HR and management
Headcount, roles, salaries, organisation structure, utilisation, recruitment and performance.
For a century, capacity meant people.
Headcount is no longer a head count.
Need more work done? Hire more people. That relationship held for as long as people were the only thing in a company that could do the work. It is starting to break.
What changed
Machine labour breaks the relationship.
Once AI can hold a goal over days, keep working when people leave, create or coordinate machine workers underneath a mission, and consume resources according to the work performed, workforce capacity is no longer the same thing as the number of employees on the payroll.
The platforms are already moving. Agents are now sold that pursue a measurable goal over days and weeks rather than completing a single task. And while the everyday assistant is still priced per seat, long-running agent work is moving to usage-based billing, because its cost depends on the work being done rather than on who is logged in.
The operating-layer problem
Capacity is planned in two places that never meet.
IT and finance
Software licences, cloud infrastructure, compute and platform budgets.
Agentic AI begins to collapse that distinction. A persistent AI system performs work, consumes variable resources, runs continuously, takes on extra workload instantly, coordinates other workers, changes how much effort it applies and produces measurable business output.
That looks less and less like software licensing, and more and more like operational capacity.
The mistake
Two budget lines that tell you almost nothing.
The easy move is to keep the old model: "We have 40 people in the commercial team and £200k of AI spend." Two unrelated budget lines. But if AI is performing 30 per cent of the team's execution, that picture tells you very little about what the team can actually do.
Budgeted by department
Commercial team40 people
AI£200k a year
Two lines. Neither says what the team can do.
Allocated to the mission
Recover £2m of at-risk pipeline before quarter end
Human capacity
Accountable owner1 person
Expert judgement3 hours a week
Relationship work20 hours a week
Machine capacity
Machine execution£4,000 a month
Specialist data and tools£500 a month
Escalated to a personunder 2% of cases
Measured byCost per successful outcome
Illustrative. The same mission as essay 012, resourced. On the left, how the budget sees the team. On the right, what the work actually needs.
Now you are planning the work, not the org chart. That is where essay 012 left us: if the work is the unit, capacity belongs to the work too, whether it comes from people or from machines.
Essay 011 made the narrow version of this argument: if AI makes production ten times faster, you do not plan for ten times more reviewers. The same logic now runs through the whole of capacity planning.
None of this is a job-replacement argument, and it should not be run as one. Some work becomes machine execution. Some human work becomes more valuable, as essay 010 set out. The operating question is the right mix of each.
When machines can do the work, headcount stops being a useful measure of how much company you have.
Where it lands hardest
This may matter most to a smaller company.
A 20-person company could soon have the execution capacity once associated with 50 or 100 people. Not because it has replaced 80 jobs, but because machine intelligence can absorb research, monitoring, administration, first-pass analysis, follow-up, reconciliation, preparation, coordination and reporting.
So the owner's question changes. Less "who do I hire next?", and increasingly "where is the next unit of capacity best bought: human or machine?"
The answer varies by task. You do not buy machine capacity for empathy. You do not hire another analyst simply to reconcile spreadsheets if a machine can do it continuously. The company increasingly assembles the cheapest appropriate combination of machine execution and human judgement around the outcome. That is a real change in management.
The governance consequence
In an agentic company, a budget is not just finance. It is a permission.
Once a persistent mission can call expensive models, invoke tools, create sub-work, retry tasks and run for days, its budget becomes a runtime boundary.
This is different money from the kind essay 009 asked about. 009 asked whose authority an agent spends when it moves the company's money. This is the money the system consumes while it works, and it needs limits of its own. Essay 012 listed budget as one property of a mission. In practice it opens into a set of decisions.
What every mission's budget has to answer
- How much may this system spend pursuing the outcome?
- Who can increase that?
- At what point does it pause?
- What is the expected value of continuing?
- What retry limit applies?
- Which model quality is economically justified?
- When should it escalate rather than spend more?
- What did each completed outcome cost?
Where Praxis stands
Give every mission two resource envelopes, and plan them together rather than in two budgets that never meet.
The judgement it needs
An accountable owner, the judgement the work requires, the relationship effort it takes, the capacity to handle exceptions, and the domain expertise it draws on.
The execution it buys
A model and runtime budget, tool and data costs, concurrency, retry and delegation limits, a maximum cost per outcome, and the threshold at which spending escalates.
Then measure the combined system against one number: cost per successful business outcome. Not the number of agents. Not token consumption. Not FTE reduction. Outcome.
The company of the future will still count people. It just won't mistake the count for its capacity.
Point of View
← All essays
Previous
The work outlives the workers. →
Begin
Where is your next unit of capacity best bought?
One conversation, no pitch deck. Bring a team and its budget, and we will work out what the work actually needs, in people and in machine capacity.