The piece I'd push on is candor. Once people assume every meeting is recorded and queryable by their boss's agent, the messy, half-formed thinking that actually produces good decisions tends to migrate to the hallway and the DM. You capture more context, but a thinner, more performed version of it.
Not record itself only, transform recorded contents and distribute as well as align to pointed team is new way of “team aligning”, its expected to see SOP and more standard push within organization happens along with the development of enterprise-level agent, and that’s how agent maximize organization efficiency.
the context layer is already live inside Chinese companies. Doubao Enterprise at Bytedance, running 120T tokens/day across Tencent and Alibaba, is basically this architecture deployed at scale. the 'whether companies are paying attention or not' framing assumes theres a window to catch up. in China that window closed maybe 18 months ago.
Verbal vs written culture is the cut that ages well, and it reorders more than people expect. I build in this space, so a couple of things I'd add.
The consent and attestation layer is what makes the valuable rooms recordable in the first place. People stuff, leadership, anything near HR. Get it in early and those rooms open up. Had a deal last quarter where the recording objection nearly killed it, and a manager attestation gate is what got us over the line. The BIPA voiceprint cases in federal court now, Otter and Fireflies, are probably going to pull all of this forward faster than six months anyway.
Recall is the obvious first prize. The one that compounds most is measurement. You get to see how someone actually leads across months instead of how they tell it in a self review. Stranger and bigger surface than search, I think. Then you watch it stack up across a whole org. That's what we're building at Parafoil.
Good article, as a follow-up, I'd like to see you talk about the downsides and risks of doing this, and how companies are addressing these risks
1. Cybersecurity: This creates a massive attack surface both from outside threats getting access to the data, and to insider threats manipulating the AI's context of the company, or just wanting to find out company secrets!
2. Missing Context: The AI won't have all the context, and so won't know, e.g., that when a meeting ended, and the recording was off, the people talked to each other correcting something that was misspoken.
3. Second Order Impacts: From years of working with systems that tried to force people into behaviors they don't like, I can tell you that people just find ways around the system - e.g., people will just use personal phones to call each other and habituate to that. And then, here, the second-order impact of the AI losing context becomes a downward spiral: it accumulates errors, and people don't want to use it even more.
The Bridgewater example holds the question everyone skips. The advantage was publicly known for decades, Dalio wrote books about it, and almost nobody copied it. Cost was never the barrier, stenographers have existed forever. What AI changed is the value side, the record finally became queryable. The social cost didn't go anywhere, it's just getting paid later now.
And later has a venue. A complete, searchable corpus of every internal deliberation is the best assistant a company ever had, and it's also the best witness, because the same completeness that makes it useful makes it discoverable. Slack in discovery was bad. Two years of synthesised meetings, queried by the plaintiff's AI rather than the defendant's, is a different order of thing. I'd say better than even odds a landmark case does exactly that within three years, and the retrofitted controls arrive the week after.
The context layer is everything described here. It just doesn't work only for you.
This resonates strongly. My friends over at Earkmark are seeing it compound on the daily.
A little quibble on terminology, if you’ll allow it: When you say ‘verbal,’ I think you really mean ‘oral.’ Yes? If verbal means word-based, then both written communication and oral communication are verbal forms of communication.
Fascinating vision but there are missing pieces today that prevent us from capturing the true value of recording every conversation.
The value of written data comes from the tools that employees can use to extract value: emails, chat documents, CRM and other systems of record. We don't have the same tools for conversations, IMHO. The value of recording conversations today comes from its ability to capture non-structured data into traditional documents and systems of records for use as text.
The missing piece for conversation recording is the creation of digital twins of the enterprise, which can create a fine-tuned model reference for the distilled conversations. That twin creation becomes the onboarding process that you were referring to, doesn't it?
The piece I'd push on is candor. Once people assume every meeting is recorded and queryable by their boss's agent, the messy, half-formed thinking that actually produces good decisions tends to migrate to the hallway and the DM. You capture more context, but a thinner, more performed version of it.
