There is a comfortable story the AI category is telling itself right now: agents are cheaper than humans, the curve is bending, the math works. I have been building and operating agentic systems for long enough to say, plainly, that the comfortable story is wrong about the most important part and that the part it gets wrong is exactly the part a CFO is going to ask about in your FY26 budget review.
This is not an argument that AI is bad, or that agents don't work, or that you shouldn't be building. I run a company whose product IS agentic infrastructure. The agents work. They do things humans cannot do at any price. That is real.
But the economics aren't industrial yet. And until they are, the way most companies are budgeting for agentic AI is going to produce some very uncomfortable conversations in finance reviews this year.
The thing nobody is saying out loud
Compare three kinds of inputs in a typical revenue tech stack.
A human SDR costs roughly the same per year whether they make 30 calls or 300. The cost is capped, contractually known, budgetable to the dollar a year in advance. Finance can model headcount capacity with high confidence.
A SaaS tool costs per-seat or per-tier with contracted usage caps. The bill at year-end matches the bill in the budget within a small variance. Finance models this without breaking a sweat.
An AI agent costs whatever its tokens cost to do whatever work it does and the per-task token consumption depends on context length, reasoning depth, tool call count, retry behavior, and which model you're routing to that week. The same workflow that ran for $0.10 last year on a previous-generation model now runs for $0.40 on a current one, because better models use more tokens to do better work. Per-token prices fell. Per-task cost rose. The bill at year-end does not look like the bill in the budget.
AI tokens are not a third item on the same axis as humans and SaaS. They are a different class of input. Budgeting for them as if they were just another SaaS line item is the mistake.
That last point is doing more work than it looks. A CFO who has spent twenty years budgeting humans and SaaS knows how both behave. They have never ever budgeted a line item whose unit cost reliably rises as the underlying technology improves. That is not how technology is supposed to work. But it is exactly what agentic AI is doing right now, and pretending otherwise is what makes the variance show up in Q3.
Why per-task cost is rising even as per-token cost falls
Three things are happening simultaneously, and the AI category prefers to talk about only the first one.
- Per-token prices are falling. True. Frontier model pricing has come down meaningfully over the last 18 months. This is the slide every AI vendor shows you.
- Per-task token consumption is rising faster. Newer models use longer context, more reasoning tokens, more tool calls per task. A workflow that took 2,000 tokens to complete on a 2024 model takes 20,000 tokens on a 2026 reasoning model because the reasoning IS the improvement. The total bill rises. This is the slide AI vendors don't show you.
- Workflows themselves are getting more complex. As agents get more capable, teams ask them to do more sophisticated work. A single “qualify this lead” agent in 2024 was one model call. In 2026 it is a multi-step agentic flow with retries, tool calls, and verification. Cost per useful unit of work goes up, even when each individual call gets cheaper.
Net effect: the curve isn't bending in the direction the category is selling. Per-token cost is following a normal commodity curve. Per-useful-unit-of-work cost is following a consumption-creep curve, and consumption creep is winning right now.
Why this isn't an indictment it's a timing observation
None of this means agents don't work. It means agents are not a utility yet.
Electricity is a utility because pricing is predictable, declining, contractual, and the per-unit cost has been stable enough for long enough that you can build a budget against it forty years out. Cloud compute became utility-like around 2014, after roughly eight years of variable pricing and surprise bills (anyone who lived through early AWS remembers this).
Agentic AI in 2026 is roughly where cloud compute was in 2008: genuinely transformative, definitely the future, and economically unpredictable. That is not a contradiction. It is the normal shape of a foundational technology before it commoditizes. Treating it as if it has already commoditized is the mistake.
Three things have to happen before agentic AI becomes a utility input:
- Per-task consumption has to stabilize, so reasoning improvements stop ballooning token use.
- Providers have to offer truly predictable contracts, fixed-cost-per-workflow or capacity reservations rather than per-token metering.
- Batched and scheduled workloads have to get utility pricing the way off-peak electricity does.
