Expertise mining, a manifesto

Enterprise AI has reached a strange moment. Adoption has never been higher. Models have never been more capable. Budgets have never been larger. And yet most projects do not work.

MIT's 2025 study of 300 enterprise GenAI deployments found that 95% produced no measurable impact on profit. Companies have spent between $30 and $40 billion on enterprise generative AI in the last two years. Most of that money produced nothing. McKinsey's 2025 survey shows the same gap from a different angle: 88% of companies use AI, but only 23% have managed to scale it inside even one business function. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

The pattern is too consistent to be a coincidence. Better models, bigger teams, and larger budgets keep producing the same result. Something structural is missing.

Two legs

A working AI agent stands on two legs: knowledge and execution. Knowledge covers what the agent needs to know, the work, the rules, the exceptions, the tools, and the reason behind every decision. Execution covers how the agent actually runs, the right tools, in the right order, with the right context, with humans in the right places.

The 80% nobody can see

Multiple independent studies find the same thing. Around 80% of how work actually gets done in a company is not written down anywhere. Harvard research finds that 42% of essential expertise lives only in employees' heads.

The official process exists. But the real work depends on something else. It depends on the experienced employee who knows when to ignore the standard path. The one who knows which customer signal matters. The one who knows which exception is common, which approval is required, which system field is stale, which action is risky.

That knowledge is what keeps the company running. It is also the knowledge that no automation platform has ever been able to reach. If you give an AI agent only the documented 20%, you get an agent that handles the easy cases and breaks on everything else. That is what most enterprise pilots look like today.

AI did not create this problem, it exposed it

Companies have been paying the cost of undocumented expertise for decades. A new employee who needs six months to become useful. An expert who becomes the answer desk for the whole company. A process that breaks the moment the one person who understood it leaves.

Before agents, the gap was hidden. A new employee asked the senior person. A manager corrected the path. A team used a workaround. The company paid the tax in meetings, delays, mistakes, and repeated questions. Agents remove the hiding place. An agent cannot silently ask the senior person. It cannot course-correct with a workaround. It either has the expertise it needs, or it fails, visibly, on every real case.

Why current approaches don't fix this

RAG and enterprise search find knowledge that has already been written down. They are good at that. The problem is that the most valuable expertise was never written down in the first place.

Process mining reads system logs. Logs show what happened. They do not show why. A log can tell you that a manager approved a refund. It cannot tell you why she approved this one and rejected another one yesterday. The decision lives in her, not in the log.

Agent builders give you a runtime. They assume you already know what the agent should do. In real enterprise work, this is almost never true. The company knows the goal. It does not have the decision logic, the exceptions, or the handoff rules written anywhere.

The "why" question

What an AI agent really needs to learn is judgment. Judgment is the answer to one simple question: why did you do it that way?

A senior nurse skips a step in the protocol because something feels wrong about the patient. A claims adjuster approves an unusual refund because the customer's history makes it credible. A retail manager moves inventory between stores because she knows a local holiday starts next week. None of this is in the system. None of it is in the logs. It is in the person.

If you cannot answer "why did you do that?", you cannot build an agent that does the same thing. You only get an agent that copies actions without understanding them.

What expertise mining is

Expertise mining is the missing leg. It is the discipline of extracting the undocumented, role-based, operational expertise that lives inside experienced employees, and turning it into something a machine can read and use.

The four pillars of enterprise agentic AI

1. Expertise Mining, extract the knowledge and judgment that lives in people.
2. Company Expertise, connect that knowledge to the company's real data, systems, and tools so agents can act on it.
3. Governed Agent Operations, run agent teams safely, with shared context, permissions, audit, and human oversight.
4. Continuous Learning System, let the system get better as the work changes, with humans approving every material change.

Pull any layer out and the chain breaks. Most teams skip layer one and start at layer three. That is why their pilots do not survive contact with real work. They built the runtime before they extracted the knowledge it was supposed to run on.

Why now

Every wave of enterprise software has separated companies into two groups. The ones that finished the previous wave's organizational work rode the next wave. The ones that had not, fell behind. A company without records could not move to PCs. A company without digital infrastructure could not move to the internet. A company without systemized processes could not move to SaaS and RPA.

A company without machine-readable expertise cannot move to agents. Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The companies that finish the expertise layer first will own the next decade.

What to do

Start with expertise mining. Not later. Not after the pilot fails. Not as a parallel project. First. Once a company can see and structure its own operational expertise, agents stop being demos and start being workforce.

The companies that mine their expertise first will operate differently from everyone else. The ones that keep buying agent demos will spend the next three years in pilot purgatory.