Catching Weak Signals Before They Become Problems
- Amy Westlake
- Apr 28
- 3 min read
Updated: May 25
Situation
I was running a large cross-functional program. Lots of stakeholders. Lots of meetings. And if you've ever managed something that size, you know the dynamic: everyone is professional in the room. Issues get soft-pedaled. Concerns get wrapped in qualifications. And by the time something surfaces as an actual problem, it's been circling for weeks.
I was sitting in meetings, taking notes, tracking action items — doing all the things a good program manager does. And I kept having this nagging feeling that I was missing something. Not because I wasn't paying attention. But because the signal was too quiet to catch in real time.
The AI Move
I started small. I uploaded one meeting transcript to NotebookLM and asked it a question I'd never thought to ask before: What unspoken risks or concerns are present in this conversation?
What came back stopped me.
It didn't just summarize the meeting. It pointed to specific moments — things that had been said in passing, hedged, or buried under other topics — and framed them as signals worth paying attention to. And because it was NotebookLM, it sourced everything. It told me who said it and what the exact language was.
I started adding more transcripts. Eventually I had every meeting transcript from that program in one notebook, regardless of topic. And I built out a small set of prompts I'd run regularly: identify unspoken risks or concerns, identify tensions between stakeholders, flag anything that probably should have been said but wasn't.
The Shift
The obvious shift: I started seeing things earlier.
The less obvious one: I started asking different questions in meetings. Not because I had a playbook, but because I'd already seen the subtext. I knew which thread to pull. I could raise something that had been dancing around the room for two weeks and create space for it to actually be addressed — without it feeling like I was blindsiding anyone.
I can't point to a specific disaster that didn't happen because I did this. That's the nature of early intervention — you don't get to see the counterfactual. But I can say that it changed the quality of my conversations. It made me a better-informed presence in rooms where information was filtered before it reached me.
The Pattern
Meeting transcripts are a record of what was said. That's not the same as what was meant.
When you're in a room, you're managing the conversation in real time — reading body language, tracking who's talking, deciding when to push and when to let something breathe. You can't simultaneously do all of that and also monitor for weak signals across a dozen stakeholders.
AI can. It's not distracted by the live conversation. It reads the transcript cold, without the social pressure to let something slide, and it surfaces what you were too busy to catch in the moment.
The signal was always there. You just needed something that wasn't in the room to find it.
The Implication
You don't need months of transcripts to try this. Start with one meeting — ideally one where you had that vague feeling something was off but couldn't name it.
Upload the transcript. Ask it: What concerns or risks are present in this conversation that weren't explicitly named?
Then ask: Is there anything that probably should have been said here but wasn't?
Read what comes back slowly. You're not looking for a summary. You're looking for the thing that makes you say: hm.
What I'm Testing Next
I'm curious what happens when I run this analysis across a full quarter of transcripts at once — not looking at individual meetings, but asking what patterns have been quietly building across the whole period. What's been circling for months that no single meeting would show me?
Comments