David Haber at a16z published a piece this week that’s been circulating among enterprise AI decision-makers. The headline claim isn’t new. Bridgewater has been recording everything since long before large language models existed. But he puts a number on something most companies still aren’t tracking: Granola grew 400% in the past twelve months. Not because meeting notes became fashionable. Because the market started to understand that an unrecorded conversation is context that’s gone forever.

The argument I want to make is more specific than “record more meetings.” Every unrecorded meeting is a training sample that will never enter any system. There is no way to recover it.

The onboarding mistake that costs enterprise AI the most

The way most mid-size companies deploy AI is the same way nobody should onboard a new employee: hand them the documentation and wait for results.

Haber puts it well. We don’t tell a new employee to sit down and read the CRM and company wiki to get up to speed. We invite them to meetings and let them learn through osmosis. That’s where culture lives, where expectations are set, where edge-case handling actually happens. A good new employee in their first three months doesn’t learn from what’s documented. They learn from what they hear, watch, and do.

The AI your company runs today, if it only has access to documents, is in that position. It has a map of what someone sat down to write. It doesn’t have the context for how your company actually works.

And the problem isn’t only that the AI is less useful. It’s that the asset that could be accumulating, the spoken history of how your company operates, makes decisions, and solves problems, isn’t accumulating anywhere. Every week that passes without those conversations being recorded is a week of corpus that nobody will ever recover.

Why verbal-culture companies now have the structural advantage

Haber makes a distinction in his piece that I think matters and that few people are pointing to.

Written-culture companies, the ones with strong traditions of communication in documents, internal memos, and recorded decisions (like Stripe or Anthropic), already had their context in a format AI can read. They didn’t have to change much. The corpus existed; they just had to point a model at it. Their AI advantage arrived almost for free.

Verbal-culture companies, the ones that work in meetings and calls and conversations that never made it into any document, historically lost that context. When a senior employee left, they took years of knowledge that wasn’t written anywhere.

Systematic recording changes that equation. For the first time, a verbal-culture company can build a corpus that nobody else has. Not generic documentation that any competitor could replicate. The spoken history of how that specific company has made decisions, managed clients, and solved problems over the past several years. That corpus is proprietary. It can’t be bought. It can’t be copied.

Bridgewater understood this before LLMs existed. Recording everything as institutional policy looked eccentric for years and turned out to be prescient. OpenAI now uses AI agents in senior leadership meetings when the executive can’t attend: the agent sits in, reasons over what was discussed, and represents the absent executive’s position. These are not cultural coincidences. They are strategic bets on which asset will matter.

What a company that records systematically is actually building

The difference between a company that starts recording today and one that has been doing it for three years isn’t technological. It’s the depth of the corpus.

A single recording is worth almost nothing. A hundred hours of organized, transcribed sales calls from 2023 through 2026 is a different kind of asset. An AI system with access to that corpus can answer questions that previously only the fifteen-year sales director could answer: which objections appear at which stage of the process, how the team responds to specific price points, what type of client historically churned before six months.

That doesn’t live in the CRM. It isn’t in Notion. It only exists in the conversations, and only if someone recorded and organized them.

Haber calls this the emergence of a new enterprise software category built around voice rather than text. The current system of record is structured data: CRM entries, tickets, documents. The highest-value context lives in conversation. LLMs are exactly the technology that can take that unstructured verbal data and make it searchable and queryable. The verbal corpus is the asset. The model is the retrieval system.

The company that has that corpus, built over three years before a competitor starts thinking about it, holds a distance that no AI budget increase can close.

On the legal side, the AI Chats Are Discoverable Evidence post covers what enterprise teams should know about recorded conversations before they scale. Worth reading in parallel with this.

What IQ Source does with this

When we work with a company on AI Maestro, the first deliverable is the Process Reality Map. That’s the written layer: what each operation does, what tools it uses, where the bottlenecks are, what decisions get made at each stage and by whom. It’s the corpus that AI can read to understand how your company works from the documentation side.

But there’s a second layer that map doesn’t capture. The verbal layer. The conversations where your teams actually make decisions, negotiate with clients, debate priorities. That layer can’t be built by consultants. Nobody can build it for you.

What we do structure during AI Maestro is what to record, how to label it, and how to organize it so that in twelve months you have a corpus your AI can actually use. This isn’t a technology project. It’s a knowledge capture decision, and it gets made now or not at all. The conversations from this week won’t be available next year if they aren’t recorded today.

The asset a team whose reasoning is searchable builds starts with both layers: the written one, which we can map together, and the verbal one, which only you can start recording.

One question to close with. If you deployed an AI system tomorrow with access to every recorded conversation from your company over the past three years, how many would you have? If the answer is “very few” or “none,” that’s the asset you’re leaving unbuilt today.

For the written layer, we can help you now. For the verbal layer, the clock has been running.

Map your company’s operational context with AI Maestro