Polanyi's "we can know more than we can tell" predicts which agents win

In 1958, the Hungarian-British polymath Michael Polanyi published Personal Knowledge: Towards a Post-Critical Philosophy with University of Chicago Press, introducing the concept of "tacit knowing." In 1966, in The Tacit Dimension, drawn from his 1962 Terry Lectures at Yale, Polanyi distilled the thesis: "we can know more than we can tell." Some knowledge is propositional and can be transmitted in language. Other knowledge is tacit, residing in skilled practice, intuitive judgment, embodied competence, and cannot be fully verbalized even by the expert who possesses it. Gilbert Ryle had made a related move in 1945 with the "knowing how" versus "knowing that" distinction. Polanyi pushed further: tacit knowing is the substrate that propositional knowledge rests on.

The Polanyi paradox, as economist David Autor named it in 2014, predicts which tasks AI will automate and which it will not. Tasks where the relevant knowledge is fully propositional, like rule-bound calculation, well-documented translation, and code generation against specifications, are vulnerable to automation. Tasks where the relevant knowledge is tacit, like customer-service judgment, complex sales, legal interpretation, medical diagnosis at the margin, resist automation because the expert cannot fully articulate what they know.

The 2026 agent-deployment data fits Polanyi exactly. Coding agents reach production because coding has been extensively codified into explicit specifications, test suites, and review protocols. Customer service agents fail at 74 percent rollback rate because customer-service knowledge is overwhelmingly tacit and the codification step destroys what made the practice work.

The implication is that the productive 2026-2030 bets are in domains where tacit knowledge can be partially codified through human-AI workflows, not in domains where automation requires full codification of tacit practice.