For most of my career, the financial template for a program had two columns and one equation: FTE × effort = cost. You sized the team, you priced the people, you added some overhead, you delivered a number. Everyone — me, finance, procurement, the CFO — understood the same picture.
In the digital transformation program, that template started returning the wrong number within the first month. Not by a little. By enough that I had to rebuild it from the ground up.
What’s missing from the old equation
The old equation captures human time. It doesn’t capture token consumption — the per-call cost of every prompt and response the agents make. It doesn’t capture model selection — the difference between a small model at a fraction of a cent per call and a frontier model at fifty times that. It doesn’t capture agent runtime — the cost of an agent that spends three minutes deliberating before producing output. It doesn’t capture experimentation — the prompts and runs you spent figuring out that the approach didn’t work.
Add it all up and you can run twenty percent over the FTE × effort budget while still being under the actual delivery cost — or vice versa, depending on how the program runs. It’s not noise. It’s a different variable in the equation.
The new shape
The financial template I now build has four columns instead of two:
Human cost. Same as before. Roles × rates × time, plus burdens.
Consumption cost. Token spend, model usage, agent runtime. Forecast monthly, reconciled weekly. Has its own line in the P&L.
Experimentation overhead. A reserve, typically ten to twenty percent of consumption, that covers the prompts, runs, and approaches you spent figuring out what works. The CFO will ask why this is separate from consumption. It’s separate because you want to know how much of your spend is going to runs that ship versus runs that taught you something.
Oversight cost. The human gate cost. Reviewer time, governance forum time, audit time. This grew, not shrank, in the AI-first model. Worth surfacing as its own line so it doesn’t hide inside "human cost."
The four columns are not arbitrary categories. They are the cost structure of the new operating model: agents drive the work; humans navigate it. Consumption and experimentation are what the agents cost. Oversight is what the navigators cost. Human cost is what the humans doing non-navigation work (architecture, decision, design) cost. Once the financial template names both species of teammate, the CFO can see where the leverage is, and where the unmanaged exposure is.
Why margins need recalibration
In a traditional delivery program, your margin lives in the gap between rate-card revenue and FTE cost. If you bid a program at fifteen percent margin, you have a knowable cushion against estimation error.
In an AI-first program, consumption cost is volatile. A model upgrade that comes out mid-program can drop your per-call cost by a third — or double it. An agent that drifts and starts retrying calls can eat your reserve in a week. If you priced with a 2024-style margin, you can hit the project end on time and still have lost money because the consumption line ran hot.
I now recalibrate margins under two assumptions. Floor margin — what the program returns if consumption runs ten percent over forecast. Realistic margin — what it returns at forecast. Optimistic margin — what it returns if a model upgrade lands and consumption runs under. The CFO sees all three. The bid is priced to the floor.
The forecasting cadence changed
Annual budgets and quarterly forecasts can’t hold AI consumption. The variables move too fast. Pricing curves change. Models get cheaper or get retired. New agents come into the workflow. The consumption side of the financial template gets a weekly review, not a quarterly one. The human side stays on its old cadence.
This dual-cadence pattern feels strange the first time you run it, but it converges fast. Within two months on the retail program we had a weekly consumption rhythm that felt natural and a monthly human-cost rhythm that felt routine. The two met at the steering committee.
The CFO conversation
If your CFO came up through the cloud era, they already know how to think about consumption — they’ve done it with AWS for ten years. Frame AI consumption as a cousin of cloud spend and you will get a much faster yes than if you frame it as a new category of unknown.
If your CFO came up before that, the conversation is harder. The mental model needs to be built. The artefact that has worked for me is a one-page comparison: same program, costed two ways. FTE × effort gives you one number. The four-column model gives you a different number — typically higher in the first quarter (when experimentation is heavy) and lower from the second quarter on (when agents are doing volume work that humans would otherwise have done). Showing both side by side, with the actuals each month, builds confidence in the new shape faster than any deck.
The financials don’t go up or down. They change variables. The CFO’s job is the same; the equation isn’t.
The practical starting point
If you are running an AI-bearing program today and you are still costing it with a two-column FTE template, redo the financials with a four-column template before your next steering committee. Even if you have to estimate consumption in the first version, the act of putting it on the page changes the conversation. The next month you have actuals. The month after that, a forecast.
It’s the cheapest leadership upgrade in this whole transition. And it stops the post-hoc surprises that come from a cost line you didn’t name.