The Most Overlooked Innovation Opportunity Is the Live Conversation
- 1 hour ago
- 3 min read
By Josh Torrey
Most of the money in AI right now is being spent to help people remember things that already happened.
Look at the meeting software market. Gong, Otter, Fireflies, the long list of post call summarizers. Billions of dollars and a lot of clever engineering, all pointed at one job: record the conversation, transcribe it, and tell you afterward what you said and what you missed. It is a genuinely useful job. It is also the wrong moment of impact.
By the time the summary lands in your inbox, the deal has moved, the objection went unanswered, and the budget number you should have pushed back on is sitting in the notes as agreed. The leverage was never in the recap. It was in the thirty minutes while the decision makers were online together, the decision itself was still open, and almost nobody is building for those thirty minutes.
That is the opening I keep coming back to. The most overlooked opportunity in AI is not a new model or a new modality. It is a shift in timing and presence.
We have spent three years teaching machines to be excellent historians of our conversations. The harder and more valuable problem is making them useful participants while the conversation is still happening.
I run a company that builds exactly this, so I will be honest that I am talking my own book. In our case the agent sits inside the meeting on Google Meet, Zoom, or Microsoft Teams, listens to the live transcript, and either feeds a private prompt to one person or speaks out loud when it is invited to. The technically interesting part was never the summary we produce afterward. It was the half second of judgment in the middle.
The pattern shows up far outside meetings. Most AI products today are forensic. They look backward. They analyze the call, the quarter, the campaign, the support ticket, and they hand you a report. The report is smart. It is also too late to change the thing it describes.
Real time is hard, which is why the field has avoided it.
A historian can take its time. A participant cannot. It has to listen, understand who is speaking and what they actually mean, decide whether it has something worth saying, and either say it or stay quiet, all inside the few seconds a human would have. Get the timing wrong and you have built something annoying. Get it right and you have built something that changes the outcome instead of describing it.
The companies that figure this out will not win by having a better model than everyone else. The models are converging. They will win by solving the unglamorous problems around the model: latency, turn taking, knowing when to speak and when to stay silent, pulling the right context from a customer's own systems at the exact second it is needed. Those are product and engineering problems, not research problems, which is precisely why a small team can still win them.
My one prediction. Within two years, the phrase "AI notetaker" will sound as dated as "online video rental." Not because notes stop mattering, but because helping after the fact will feel like the least interesting thing a capable system could do.
The interesting question is no longer what did I miss. It is what should I do, right now, while it still counts.
The future of applied AI is not better hindsight. It is presence at decision making time. If you want to see how, check it out our AI Meeting Agents @ CoAgentor.com
Connect With Josh
Comments