The scene repeats in any meeting with an open headcount on the table. Somebody says the role runs about eighty thousand a year, and somebody else says an agent does "that same thing" for two hundred a month. The room goes quiet for two seconds, because the number is so good it sounds like the discussion is over. It is not over. The subtraction was done wrong.
The thesis in one line: you are not comparing two prices, you are comparing two assets with different curves. The person starts expensive and clumsy, compounds over time and absorbs the unforeseen without being asked. The agent starts cheap and competent, stays flat forever and breaks the moment something shows up that you did not plan for. So the question is not which one costs less, it is what percentage of your process is predictable. That percentage decides, and it is almost never zero or a hundred.
Hire a person or deploy an AI agent: why the standard comparison is framed wrong
The mistake has a very specific shape: the annual loaded cost of a role goes head to head with the monthly invoice for a tool, as if the two numbers measured the same thing. They do not. Salary is the price of a complete capability — judgment, initiative, memory of how this company works, the ability to improvise when the process leaves the rails — and the subscription is the price of a narrow one: running a bounded procedure many times without getting tired. Comparing them is comparing the price of a car to the price of a train ticket without asking where you are going.
There is a second, more expensive mistake underneath. Both numbers are incomplete, but they are incomplete on different sides. The human side forgets everything that is not salary. The agent side forgets almost everything that is not the license. And because the two omissions do not cancel out, the comparison is not merely imprecise — it is biased, always in the same direction.
What follows is an attempt to put both complete numbers side by side. Not to hand you a universal verdict — there is none — but so the decision gets made on the variable that actually governs it, which is not price.
What hiring a person really costs (the part that is not salary)
Start with the side that feels familiar. The first lump is that salary is not the cost. In Spain, according to the Quarterly Labour Cost Survey published by the national statistics institute INE for the first quarter of 2026, average labour cost per worker per month was €3,278.01, of which €2,403.80 was wage cost. In other words, roughly 27% of what the company pays for a person never shows up on that person's payslip. That figure is Spanish and it sizes the problem rather than promising a result: in the US, in Germany or anywhere else the ratio moves, and the number worth arguing with is your local one.
On top of that running cost sit three more things that belong to the role and rarely make the slide. Recruiting: the time of whoever screens, interviews and decides, plus whatever the candidate source costs. Ramp: the weeks or months where the person is paid in full and performs in part, which in a judgment role is not two weeks. And turnover: the very non-theoretical chance of repeating both of the above sooner than planned, losing along the way the knowledge that person had built.
But the important part of the human side is not a cost, it is a property: the curve goes up. Six months in, that person knows things about your company that are written down nowhere — which client is sensitive to what, which exception gets approved and which does not, who needs a heads-up before anything moves — and that accumulation is exactly why year two costs the same as year one and delivers considerably more. No agent does that on its own.
What deploying an agent really costs (the part that is not the subscription)
The agent side has the opposite problem: the visible price is real and it is low, which is why it gets mistaken for the total. Subscription and model usage are the cheap part and the only part that shows up on an invoice. The expensive part lives in three places nobody bills you for.
- Design. Before the agent does anything, somebody has to write down the procedure that currently lives in one person's head: what comes in, what gets decided, what happens to each exception. That work is done by your people, paid in hours from whoever knows the most, and it is what actually determines whether the agent works.
- Supervision. An agent in production needs somebody watching what it returns, approving what needs approving, and catching what went off the rails. This is not a launch phase that switches off: it is a permanent cost line, smaller than the work it replaces but never zero.
- Maintenance. A vendor changes a format, a screen moves in the source system, the underlying model changes and the output is no longer quite the same. An agent does not degrade loudly. It degrades in silence, and finding out late costs more than the fix.
And again, the property matters more than the cost: the agent's curve is flat. It does today what it will do in two years, at roughly the same accuracy, learning nothing from what happened in between unless somebody sits down and improves it. In exchange it offers something a person cannot: the margin. Task number one thousand costs about what task number ten cost, and that — not the entry price — is the real argument for an agent.