Not record itself only, transform recorded contents and distribute as well as align to pointed team is new way of “team aligning”, its expected to see SOP and more standard push within organization happens along with the development of enterprise-level agent, and that’s how agent maximize organization efficiency.
Gataca
Spot on. Voice conversations are the gold. Most companies still missing most of it.
the context layer is already live inside Chinese companies. Doubao Enterprise at Bytedance, running 120T tokens/day across Tencent and Alibaba, is basically this architecture deployed at scale. the 'whether companies are paying attention or not' framing assumes theres a window to catch up. in China that window closed maybe 18 months ago.
Verbal vs written culture is the cut that ages well, and it reorders more than people expect. I build in this space, so a couple of things I'd add.
The consent and attestation layer is what makes the valuable rooms recordable in the first place. People stuff, leadership, anything near HR. Get it in early and those rooms open up. Had a deal last quarter where the recording objection nearly killed it, and a manager attestation gate is what got us over the line. The BIPA voiceprint cases in federal court now, Otter and Fireflies, are probably going to pull all of this forward faster than six months anyway.
Recall is the obvious first prize. The one that compounds most is measurement. You get to see how someone actually leads across months instead of how they tell it in a self review. Stranger and bigger surface than search, I think. Then you watch it stack up across a whole org. That's what we're building at Parafoil.
I would say all you wrote here is very aligned with our vision and what we do at https://collabute.ai
Good article, as a follow-up, I'd like to see you talk about the downsides and risks of doing this, and how companies are addressing these risks
1. Cybersecurity: This creates a massive attack surface both from outside threats getting access to the data, and to insider threats manipulating the AI's context of the company, or just wanting to find out company secrets!
2. Missing Context: The AI won't have all the context, and so won't know, e.g., that when a meeting ended, and the recording was off, the people talked to each other correcting something that was misspoken.
3. Second Order Impacts: From years of working with systems that tried to force people into behaviors they don't like, I can tell you that people just find ways around the system - e.g., people will just use personal phones to call each other and habituate to that. And then, here, the second-order impact of the AI losing context becomes a downward spiral: it accumulates errors, and people don't want to use it even more.
While I agree, I think AI can play a more active role than only listening in and transcribing/summarising.
https://tchbytes.substack.com/p/everything-is-recorded-the-intelligence
The Bridgewater example holds the question everyone skips. The advantage was publicly known for decades, Dalio wrote books about it, and almost nobody copied it. Cost was never the barrier, stenographers have existed forever. What AI changed is the value side, the record finally became queryable. The social cost didn't go anywhere, it's just getting paid later now.
And later has a venue. A complete, searchable corpus of every internal deliberation is the best assistant a company ever had, and it's also the best witness, because the same completeness that makes it useful makes it discoverable. Slack in discovery was bad. Two years of synthesised meetings, queried by the plaintiff's AI rather than the defendant's, is a different order of thing. I'd say better than even odds a landmark case does exactly that within three years, and the retrofitted controls arrive the week after.
The context layer is everything described here. It just doesn't work only for you.
Curious why you didn't include the native note takers of Google Meet, Zoom in your assessment
Great article David! Does a16z use any tools other than Granola?
This resonates strongly. My friends over at Earkmark are seeing it compound on the daily.
A little quibble on terminology, if you’ll allow it: When you say ‘verbal,’ I think you really mean ‘oral.’ Yes? If verbal means word-based, then both written communication and oral communication are verbal forms of communication.
Fascinating vision but there are missing pieces today that prevent us from capturing the true value of recording every conversation.
The value of written data comes from the tools that employees can use to extract value: emails, chat documents, CRM and other systems of record. We don't have the same tools for conversations, IMHO. The value of recording conversations today comes from its ability to capture non-structured data into traditional documents and systems of records for use as text.
The missing piece for conversation recording is the creation of digital twins of the enterprise, which can create a fine-tuned model reference for the distilled conversations. That twin creation becomes the onboarding process that you were referring to, doesn't it?