What agent builders assume you already know

I have been building AI applications for enterprises since university. Not advising on them, building them. Sitting in the room with the operations team, shipping under real deadlines, and watching what happens when an agent meets an actual business process.

After two years of agentic AI delivery across several verticals, one pattern showed up in every single project. And it was never about the technology.

The promise vs. the reality

Agent platforms are impressive. They have gotten very good at the runtime layer, orchestration, tool use, memory, multi-step reasoning. The demos are convincing. The documentation is solid. The frameworks keep getting better.

And yet, many enterprise AI pilots still fail to produce measurable business impact. MIT NANDA reported in 2025 that 95% of enterprise GenAI pilots produced no measurable P&L impact. McKinsey found that while 88% of companies use AI, only 23% have scaled it inside even one business function. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.

Better models. Better frameworks. Bigger budgets. Same result. The problem is not the runtime. The problem is what the runtime is running on.

The hidden assumption

Every agent builder makes the same assumption: the customer already has the knowledge the agent needs. It sounds reasonable. The platform provides the execution layer. The customer provides the domain knowledge.

Here is where it breaks. In most companies, that domain knowledge does not exist in a form the agent can consume. Roughly 80% of how work actually gets done is undocumented. (Tallyfy) The real operating logic, the judgment calls, the exception handling, the "call this person before you escalate" rules, lives in a few experienced employees' heads.

The agent builder hands you the engine. But the fuel, the operating knowledge, is missing.

Three gaps configuration cannot close

Decision logic. Every process has branch points where a human makes a judgment call. A senior operator does not follow a decision tree, they weigh factors that are not in any tree. Past experience with this client. Time of year. Backlog pressure. You cannot configure this into an agent because it has never been written as a rule.

Exception handling. The official process covers the happy path, maybe 20% of the work. The other 80% is exceptions. When an agent hits an exception it was not configured for, it guesses (dangerous), stops and asks a human (defeats the purpose), or silently routes incorrectly (expensive to catch later).

Source trust. An experienced operator knows which data sources to trust and which to verify. They know that a particular field is often outdated, that a certain report is reliable for aggregates but not for individual records. An agent treats every source equally unless told otherwise, and nobody has told it otherwise.

Why more configuration is not the answer

The instinct is to solve this with more configuration. More rules. More decision trees. More documentation fed into the agent. This does not work for two reasons.

First, it does not scale. The number of judgment calls, exceptions, and contextual factors in even a single operational process is enormous. Trying to anticipate and configure all of them is a project that never ends, and the result is brittle.

Second, it starts from the wrong substrate. Configuration starts from what someone can write down today. But the hardest, most valuable operating knowledge is precisely the knowledge that was never written down, because it is tacit, contextual, and accumulated over years.

The layer that is missing

There is a step that comes before any agent platform, any framework, any orchestration layer. A step that most companies skip: expertise mining.

Expertise mining extracts the real decision logic, exceptions, trusted sources, and operating judgment from the people who hold the work. Not through documentation. Through structured extraction, asking the right questions, in the right order, validating across experts, and packaging the result as structured, machine-readable operating knowledge.

The companies that will scale their AI operations are the ones that capture the expertise first. They mine the knowledge that makes their operation work. They structure it. They validate it. And then they deploy agents on top of it, agents that know what the experts know.

The agent is the easy part. The expertise underneath it is the hard 90%. Whatever platform you use, whatever framework, whatever model, you still have to do expertise mining first.