You will be able to catch the errors AI meeting summaries typically make before they reach colleagues or clients.
Before an AI meeting summary goes to anyone, read it against the transcript and against your own memory, beginning with the action items. A neat layout makes a summary quicker to read and quicker to check, but the layout says nothing about whether the content is correct.
Take Jun Wei, whose summary from lesson 7.2 seemed ready to go. It was short and split into four sections. Every task had an owner, and each point carried a timestamp. She came close to sending it to the client as it stood. Instead she opened the transcript at the timestamp beside the first decision. There, the client's HR manager had said, "Maybe we could look at moving it to July, I'll need to check with my director." The summary had turned this into "Decision: the workshop moves to July." Had she sent it, the client's director would have learned of a decision that nobody had actually made.
Errors in a summary have two sources. The assistant can introduce them while it condenses the meeting, or they can be sitting in the transcript already and pass straight through.
Among the first kind, the one you will meet most often is the suggestion turned into a decision. Someone floats an idea or says "we could", and the summary records it as agreed. Asking an assistant to list decisions sets it searching for them, and hesitant wording reads easily as consent. The mistake can run the other way too, with a real decision written up as a discussion point.
Another is the dropped disagreement. Plenty of meetings close with a difference still open: the sales head wants one thing, finance wants another. A summary may tidy this away by keeping only the view voiced last or voiced most, so the other position vanishes from the record. Along with it goes any reason for the group to return to the question.
Recall from lesson 7.1 that transcription errors gather around names, numbers, accents and local terms. Whatever errors the transcript contains, a summary built from it will contain too. In the transcript they sat in messy text. In the summary they appear inside clean sentences that seem more trustworthy than the original.
Because of that, they rarely stand out. You might see a client's name spelt as the transcription service heard it, a date that has slipped by a week, or "fifty" where someone said "fifteen".
Work through the summary one item at a time and take the action items first, since those drive what people go and do. Open each one at its timestamp and read or listen to the actual words. Put three questions to it. Was it agreed, or merely suggested? Did the named owner really take it on? Was the deadline spoken in the meeting, or did the assistant infer it?
After that, set the summary beside your own recollection. Anything that sits wrongly with what you remember deserves a second look. If something important is missing, say a promise you made or a concern the client raised, locate it in the transcript and put it in. So the person doing the check needs to have attended the meeting. A colleague who was absent can compare summary and transcript, but has no way of knowing what the transcript failed to capture.
Then make a separate pass for names, numbers, dates and money. With each one, ask yourself whether you actually know it to be correct, because an item that merely seems plausible is the kind that slips through. Wherever possible, confirm it from a source other than the transcript: the meeting invitation for names, the proposal for prices, the calendar for dates. Since the transcript produced the error in the first place, it may not be the right place to verify the fix.
If the meeting was internal, you can stop at this point. A careful check by you is usually sufficient.
Anything bound for a client or another outside party needs two more steps.
The first is that a person reads the complete summary before it goes out, on every occasion. Emailing an automatic summary to everyone who attended, right after a client call, is a risk your company may never have agreed to take.
The second is taking out whatever was said off the record. Meetings contain asides, candid comments about third parties and remarks people would rather not see in writing. The transcript picks all of these up, and the summary can carry them forward. In Jun Wei's call, the client had said something frank about a colleague. That had no place in a summary returning to the client's company, and she made certain it was left out.
Running the whole routine, Jun Wei turned up three errors: the July "decision" that had only been a suggestion, a "next Friday" deadline that should have read "the Friday after", and a disagreement about the venue that the summary had omitted. She corrected all three. The check took her only a few minutes, against the forty minutes or so she once spent writing notes from scratch.
Now pick a meeting you attended recently, preferably one that has an AI summary, and compare that summary with your own notes. Doing this is the clearest way to see these errors turning up in your own meetings.
Compare an AI summary of a meeting you attended with your own notes and list every item that was wrong, missing or overstated.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).