Commercials · 7 min read · Aug 19, 2026
Three cost shapes, not three prices.
The finance manager is at capacity and the group is opening two more outlets. The three options are not three prices for the same thing — they have different shapes, and the shape is what determines which one is right.
The comparison people make
The usual analysis puts a loaded salary next to a monthly agency retainer next to an estimated build cost, and picks the smallest number. It is the wrong comparison, because the three options behave differently as volume moves, and volume is the only thing you can be confident will change.
Three shapes
| Headcount | Outsourcer | Agent | |
|---|---|---|---|
| Cost curve | Step function — nothing, then a whole person | Roughly variable, with a floor | Build cost up front, then running cost plus supervision |
| Time to useful | Weeks to hire, months to competence | Days to weeks | Weeks to a pilot, longer to trust |
| Handles the unexpected | Well — the main argument for it | Within scope; escalates outside it | Badly, unless the exception path was designed |
| Where knowledge sits | In a person, who can leave | With the provider, and leaves with them | In code and prompts you own, if the contract says so |
| What happens at 3x volume | Hire again | Bill goes up roughly linearly | Marginal cost barely moves |
| What happens at 0.5x volume | Cost stays | Bill falls, floor remains | Running cost falls; build cost is spent |
Read down the last two rows. That is the actual decision. An agent is a bet that volume holds or grows and that the process stays roughly stable; a person is a bet that judgement and variation matter more than throughput; an outsourcer is a bet on neither, bought for flexibility, and priced accordingly.
Supervision is a permanent line, not a transition cost
The most common error in an agent business case is treating oversight as a first-year expense. It is not. Someone reviews the exception queue, samples what was auto-approved, and re-tests after model changes, for as long as the system runs. Budget it as a standing share of a role rather than a project line, and the comparison becomes honest.
An agent does not remove the work. It changes who does what: the routine volume moves to the system, and the person moves to exceptions, judgement and the cases the system was designed to hand back.
Volume and variance decide it
Two axes settle most of these decisions faster than a spreadsheet:
- High volume, low variance — invoice capture, statement reconciliation, item normalisation, report assembly. Agent territory. The output is checkable and the same shape every time.
- Low volume, high variance — supplier negotiation, unusual transactions, anything requiring a relationship or a judgement call. A person, and automating it is a way to spend money slowly.
- High volume, high variance — split it. The stable share goes to an agent, the residue to a person whose day just got more interesting.
- Spiky, seasonal or temporary — an outsourcer. You are buying elasticity, which is precisely what a build does not give you.
Does it replace the finance team?
In the deployments that reach production, the pattern is displacement of tasks rather than of people: routine processing volume falls, exception handling rises, and the roles shift toward review and analysis. That is also why the payback is slower in finance than in sales — the outputs touch audit trails and contractual obligations, so organisations deliberately keep humans in the loop. The figures are in where back-office agents pay back.
The honest version for a group director: expect the same headcount doing higher-value work before you expect fewer people. If the business case only works through redundancies, it probably does not work.
What to compare, if you compare properly
Put all three on the same terms over three years: total cash out, supervision load in hours, what happens at your realistic high and low volume, and what remains if the arrangement ends. That last column is where the differences are widest — a departing employee takes the knowledge, an ended contract takes the provider's process with it, and a built system leaves whatever the contract says you own.
Which is the same reason ownership sits near the top of the questions to ask an AI vendor, and why the running-cost lines are worth sizing before signature rather than after — those are set out in what an AI operating system costs to run.
Dealing with this in your own group?
We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.