Every company has two handbooks. One is written and nobody reads it. The other is not written and everybody works by it.
An agent hands back everything a company never wrote down, in the form of errors.
The story of enterprise automation is usually told through technology: which model, which integration, which licence. Yet every team that tries to hand a process to AI hits the same wall in its first week, and that wall has nothing to do with technology. You tell the agent to run the invoice approval process. The agent applies the written procedure. Then someone from accounts comes over and says: that supplier’s invoice is never approved the way the procedure says; first you ask a particular person in purchasing, because there was a problem three years ago.
That knowledge is written nowhere. The company has run on it for years all the same.
We cannot say what we know
In The Tacit Dimension, published in 1966, Michael Polanyi described this in one sentence: we know more than we can tell. Someone who can ride a bicycle cannot explain how they keep their balance. An account manager of twenty years knows which client really means urgent when they say “urgent”, but cannot turn that into a rule. Polanyi called this tacit knowledge.
Organisations run largely on tacit knowledge. A new hire’s first six months go on learning the unwritten handbook: who to ask what, which rule bends, which signature is a formality and which one is actually read. People pick this up by instinct, by watching, by talk in the corridor. Nobody thinks of it as a cost, because people do it unprompted.
An agent does not. An agent has no corridor. Whatever written rule you give it, that is the company. This is why automation has become the first instrument that measures the gap between what an organisation thinks it knows about itself and how it actually works. Trying to hand a process to an agent is the fastest way to count how many exceptions really hold that process up.
Automation is a writing job
The conclusion is the reverse of what most executives expect. The longest and most valuable part of an AI project is the company writing itself down. Setting up the model is the shorter part. Which decision is made on which information? When is an exception an exception, and when is it really the rule? For a job to count as done, who looks at what?
Teams that sit down to write the answers discover two things. First, some rules no longer have a reason. The problem from three years ago was solved, the person left the company, but the habit remains. Second, some critical knowledge lives in one person’s head. When that person goes on leave the process slows down and everyone treats that as normal. Both discoveries are valuable with or without AI. The agent was only the excuse that made them visible.
This is what companies that come out of automation disappointed have in common. They bought the tool, connected the process, watched the agent make “absurd” mistakes and decided the technology was not ready. Most of the time it was not the technology that was unready. The agent applied the rule it was given to the letter, and the rule turned out not to match the company’s reality. The mistake does not belong to the agent. It is the first time anyone has felt the absence of knowledge that nobody, for years, saw any need to write down.
An organisation’s unwritten rules are its real culture. Values are written on the wall; culture lives in the exceptions. In the age of AI the competitive advantage will sit with the organisation that manages to write those exceptions down first. A company that can describe itself can hand its work over. One that cannot will hit the same wall again with every new tool.
In your company, what is the first rule that everyone knows and that is written nowhere?