Everything Recorded: Capturing Enterprise Tacit Knowledge

TL;DR: Everything-recorded enterprise tacit knowledge capture by default is the quiet shift behind AI at work: most meetings and calls are now recorded, turning spoken context into searchable data. That captured context becomes a living company memory AI agents can reason over. The companies that connect those recordings into one queryable layer get a force multiplier; the ones that don’t keep losing their best institutional knowledge the moment a call ends.

Most companies don’t lose knowledge because nobody created it. They lose it because the important part was said out loud, in a meeting, and then it evaporated. The real argument in a product review, the offhand comment that changed the roadmap, the way a senior rep handles a tricky renewal: none of that reliably makes it into the wiki.

That is starting to change, and not because anyone voted on it. As a16z’s David Haber puts it in Everything is Recorded Now, recording “wasn’t debated – it just happened,” and you should now assume most work discussions are captured by default. This post looks at why that default is flipping, what it means for capturing tacit institutional knowledge, and how recordings turn into context an AI agent can actually use.

What does “everything is recorded now” actually mean?

Recording by default means the assumption inverts: instead of “don’t record unless someone opts in,” the norm becomes “assume you’re being recorded unless a meeting is explicitly marked private.” Haber predicts this will be far less contested within six months and frames the underlying principle as an old one applied to speech: never say anything you wouldn’t want on the record, the same way professionals already treat email and Slack (a16z, 2026).

The point is not surveillance. The point is tacit institutional knowledge: the unwritten rules for how a company operates. Meetings are where culture lives, where expectations get set, and where edge cases get resolved. Capturing them turns that latent context into something durable instead of something that walks out the door at 5 p.m. or when an employee quits.

Why does capturing spoken knowledge matter so much?

Because the highest-value context was never written down in the first place. Structured systems hold the outputs; conversation holds the reasoning.

The cost of leaving that knowledge uncaptured is measurable. APQC found the average knowledge worker spends only about 30 of every 40 hours on productive work, with roughly 2.8 hours each week just looking for or requesting information, plus more time recreating information that already exists (APQC, 2021). McKinsey Global Institute earlier estimated knowledge workers spend about 20% of their time, roughly one day a week, searching for and gathering internal information (McKinsey Global Institute, via Nakash & Bouhnik, 2024). A lot of that hunting is for answers a colleague already gave, out loud, in a meeting nobody captured.

Large language models change the economics here. They are unusually good at taking messy spoken input and making it structured, searchable, and queryable (a16z, 2026). That capability is what makes a new category of voice-first enterprise software possible, organized around conversation instead of text.

How do you onboard an AI agent on company context?

The same way you onboard a person. You don’t hand a new hire the CRM and the wiki and expect them to understand the company. You put them in meetings and let them absorb how things really work.

Haber’s argument is that AI learns by the same osmosis, except an agent can attend every meeting, reason over every interaction, and never get bored. A model that has ingested two years of internal discussion is simply a better assistant than one trained only on documentation (a16z, 2026). He points to concrete cases: Bridgewater recorded everything as policy years before it looked smart; OpenAI now runs with nearly everything recorded, with agents standing in for leaders who can’t attend; and Granola has better context on a16z’s culture and investments than most of the firm’s other tools, because it has been in the room.

There is a second, less-discussed reason leaders want this. Recording is not only a bottom-up boost for individual contributors. It is also top-down oversight: an exec’s agent can sit in the meetings they miss and flag what matters, which helps with the alignment problem of knowing what is actually getting built (a16z, 2026). Both advantages compound, because every recorded meeting makes the system a little smarter.

Why verbal cultures gain the most

Companies tend to cluster into two types. Written cultures like Stripe or Anthropic already capture most of their context by construction, in long documents and memos. Verbal cultures like Shopify or OpenAI move fast in conversation, and historically their best context vanished as soon as the call ended (a16z, 2026).

When AI can attend every meeting and synthesize what happened, verbal culture finally scales. The bottleneck that penalized talking instead of writing goes away. Written-culture companies still benefit, since feeding good writing to an AI is also a fast way to get it up to speed. But the relative jump is largest for the talkers, whose knowledge was the most fragile.

This is also where the competitive wedge appears. Recording is the obvious default for small, AI-native companies and a fight against inertia for incumbents, who tend to respond with “have you ever been sued before?” Fair question. Haber’s bet is that recording happens anyway because it is too hard to stop, and the controls (special “do not record” designations for sensitive HR and legal meetings) get retrofitted on top (a16z, 2026).

From recordings to a living context layer

Recording is step one, not the finish line. A pile of transcripts is just more unsearchable content, scattered across a meeting tool, a call platform, and a dozen SaaS apps. The value shows up only when that spoken context gets connected to everything else: the docs, the tickets, the deals, the people.

That connected, continuously updated record is what a16z calls the living company context layer (a16z, 2026). A useful way to think about it: a knowledge graph that links entities (people, projects, customers, decisions, and now the meetings where those decisions were argued) so that one query can traverse all of it. This is the layer SemanticOS builds, a unified semantic layer and AI search across fragmented enterprise tools, so a recorded conversation stops being a dead transcript and becomes a node a person or an agent can reason from.

A concrete example: Vantage Health

Vantage Health, a mid-size health insurer, used to lose decisions in the gaps between tools. A pricing exception got approved on a Tuesday leadership call. The reasoning lived in three people’s memories, the approval lived in a Slack thread, and the policy doc never got updated. Six months later, a renewals analyst hit the same edge case, couldn’t find the precedent, and spent an afternoon pinging teams who half-remembered it.

After Vantage Health moved to recording leadership and customer calls by default and feeding those recordings into a connected context layer, that afternoon became one query. “Why did we approve the off-standard rate for the Northwind account?” returns the moment in the recorded call where it was decided, the people who argued for it, the Slack approval, and the related policy doc, all linked. The tacit knowledge that used to live only in a few heads is now part of the company’s searchable memory, available to the analyst and to the AI agent drafting the renewal.

Nothing about that requires a heroic documentation effort. The team kept talking the way it always had. The difference is that the conversation was captured and connected instead of lost.

Key takeaways

• Recording by default is flipping the norm from opt-in to opt-out, and it is happening without anyone formally deciding it (a16z, 2026).
• The prize is tacit institutional knowledge, the unwritten context that lives in meetings and usually evaporates afterward.
• You onboard AI agents like new hires: a model that has sat in two years of meetings beats one that only read the docs.
• Verbal cultures gain the most, because the spoken context that used to vanish can finally scale.
• Recording is step one; the payoff comes when conversations are connected into a living context layer, like the knowledge graph and AI search SemanticOS provides.