The Knowledge Bottleneck: Why AI Value Stalls Where Your Expertise Stays Tacit

The platform industry just admitted the models were never the problem. What they left out is the part only leaders can fix.

In June, the CEOs of the two largest data platforms on the planet stood on their own keynote stages and made a confession.

Ali Ghodsi of Databricks put it bluntly: "Most enterprise AI today is just guessing with false confidence. That is not good enough for business. If you're a CFO and AI can't tell you why margins changed, or you're a sales leader, and it can't find your next upsell, that's not an AI problem, that's a context problem."

A week earlier, Snowflake's Sridhar Ramaswamy built his entire Summit keynote on the same admission. The model is not your advantage. Every competitor you have can buy the same models you can. The only durable edge is what happens when those models meet the knowledge that is uniquely yours.

Read that again. The companies selling the machines just told you the machines were never the problem.

Then Snowflake put a number on it. In benchmarks the company shared at Summit, frontier AI agents answering hard business questions on their own scored roughly 24 percent accuracy. Add a standard semantic model and accuracy climbed to about 47 percent. A heavily tuned internal model reached about 73 percent. Ground the same AI in the organization's full governed business context and it hit roughly 86 percent.

Same model. Same questions. A three-and-a-half-times improvement — and not one point of it came from better AI. Every point came from what the organization knew about itself.

The Failure Pattern Everyone Is Misreading

That number explains a pattern most leadership teams are living through right now but describing incorrectly. MIT's NANDA initiative found that 95 percent of custom enterprise GenAI pilots produce no measurable P&L impact. S&P Global reports that 42 percent of US companies abandoned most of their AI initiatives in 2025 — up from 17 percent the year before — walking away from an average of $7.2 million in sunk cost per failed effort. Gartner projects that 60 percent of AI projects unsupported by AI-ready data will be abandoned through 2026. And in a global study of more than two thousand leaders, Snowflake and Omdia found that only 7 percent of organizations say the majority of their unstructured data is ready for AI at all.

Faced with numbers like these, most executive teams reach for the same explanation: the technology isn't mature yet. So they wait. They watch the model release cycle. They tell the board the next generation will close the gap.

The next generation will not close the gap. The gap was never in the model.

The Knowledge Bottleneck: The point at which AI value stalls because an organization's most valuable knowledge — the judgment, pattern recognition, and hard-won expertise of its people — remains tacit: never captured, never encoded, and unreachable by any platform. AI value doesn't stall where your technology stops. It stalls where your expertise stays tacit.

What the Platforms Left Out

Here is the part the vendors left out of their confession. When they say "context," they mean whatever their systems can catalog — schemas, semantic layers, documents, metadata. That is real and it matters, and the benchmark proves it. But it only reaches the knowledge your organization has already written down.

Gartner estimates that 70 to 80 percent of enterprise knowledge is fully tacit — never captured in any written or retrievable form. The philosopher Michael Polanyi named this reality six decades ago: we know more than we can tell. The judgment your best underwriter applies to a borderline case. The pattern your top salesperson recognizes in a stalling deal. The thirty years of reasoning your most experienced operator uses to know which alarm matters and which one doesn't. None of it lives in a system waiting to be indexed. It lives in people. And no platform purchase — at any price — extracts it.

The constraint on AI value is no longer what the model knows. It is what your organization knows about itself, what has been extracted from your people's expertise — and how little of it has ever been made explicit.

Why This Is Leadership Work

The bottleneck is not a procurement problem. It is a leadership discipline. Converting tacit expertise into explicit, loadable knowledge is curation work — deciding which judgment matters most, drawing it out of the people who hold it, encoding it so AI can use it, and keeping it current as the business changes. This is the work of the Knowledge Architect, and it cannot be delegated to IT, because the raw material isn't in IT's systems. It's in your people's heads.

And here is why leadership is the operative word. Research published in the Journal of Knowledge Management found that tacit knowledge moves on trust — and on two different kinds of it. People share what they know when they trust the relationship. People use what others share when they trust the competence behind it. Break either one and the exchange stops. You cannot mandate that kind of trust, and you certainly cannot license it from a vendor. It is built by leaders who make knowledge-sharing safe, valued, and visibly rewarded — or it doesn't exist.

What Good Looks Like

Bridgewater Associates offers a clean example of a leadership team actually working this problem. Their AIA Labs needed AI to filter financial documents with expert-level judgment — and found that frontier models on their own performed at roughly 50 percent accuracy. A coin flip. Instead of waiting for a smarter model, they did the harder thing: they sat their domain experts down and codified the evaluation criteria those experts had been carrying in their heads — the distinctions, the exceptions, the reasoning. With that expertise encoded into the system, accuracy reached 84.7 percent, at nearly fourteen times lower cost than the frontier alternative.

The model didn't get smarter. The organization got more explicit.

That is the move, and it compounds. Every piece of expert judgment you capture makes the next AI application better, faster, and cheaper to build. The knowledge doesn't expire when the project ends or walk out the door when the expert retires. Organizations doing this work are stacking capability on capability while their competitors restart from zero with every pilot — and every departure. The ones who wait are compounding something too: the same 24 percent ceiling, repainted with each new model release.

The Leadership Imperative

So here is the challenge. This week, pick the three most consequential decisions your organization makes repeatedly. For each one, ask a single question: where does the judgment live? If the answer is a name rather than an asset — if the reasoning exists only in someone's head — you have found your bottleneck. It was never in your tech stack. No vendor is coming to fix it, because no vendor can.

Your AI is exactly as smart as what your organization knows about itself — and what it has drawn out of its people. Most of that knowledge has never been written down.

That's not a technology gap. That's your job.

Scott Wise brings 30 years of transformation consulting experience to the most important leadership challenge of our time. Author of AI4Leaders: Amplify Your Impact and certified in AI by MIT and Oxford, he helps executive teams and organizations move from AI-Curious to AI-Capable. Explore his work at ScottWise.ai.