Precision and range in knowledge are constitutively incompatible. Schooler’s 1990 verbal overshadowing showed that describing a face degrades recognition. The same displacement is appearing at organizational scale: formalizing knowledge sharpens the explicit representation at the cost of the tacit original.
Interloom raised $16.5 million in March to solve a problem that every enterprise recognizes: seventy percent of operational decisions are never formally documented. Their approach ingests millions of records to build "context graphs" that map how problems actually get resolved. At Commerzbank, they reduced the gap between documented and actual knowledge from roughly fifty percent to five. The bet is straightforward. Capture what workers know by converting it from implicit to explicit form.
In 1990, cognitive psychologist Jonathan Schooler asked participants to describe a face they had just seen, then pick it from a lineup. Participants who wrote descriptions performed significantly worse than those who did nothing. The descriptions were accurate. The act of describing forced a processing shift from holistic to analytic mode, and the analytic representation displaced the richer original. Schooler called the phenomenon verbal overshadowing.
The effect replicated across domains. Wine tasters who described a wine they had just sampled were worse at identifying it afterward. Golfers who described their putting technique putted worse. People who verbalized their approach to insight problems solved fewer of them. Which expertise level suffered most varied by domain — untrained wine drinkers were most vulnerable, while higher-skill golfers paid the steeper price — but the displacement itself was consistent. Wherever holistic processing carried the knowledge, describing it degraded it.
Thirty-six years later, the same structure is appearing at organizational scale.
The Pattern
The OECD Digital Education Outlook 2026 found that students using generative AI produced higher-quality outputs and were forty-eight percent more successful at completing tasks. When AI access was removed during exams, the advantage "disappeared and sometimes reversed." The report's language is precise: offloading cognitive tasks to general-purpose tools creates "metacognitive laziness and disengagement that may deter skill acquisition in the long run." The explicit output improved. The tacit learning substrate degraded.
HFS Research named the enterprise version this year: process debt. Decades of undocumented decision logic and exception handling remained hidden inside human execution. Teams absorbed ambiguity, interpreted context, and managed exceptions that formal workflows never captured. Agentic AI does not absorb this debt. It surfaces and scales it. Gartner predicts over forty percent of agentic AI projects will be cancelled by the end of 2027. The pattern is the same: organizations are automating processes whose value lived in the tacit layer they never documented.
Boeing offers the starkest organizational example. After two crashes that killed 346 people, the FAA found Boeing failed thirty-three of eighty-nine audit tests. The company had a formal safety culture program called "Seek, Speak and Listen." Employees were uncomfortable speaking up. The formalization of safety culture had displaced the actual safety culture.
The Structure
In physics, conjugate variables are pairs of measurements where precision in one necessarily reduces precision in the other. Position and momentum. Energy and time. You can measure either with arbitrary accuracy, but not both simultaneously. The limit is structural, inherent to the mathematics of wave functions.
Precision and range in knowledge behave the same way. Formalizing knowledge sharpens the explicit representation at the cost of the holistic original. Describing the face improves the description and degrades the recognition. Documenting the process improves the documentation and degrades the judgment. Implementing the safety program improves the program and degrades the safety.
This loss is structural. Converting tacit knowledge to explicit form forces a processing shift that displaces what it claims to capture.
The Boundary
The pattern holds when holistic and analytic processing compete for the same cognitive substrate. When they operate on different substrates, the displacement does not occur.
This is why the OECD found that pedagogically designed AI tools showed sustained learning gains while general-purpose tools degraded performance. The designed tools scaffolded human effort rather than replacing it. They structured when AI assistance arrived, keeping the learner's own processing intact during the critical formation period.
It is also why large language models can extract pattern-level tacit knowledge from operational records without triggering the displacement. The model processes in a different substrate than the humans whose knowledge it maps. The humans are not forced into analytic mode because they are not doing the describing. Interloom's approach may work precisely when it reads what workers did rather than asking workers to explain what they know.
What Precision Costs
Every knowledge management initiative in enterprise history has assumed that tacit knowledge can be made explicit without structural loss. The Conjugate Pair says the loss is inherent. You cannot sharpen the description without narrowing the thing described.
Interloom's Commerzbank metric measures the explicit residue. Reducing the documentation gap from fifty percent to five represents real value. Whether it also represents displacement of the judgment that filled that gap is the question the metric cannot answer.
The forward-looking implication is specific. Organizations deploying AI to capture institutional knowledge face a design choice that most have not recognized. Systems that observe what workers do, learning from the artifacts of practice, preserve the tacit substrate. Systems that ask workers to explain what they know, or that replace the cognitive work entirely, trigger the conjugate displacement. The difference between these two architectures is the difference between a tool that augments knowledge and one that consumes it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
