Tacit knowledge cannot be reliably codified for AI because it is a unique product of human experience and is, by definition, impossible to translate into words. Implicit knowledge on the other hand can be codified for AI, even though it is unwritten, so long as trusted documentation exists from which it can be inferred.
Unlocking the value of undocumented knowledge is one of the great potential prizes of AI. In our experience, substantial enterprise value lies in the tacit and implicit knowledge of its experienced workers, which is lost when they depart. If that knowledge can be captured in a machine-readable way, and retained by the business for AI or human training, then the benefits of domain intelligent systems at scale are within reach.
Tacit knowledge vs implicit knowledge
Implicit knowledge is what has not yet been documented, while tacit knowledge is what cannot be documented. Tacit knowledge is often mythologised for the purposes of personal or corporate branding. A salesman who always closes a deal, or a leader with skills that inspire teams to achieve great things, is objectively valued by markets for their tacit knowledge, even if their method is not entirely clear or copiable.
Implicit knowledge, by contrast, can generally be obtained through a logical process of observation and deduction – which sounds less glamorous but is good news for anyone wanting to get value out of reliable AI inference.
How is tacit knowledge transmitted?
Traditionally, tacit knowledge has been passed on through mentorship and setting an example. This can have, shall we say, uncertain outcomes. It’s the kind of expertise acquired by living in the real world, being exposed to a company’s culture, and understanding how they really operate. A corporation’s vaunted ‘instincts’ or ‘tribal lore’ are flags for tacit knowledge.
Mentoring can have, shall we say, uncertain outcomes.
Because transmitting tacit knowledge is hard, it is considered a useful if limited moat against business competitors. Short of poaching an employee, tacit knowledge cannot easily be copied. It’s not a secret that can be usefully divulged to a rival in a purloined training manual, say. That’s because success largely depends on the capabilities of the person reading the manual, and the qualities of the organisation within which they work.
For now at least, tacit knowledge cannot be automated by AI because it is so deeply tied to the lived experiences, and possibly even the biology, of human beings.
Why is implicit knowledge more valuable for AI?
Unlike tacit knowledge, implicit knowledge can be inferred and replicated. This makes it far more valuable because it can drive growth at scale, and potentially be used to train and govern AI agents for example.
Tacit knowledge will always be valuable, of course, and unique. But it is principally defensive; it serves to protect market share, rather than expand it. In other words, it cannot scale.
Tacit knowledge is brand. Implicit knowledge is scale.
Why does so much valuable enterprise knowledge go undocumented?
Much implicit knowledge is undocumented simply because it was not considered valuable in the past. Programmers document APIs but don’t bother to record debugging processes. Doctors record successful diagnoses but discard hypotheses that turned out to be wrong. Sometimes, knowledge remains undocumented simply because it protects somebody’s interests. Medical researchers for example have shown themselves to be keener on publishing exciting new findings than they are about reporting a failure to replicate results.
The main point is that human experts tend to be aware of implicit knowledge in their domain, and are able to navigate its risks and opportunities; whereas an AI mostly doesn’t, and can’t. The only way to fix that, is to make implicit knowledge explicit.
How do I capture implicit knowledge for AI?
The job of making implicit knowledge reliably explicit, a process known as externalisation or elicitation, is not pain-free. To begin with, it needs humans. Someone’s got to verify that the documented knowledge being used to infer from, is of sufficient quality. Humans-in-the-loop then need to check those inferred outputs to ensure they are both accurate and aligned with the enterprise’s requirements. Ongoing monitoring and updates are needed to guard against issues such as semantic drift.
A 2026 paper by Hengyu Liu et al. is worth a look on this topic. It sets out how reliable AI needs to externalise implicit knowledge using structured artefacts that humans can inspect, verify, and endorse. This involves a combination of explicit knowledge, such as documentation and data, and human involvement in the definition and validation of implicit knowledge.
How does consensus-building support externalisation?
The externalisation of implicit knowledge is burdensome whichever way you look at it, although consensus-building methodologies do exist to lighten the load, sometimes as a feature of ontology-as-a-service (OaaS) platforms. These offer ways for human domain experts to agree on the design and definitions of structured knowledge as a basis for grounding and constraining vertical AI agents.
Potentially, this enables enterprises to capture and operationalise implicit knowledge, and actually create measurable value from their AI, while reducing the risks of it hallucinating or making mistakes.