I ran six real meetings through my transcript skill. It refused all six.

A talk show host points at the audience, captioned 'You get a meeting! And you get a meeting!'

What you'll learn:

  • How to decide which meetings your AI should learn from, and which stay out

  • How to rank your sources so the AI answers from verified information first

  • Why a transcript should never count as an answer until a person confirms it

I built a skill that turns meeting transcripts into decisions and updates for our knowledge base. Then I tested it on six real transcripts from my own tools.

It stopped on all six.

A 1:1. Calls with people outside the company. A two-person Legal session. Not one of them should end up in a company knowledge base, and the skill caught them all. So it passed round one of testing.

In this article, I’ll briefly cover what I’m thinking about and building when it comes to incorporating meetings into your AI context layer. The full article is in the Field Guide.

The week the tool decision got interesting

While we were testing transcription tools, Claude launched Meetings.

It would be easy to call that bad timing. I think it was a useful reminder of what we were actually doing. We were building a context layer: the information AI needs to answer questions without us answering them twice.

Our code base is one source. AI can read it, so it can tell me what our product can and can't do. Now, I don't have to pull in Product and Engineering for every question. A lot of it is self-serve with AI’s help, which they engineered that way.

But what is one of the biggest sources of missing information and context within an organization? Meetings. Too often, information stops in the meeting and doesn't leave the room.

Which meetings should AI learn from?

We chose transparency as the default.

  • Captured: weekly team meetings, all-hands and similar recurring internal calls.

  • Not captured: 1:1s, HR meetings, Legal meetings, calls with outside counsel and any call with someone outside the company unless we have their consent.

Exceptions are deliberate, and we write them down.

A transcript is not an answer

Here's the rule that matters most. Transcripts feed drafts. They never answer questions on their own.

The skill pulls out two things only:

  1. decisions and

  2. updates.

Every entry cites the exact line it came from and requires human confirmation.

An invented example.

In a weekly support meeting, one manager says she's raising the refund limit from DKK 500 to DKK 1,000. The skill writes it up with her name, the meeting, the line number and her exact words. Then she gets a Slack message with three options: confirm, correct or reject. Until she replies, it isn't a decision recorded in the database. Silence is never confirmation.

What happens when a transcript contradicts an approved answer?

When sources disagree, I rank them by how verified they are:

  1. Code answers what the product can and can't do.

  2. Approved entries in the database answer what we say.

  3. Confirmed meeting entries come next.

  4. Slack and raw transcripts are last. They can point to something. They can't answer definitively.

A lower source never overrides a higher one. If a transcript contradicts an approved position, the AI answers from the approved position and flags it for review.

What I'd do first

Write the problem in one sentence before you pick a tool. Ours: information stops in meetings and doesn't reach the AI.

FAQ

Do we need video recordings of meetings for AI?

No. Text is what AI reads, what people search and what someone can check. Transcripts are enough.

Which meetings should a legal team keep out of transcription?

1:1s, HR meetings, Legal meetings, calls with outside counsel and any call with someone outside the company.

Can AI answer questions from a raw meeting transcript?

No. Transcripts feed drafts. Every entry cites the exact line it came from, and a person confirms anything that changes how we work before it counts.

Which source should AI trust first?

Code first, then approved entries, then confirmed meeting entries. Slack and raw transcripts come last. They can point to something, but they never answer on their own.

If you want the actual thing

That's the what and the why. The how is in The Field Guide, for paid members:

  • The skill with the source ranking, the confirmation flow and the never-process list built in

  • A setup guide PDF for adding your approver and your own exclusions

  • The example database layout, with invented entries you can copy

  • The full post, with the line-numbered example and the Slack message

Put it on the company card. Use code EXPENSEIT for 25% off the monthly plan for as long as you stay subscribed. Join by 31 Dec 2026a here.

Speak soon!

Field Notes is the what and the why. The Field Guide is the how — downloadable skills, explanations, lessons from what didn’t work, and a monthly video Q&A.

Get the full version
Next
Next

NDAs Are the Best Contracting Function to Automate First