Why Machines Need to Know Why Before They Can Decide How

This post is part of an ongoing series examining the forces reshaping how organizations establish truth, make decisions, and maintain control in an AI-driven world — from the economics of intelligence to the architecture of what comes next.

In 1966, philosopher Michael Polanyi coined a phrase that should haunt every enterprise AI strategy: "We can know more than we can tell."

A radiologist spots a tumor that less experienced doctors miss but can't explain what tipped them off. A master chef knows a sauce needs adjustment before tasting it. A veteran trader senses a market turning point before the data confirms it. The expert is operating on tacit knowledge — institutional understanding built from thousands of informal observations that never made it into a manual.

Tacit knowledge is most of what makes expertise valuable. And it's invisible to current AI.

The 90/10 problem

Organizations document maybe 10% of what they actually know — policies, procedures, org charts, contracts, meeting notes. Most estimates suggest the remaining 90% is tacit: why decisions were made (not just what), context that made a decision sensible at the time, informal rules ("never schedule launches before holidays"), historical failures that shaped current practice, unspoken assumptions baked into every process.

When an experienced supply chain manager says "don't ship through that port," they aren't running calculations. They're drawing on hundreds of informal observations: political instability indicators, seasonal weather, relationships with local officials, anomalies that never made it into reports.

Current AI — even with context graphs — only reads the documented 10%. This is why it produces recommendations that are technically correct but contextually wrong. It doesn't know the informal reasons why "technically correct" won't work in your operation.

You can't ask experts to document tacit knowledge

That's the definitional trap — if they could articulate it, it wouldn't be tacit. The hard problem isn't writing more documentation. It's capturing the uncapturable.

Dynamic autonomous graphs solve this by observing behavior and inferring the hidden logic from patterns:

• Decision traces — Who was consulted before a major call? What information was reviewed? What was ignored? Fast deliberation signals routine; slow deliberation signals uncertainty. What alternatives were rejected, and what happened after the decision?
• Communication patterns — When experts are worried, who do they talk to first? What language signals high confidence vs. hedging? Which documents get read carefully vs. skimmed?
• Behavioral anomalies — When do experienced people override standard procedures? What patterns precede a decision to escalate vs. handle locally?
• Outcome correlations — Which informal factors predict success or failure? Do certain combinations of people produce better results? Are there temporal patterns affecting decision quality?

The output is an organizational memory that never forgets why decisions were made, what context mattered, and which informal rules actually predict success. Not replacing experts — extending them.

This is Asimov's Hari Seldon made real. Not predicting individual decisions through mysticism, but encoding the tacit patterns that drive decisions into a model that can see further than any single consciousness.

The existential question: who controls the knowledge engine encoding your organization's tacit knowledge? If you build it, machines work for you. If someone else builds it, you operate blind while they see everything.

Next up: how knowledge engines become the cognitive substrate that ends the SAP era and rewires what enterprise software actually does.

Works Referenced

Polanyi, Michael. The Tacit Dimension. University of Chicago Press, 1966.

Asimov, Isaac. Foundation. Gnome Press, 1951.

← Back to insights