Finance Will Budget for Capacity, Not Headcount

In a McKinsey analysis published in August, tokens often made up just 20 to 25 percent of the variable cost of an AI agent doing customer service work in banking. Human oversight made up 70 to 75 percent.

Finance work is review-heavy too, so I would expect a similar split in the close. The visible cost of AI is the model. The larger cost is the people who check its work. A headcount plan hides both. The compute usually sits in the IT budget, and the review disappears into salaries finance already pays.

I think that changes how finance budgets, not only how it buys. Once agents run long stretches of the close, collections and reporting, the unit finance plans in stops being heads or seats. It becomes capacity. Whoever owns capacity decides which AI stays, and I expect that to be the CFO more and more often.

A headcount ratio measures the wrong thing

Finance has always counted itself in people: heads, seats, finance staff as a share of the company. I expect that ratio to keep falling as AI takes on more of the work.

My problem is with the unit. A headcount ratio shows how fast work leaves the payroll. It says nothing about whether the close is right, whether the forecast holds, or what the work now costs. A team can hit its target by handing reconciliations to an agent that a senior controller quietly rechecks every night. The ratio improves. The cost of finance may not.

The work has not gone. It has moved onto a usage bill and into someone’s review time. That is capacity, and it is what finance should budget.

Capacity has three parts, and cheaper tokens reset none of them

In an agent-run finance team, I see capacity in three parts:

• Throughput: the reconciliations, collection follow-ups and variance explanations the agents complete. Agents that chase overdue invoices and match payments, the kind Paraglide builds, add capacity that no headcount ratio records.
• Reviewer hours: the time people spend checking that output and fixing what is wrong.
• Compute: on its way to becoming a real line in the P&L, not a rounding error inside the IT budget.

Falling token prices do not reset this meter. The McKinsey authors name the paradox: “While the cost of individual tokens continues to fall, enterprise spending on AI is spiking.” A cheaper model saves nothing if its work has to be redone. An agent that is almost free per run but sends every third reconciliation back to a controller can cost more than the work it replaced.

The three parts trade off against each other, and against contractors and overtime. In most companies, only one executive sees all of them on one page. That is why I think AI economics belong with the CFO.

The payback question comes at the second renewal

In my view, finance AI is easy to renew the first time. The pilot has worked, a sponsor vouches for it, and the spend sits in the technology budget where nobody has to defend it. The second renewal is where the payback question gets asked, and I expect it to land on the CFO’s desk.

Payback follows the capacity meter, not the go-live date. I expect reviewer hours to run high in the first months of a rollout. Until the people who sign trust what the agents produce, the team does much of the work twice. On the meter, finance then costs more than before the tool arrived. The saving appears only when review time falls and the manual version is switched off. That lag is normal. What is changing is who gets asked about it.

The CFO’s remit is moving the same way. In the IBM Institute for Business Value’s 2026 CFO Study, a survey of 1,500 CFOs and equivalent senior finance leaders, 62% said their role had expanded into enterprise technology or AI strategy leadership over the past two years. Only 8% said finance leads the enterprise-wide AI value ambition and has established automated investment triggers tied to results.

As I read them together, those figures describe the CFO becoming the gatekeeper for AI spend, mostly before the means to judge it exist. I suspect many companies never wrote down a starting point. When budgets tighten, an AI tool whose effect the CFO cannot see gets treated like any other unproven cost.

A smaller team is a capacity cut, not a saving

Headcount is, I think, the easiest number to move in front of a board, which is why it proves so little. Removing people removes capacity from the payroll. Whether anything was saved depends on where the work went:

• onto compute;
• onto contractors;
• onto the evenings of the senior people who stay.

The last one hurts most, because senior time is also what the next controller learns from.

I think many job cuts blamed on AI are really paying for compute. That is a financing decision, not proof that the technology works. The capacity budget falls only when the work itself shrinks, because a handoff disappeared or a manual workaround was retired. A smaller team is then the result, not the method.

A forecast back in hours instead of days is worth having, and so is a quarter end without weekend work. Neither is payback the P&L can show. Folding them into the payback number is how AI budgets lose credibility.

What renews is what finance can point to

Budget season favors value the CFO can find without help. If an agent’s work shows up in a number finance already reports every month, it defends itself. If it needs its own slide and a sponsor to explain it, it gets argued over every year until someone stops arguing for it. The full cost belongs on the same page, review hours included.

I sat on the vendor side of that conversation at CXO Software, as CTO and later running the Americas business. At one retail group, the controller could name what changed: four days of board-pack assembly saved every month. That renewal took one call.

I expect AI renewals to turn on the same question: can someone in finance say what changed? Anyone who has run a post-merger integration knows the gap. Synergies promised at signing and synergies that reach the P&L are different numbers, and only the second survives the next budget.

Prices will follow the capacity line

Pricing will have to follow the budget. Seat pricing shrinks with the team. Pure usage pricing has the opposite problem: it hands the CFO a meter that runs fastest exactly when adoption succeeds, which makes it hard to plan and easy to cap. What I see gaining ground is outcome-linked pricing: a predictable price tied to something finance can see, such as a close completed, invoices processed or cash collected.

That also changes how to read adoption claims. A pilot is cheap to start and cheap to stop, so a count of pilots says little. A contract that survived the CFO taking over the budget says much more.

Next year’s finance budget will still have a headcount line. I expect the better ones to carry a second line beside it: the work the agents did, the hours people spent checking it, and what the compute cost. That line shows whether the capacity a company bought replaced the capacity it used to hire. The AI that shows up on it keeps its budget.

Wouter Born is the founder and General Partner of Born Capital. He writes AI CFO Office, a newsletter on AI for CFOs and finance leaders with more than 50,000 subscribers.