Judgement
The call that the evidence does not quite make for you.
Every operating model in use today was written for one kind of worker.
Better intelligence in the same company changes nothing.
Roles, permissions, working hours, queues, approvals, escalation, training, management ratios. Every one of them was designed on the quiet understanding that a person would be the one doing the work. That understanding held for as long as it was true.
What actually changed
The thing doing the work stopped being a person.
Models now shipping will install and test software on their own, drive professional applications through their interfaces, run a piece of research from question to summary, and fill in the routine gaps in an instruction rather than stopping to ask. They plan, check their own output, and carry a task across several systems without a person at each step.
That is a different kind of thing from software that helps somebody work faster. It is closer to a second category of worker, and it arrived without anybody redesigning the company to receive it.
So most organisations do the obvious thing. They take the process they already have and look for a step an agent could take over. Human process, plus AI.
Very little changes, and the reason is structural rather than technical.
The buried assumption
Every primitive in the operating model assumes a person.
None of these were decisions anybody made. They are the residue of the only kind of worker there has ever been, and they sit underneath the process design rather than inside it.
Roles
One person, one job, one set of hands.
An agent holds several at once, and none of them get tired.
Permissions
Granted to a named individual you could ask about it afterwards.
Inherited down a chain, spent by something that was never asked.
Working hours
Work stops at night and picks up in the morning.
It does not stop, so the overnight backlog never forms.
Queues
A backlog is the honest cost of finite attention.
Attention is no longer the scarce thing, so the queue is now a choice.
Approvals
Somebody reviews it before the consequence lands.
The consequence can arrive first, and often does.
Escalation
Hard cases go up to somebody with more context.
It may hold more context than the person it escalates to.
Training
Capability is built slowly and then stays where you put it.
Capability moves on a release note nobody read.
Span of control
One manager can hold about seven people in mind.
Nobody has said what the number is, and most companies are past it.
Read the amber line down the page and you have a specification nobody wrote. Read the blue line and you have what is already running inside the business. The distance between the two columns is the reason the pilot worked and the programme did not.
Where this differs from essay 001
The first essay in this series argued that AI dies in the operating gap: that roughly eighty per cent of the work is operating work, and that companies skip it and get pilots instead of capability. That argument holds and we have not moved off it.
This is a different claim, and a less comfortable one. It is not that the operating work gets skipped. It is that the operating work, done properly and in full, still encodes an actor who is assumed to be human. You can map the workflow, agree the definitions, assign the ownership and build the measurement, and the result is a beautifully engineered process for a kind of worker that is no longer the only kind you have.
001 says you did not do the work. This says the work itself has an assumption inside it that has quietly expired.
Where could an agent slot into this process?
That is the question almost every organisation is currently asking, and it accepts the process as given. The better one is harder and considerably more useful.
If this were being designed today, for people and intelligent agents working together, would we design it this way at all?
Ask it of anything real and the questions come quickly. Why does this sit overnight. Why does it pass through three people. Why is somebody preparing a thing that somebody else then checks. Why does this decision need a meeting. Why does this run only in office hours. Why does every exception land on the same manager. Why are there five handoffs where one human decision and three machine actions would now do.
Most of those answers are not process design. They are fossils of the constraint that a person had to do it.
The next phase is not putting better intelligence into the company we already have. It is redesigning the company around the intelligence we now have.
The human edge
Redesign it honestly and the person becomes worth more, not less.
This is the part that gets lost when the conversation turns to headcount. Taking the machine-shaped work away from people does not diminish the role. It concentrates it on the eight things a company cannot delegate to anything else.
Accountability
Someone whose name is on it when it goes wrong.
Relationships
Trust that was built before it was needed.
Ambiguity
Deciding what the question actually is.
Ethics
The line that is legal and still wrong.
Novel exceptions
The case that has genuinely never happened before.
Negotiation
Getting to an outcome with someone who wants a different one.
Direction
Choosing which problem is worth the company's attention.
Notice what has happened to the unit of work along the way. It used to be a person and a tool. It is becoming a person directing a set of capabilities that operate largely without them, holding the exceptions and carrying the accountability. That is a change in what a job is, not a change in how quickly one gets done.
Where Praxis stands
Do not start with an operating model review. They take a quarter, they produce a diagram, and the assumption we are describing survives the exercise untouched because it is never written down anywhere to be challenged.
Start with one process that matters and one question: if we were designing this today, knowing what can now do the work, would it look like this? Then follow only what that answer breaks. Usually it is the handoffs first, the queue second, and the approval last, because the approval is where somebody has to decide what a person is actually still for.
The honest version is that this is harder than adopting a platform, and slower to show, and it is the only part a vendor cannot sell you. It is also the part that decides whether any of the rest of it was worth buying.
You can put a remarkable new kind of worker into a company built for the old one. It will be quietly wasted there.
Point of View
← All essays
Begin
Which of your processes would you design this way again?
One conversation, no pitch deck. Bring a process you know well and we will ask what it would look like if it were being designed now.