A company contains many small operating systems. How purchasing works. How a product gets approved. How a customer complaint turns into a decision. How someone knows when to escalate, and who they escalate to. How knowledge moves from one person to the next. How exceptions get handled. How the founder decides things that never get written down.
None of that is on the org chart. The chart shows boxes and reporting lines, which is a picture of who reports to whom, not a picture of how work moves.
I learned this on a factory floor
In a factory, when output is too low, you do not start by adding a person. You look at the line. Where is the blocked station? Where does work pile up? Which step creates the delay, and what happens right before and right after it?
Most of the time the problem is not headcount. It is the shape of the flow. A second worker at a station that was never the constraint just produces a bigger pile in front of the station that was.
I think knowledge companies have the same problem. We just cannot see the line. Nobody walks past the desk and notices that work has been sitting there for three days.
Most of it is stored in people
A lot of what a company knows is informal. Someone knows that a particular supplier needs reminding twice. Someone else knows that one customer always needs a different process. A founder knows why a market was abandoned, and the reason was not the one in the minutes. A purchasing manager looks at a stock number and feels uneasy while the spreadsheet still looks fine.
Those are real operating rules. They are just stored in heads, or buried in email threads and meetings nobody will read again.
That works while the same people stay in the same roles. It stops working when the company grows, when someone leaves, or when you start pointing AI at the work.
The wrong first question
This is where I think companies will get AI wrong. They start with the technology. Let us build a purchasing agent. Let us build an HR agent. Let us build a marketing agent.
Before you build the agent, you have to understand the system it walks into. What is the process? Which rules matter? What data does it need, and is that data true today? Who can approve what? Where are the exceptions? What happens when it goes wrong at three in the afternoon on a Friday?
If that is unclear, AI does not remove the mess. It runs the mess faster and more often.
What I mean by a genome
I think of a company as having an operating genome. Not one giant handbook that nobody opens. A connected structure: processes, strategies, SOPs, skills, decisions, roles, data, signals, rules, exceptions, permissions.
The valuable part is not the individual document. It is knowing how the pieces connect. A process depends on a strategy. A skill gets used inside a process. A decision overrides a normal rule. A signal triggers a review. A person or an agent is allowed to act, but only inside part of that structure.
That is much closer to how a company actually behaves than a folder of PDFs.
Not everything is company truth
This matters as soon as AI starts reading and writing into the system. A company holds very different kinds of information at once. Something we know. Something someone said. Something we observed. Something we assume. Something we simulated. Something management actually decided.
Those should not collapse into one pile. If someone says in a meeting, we should probably stop selling this product, that is not company strategy. It is a statement. Maybe an idea, maybe a signal. It becomes something the company operates on only after a decision.
Without that boundary, AI turns a passing thought into a permanent rule, and then defends it with a citation.
What opens up
Once the operating systems are visible, the questions get better. Which processes exist. Which depend on one person. Which have no documented decision logic. Where two departments run on conflicting rules. Where work gets blocked again and again. Which process should be improved before it is automated. Which part of the company has enough structure to hand to an agent at all.
That is a different exercise from buying another tool.
I do not think the interesting question is how many agents a company runs. It is how well the company understands its own operating system. What already exists, what is informal, what is missing, what is broken, and what should actually be built next.