None of these is here yet. Some are coming. The smart guess is 12 to 24 months for the first signs and probably longer for the full transition. Until then, you are buying a foundational input that has not finished pricing itself.
So what does a CFO do with this in FY26?
Here is the practical part the thing the rest of the post has been building to. You can't wait for agentic AI to commoditize before you deploy it; the capacity upside is too large to defer. But you also can't budget for it like SaaS without producing a variance Finance will not forgive. Three moves that work in the interim:
MOVE 1: Budget agents as a separate line item, not under SaaS.
Stop nesting agent tokens inside your existing SaaS budget. Tokens behave nothing like SaaS the variance profile is different, the unit economics are different, the bill-prediction model is different. Give agents their own line item with explicit assumptions about per-task consumption and expected workflow volume. When the variance comes and it will it shows up in a place finance can isolate and reason about, instead of contaminating your SaaS forecast.
MOVE 2: Budget a variance reserve not a buffer, a reserve.
A buffer is padding inside a line item. A reserve is a separately-tracked allocation specifically for the unpredictability of the underlying input. For agentic AI in 2026, plan a 20 to 40 percent reserve on top of your modeled token spend, and report against it explicitly. If you don't use it, you over-reserved fine, it returns to the pool. If you do use it, Finance saw it coming. The reserve isn't pessimism; it is *honesty about pricing maturity.*
MOVE 3: Push providers for outcome-based or workflow-based contracts.
Per-token pricing is the source of the unpredictability problem. Some providers including some category-defining ones are starting to offer fixed-cost-per-workflow or capacity-reserved contracts. These remove the variance. If your AI vendor cannot quote you a fixed-cost-per-defined-workflow today, ask them when they will, and weight that timing in your renewal decision. The vendors who move to predictable pricing first are the ones whose economics will scale into a utility model. The ones who don't are betting consumption creep keeps generating revenue.
MOVE 4: Measure capacity, not spend.
This is the part that connects back to the work itself. The question isn't “what did we spend on AI this quarter,” it's “what capacity did we unlock per dollar of AI spend, and how does that compare to the capacity we'd have had to hire to produce the same output?” Done honestly, that comparison usually favors the AI even at today's unpredictable prices because the capacity an agent produces is capacity humans literally cannot supply (the volume, the always-on, the parallelism). The point is to measure it explicitly so the variance is being held against a real return, not floating in finance's mind as “we spent how much on what?”
Why this still favors building just with eyes open
The honest read of the unit economics in 2026 is not “don't deploy agents.” It is “deploy agents, budget for them like the unstable input they currently are, and make sure the capacity you're unlocking is real and measured.”
Agents do things humans can't. Read every account in your pipeline before every call. Watch every deal for risk signal continuously. Draft every follow-up at scale. Surface patterns across hundreds of conversations a quarter. No headcount you could hire would produce that volume the question isn't whether agents are cheaper than the equivalent humans, the question is whether agents do work that humans were never going to do at any price.
The right comparison is not agent versus human on cost. It is agent versus capacity-that-doesn't-otherwise-exist. On that comparison, even unstable token pricing wins. But Finance has to see the math you're making, or they see only the variance.
That is the doctrine. Build the agentic function. Budget for it with eyes open. Measure capacity, not spend. And push your providers hard on the pricing model, because that is the part that has to change before the math becomes obvious to everyone.
One more honest thing
This post has a half-life. Everything in it is true as of mid-2026. Some of it will be wrong by mid-2027 specifically the variance numbers and the timing estimates. When per-task consumption stabilizes and the predictable-contract market matures, the budgeting moves above stop being necessary and the math becomes utility-shaped.
I will write a follow-up when that happens. The doctrine underneath measure capacity, not spend; treat agentic AI as a different class of input; demand predictable pricing from your providers will still hold. The interim moves will not.
If this resonates and you want the full operating model behind it how to actually build the function, measure the capacity, and run agentic work as infrastructure rather than as a series of pilots read the Revenue Activation Manifesto. That's where the longer argument lives.
.png)