Making Meeting Memory Compound — AI That Prepares the Next Agenda for You

Summary: MUSICO — which operates corporate dining and central kitchens for global companies — rebuilt its own meeting workflow with AI. By accumulating meeting notes and having an AI agent read them with full context, minute-taking dropped from 30–60 minutes to near zero, and action-item follow-up from 60 minutes to 10 (final review only).

The Challenge

Running food services across many projects and stakeholders left a "homework" of minutes after every meeting. The real burden was starting from scratch each time — recalling what was discussed, what was deferred, and whose action items remained. Relying on human memory meant things slipped through in the next round of preparation.

Our Approach

Rather than one-off summaries, we designed an accumulating minutes pipeline where the AI builds up and retains context across meetings. The key is deciding what to automate versus what humans judge — and, above all, designing how context is stored. The tools are just means.

What We Built

Record and transcribe meetings, and accumulate the notes in cloud storage without deleting
An AI agent reads across the accumulated notes and builds up context
From that context, it generates the next agenda and action items in advance
Unfinished action items surface automatically at the start of work
Humans give final review to transcription errors and confidential content

Results

Minute-taking: 30–60 min → near zero
Action-item follow-up: 60 min → 10 min (final review only)
Zero dropped items — deferred topics and delegated tasks no longer slip
Less indirect work by design, so people focus on creating "delicious and delightful" moments

This is not "leaving it all to AI" but "building it into the system." How you accumulate context decides the outcome — and we rebuilt our own back-office work first.

The Challenge

What We Built

Record and transcribe meetings, and accumulate the notes in cloud storage without deleting

An AI agent reads across the accumulated notes and builds up context

From that context, it generates the next agenda and action items in advance

Unfinished action items surface automatically at the start of work

Humans give final review to transcription errors and confidential content

Results

Minute-taking: 30–60 min → near zero

Action-item follow-up: 60 min → 10 min (final review only)

Zero dropped items — deferred topics and delegated tasks no longer slip

Less indirect work by design, so people focus on creating "delicious and delightful" moments

This is not "leaving it all to AI" but "building it into the system." How you accumulate context decides the outcome — and we rebuilt our own back-office work first.