody agreed to. Nobody took the holiday rota. The AI gave it to Leah Bennett, the Operations Assistant, who wasn't at the meeting. Leah's name was in the team list Maya pasted above the notes, and "assistant" plus "rota" is a common pairing, so the tool made a likely-sounding match. The venue booking went to Maya, because the notes had her sending a date check; she hadn't agreed to book anything.
• A decision that wasn't made. Tom asked for sign-off on a discount. Priya asked him to bring the numbers first. The summary says "the approved discount". A discussion had become a decision.
• Deadlines nobody set. The only dates in these lines are the social (22/08) and next week's meeting. All three deadlines in the table were invented, and all of them are plausible, which is the problem.
• The honest markers were removed. Maya had written UNASSIGNED twice. The summary had no gaps, because Maya had asked for owners and deadlines and the tool supplied some.

The summary went to the team. Leah assumed she had been given the rota and started building it from memory, without knowing who had already booked leave. Priya, seeing an owner against it, assumed it was in hand. Tom read "approved discount" and began working it into his renewal proposal. It came out at the next sync, a week later, when Priya asked to see the discount numbers she thought she was still waiting for.

This differs from an invented statistic or a fake reference. The AI wasn't reaching for facts outside the notes. It was completing a pattern inside them: an action table normally has an owner and a date in every row, so it gave every row one.

A better approach

Change the instruction so that gaps are allowed, and marked. Then check the output against the notes before anyone else sees it.

Then run a quick check. It takes two minutes against the original notes:

The UNASSIGNED rows are useful. They are the items the meeting failed to settle, and putting them at the top of the summary with a line such as "Can someone take these?" is more likely to get them done than a made-up owner.

Rule to remember