The two curves, side by side
| Axis | Person | Agent |
|---|---|---|
| Entry cost | High and delayed: recruiting, ramp, months before full output | Low on license, medium on design: weeks until it runs the job well |
| Marginal cost | Linear: twice the volume asks for roughly twice the people | Nearly flat: volume climbs much faster than cost |
| Curve over time | Rising: accumulates unwritten context about the company | Flat: improves only when somebody deliberately improves it |
| Facing the unforeseen | Improvises, asks, sometimes decides badly, but resolves it | Fails, and sometimes fails with apparent confidence |
| Main risk | Leaves, and takes what they knew with them | Degrades in silence and nobody notices until there is a problem |
| What it demands of you | Management: goals, feedback, a career path | Engineering and vigilance: a written procedure and somebody reviewing |
The three cases where the agent wins outright
There are situations where the comparison is not close, and it is worth saying so plainly. The agent wins when these three conditions hold, and it wins harder the more of them hold at once.
- High volume of one task with stable rules. If the same procedure runs hundreds of times a month and the variation between cases is small, the agent's flat marginal cost is unbeatable: the person scales in a straight line and the agent does not.
- Work nobody wants to do, the kind people quit over. When a role exists only to hold up a mechanical task, hiring for it is buying turnover. The agent does not get sick of repetition, which is precisely what burns out a person with judgment.
- Peaks and hours a payroll cannot cover. Irregular demand, weekends, three time zones. Sizing headcount for the peak means paying for the peak all year; the agent covers the peak and switches off.
The three cases where hiring is still the answer
And the other way around. There are roles where deploying an agent is buying an expensive problem that looks like a saving.
- When the exception is the job. If half the cases are odd ones, the agent spends more time escalating than resolving and supervision eats the gain. A process with many exceptions is not an automatable process: it is an undefined one.
- When you need accountability, not just execution. Decisions that commit money, reputation or compliance need somebody who answers by name. You can automate preparing the decision; you cannot automate whoever signs it.
- When the value is in the relationship. A large account, a negotiation, a team that needs leading. What you buy there is not executed tasks: it is accumulated judgment and trust, and that still has no flat version.
The right question: what share of your process is predictable
If you keep one variable out of all this, keep this one. Not "what does each option cost", but what fraction of the cases entering that process gets resolved by following a procedure you already know how to write down. That percentage is the ceiling on what the agent can take, and what is left over is exactly what needs a person. When it is ninety, hiring somebody to do the ninety by hand is throwing money away. When it is forty, deploying an agent and walking off is buying yourself a new bottleneck.
The uncomfortable bit is that almost nobody knows their percentage, because nobody has looked at the process through that lens. Measuring it is not hard — take a month of cases and sort them into "followed the expected path" and "somebody had to decide something" — but it does have to happen before the decision, not after. The order and the criteria for doing that sorting are in which processes to automate with AI, which is the step that turns this argument into a defensible decision.
Why this is almost never replacement — it is a split
The question gets framed as either/or — either you hire or you automate — because that is how it fits in a budget, but the outcome is almost never one of the two in pure form. What works in practice is splitting the role: the agent takes the predictable volume, the person keeps what requires judgment plus supervision of what the agent does. The result is usually not one head fewer; it is the same head doing higher-value work and absorbing growth that would otherwise have demanded a second hire.
Said in P&L terms, that means something worth admitting up front: most of these projects do not lower a cost line, they raise the capacity of the line you already pay for. That is good news, but only if you say it at the start. Sold as savings, it will be judged a failure at six months even if it worked. And who does the work — your own people, outsiders, or a mix — is a separate problem, covered in AI consultancy, in-house team or freelancer.
How to decide it in one afternoon
You do not need a study. You need an afternoon and the discipline not to skip a step.
- Take a month of real cases from that process and split them into predictable and exceptions. Write down the percentage. That is the variable.
- Cost the human side in full: loaded labour cost, not salary, plus recruiting, ramp and an honest turnover estimate.
- Cost the agent side in full: license and usage, plus design hours, plus monthly supervision, plus a maintenance reserve.
- Look at projected volume growth over two years. If volume is growing, the agent's flat margin outweighs any difference in entry price.
- Write one sentence about what happens to the freed capacity: more volume, new work, or lower cost. If you cannot pick one of the three, the project is not ready to start.
If that exercise says a good share of the role is predictable and what remains deserves a person with judgment, the job is building the split: what the agent does, who supervises it, and what metrics prove it. That is exactly what we ship as AI workforce — specialised agents with a human owner, an SLA and a cost per task — precisely because the failure mode here is not technical, it is governance.
And the line worth bringing to the next meeting where somebody puts eighty thousand next to two hundred a month: you are not buying cheaper hours. You are buying a different way to fail. Pick the one you know how to manage.