The seat is dead: what happens when software bills like electricity
As vendors reprice AI features on tokens and consumption instead of seats, the most predictable line in the opex budget is starting to behave like a utility bill - and finance processes built for annual renewals aren't ready.
For fifteen years, the software line was the easiest thing in the operating budget to forecast. You knew the seat count, you knew the per-seat price, you knew the renewal date, and the accrual was the same every month until it wasn't. That arrangement is quietly coming apart. As vendors attach AI features to their platforms and price them on tokens, calls and compute rather than headcount, the most predictable item in opex is starting to behave like a utility bill - variable, consumption-driven, and largely outside the control of the person who signed the contract.
The billing model changed before the budget process did
The shift is no longer hypothetical. PYMNTS ran a piece in February 2026 headlined "CFOs Scramble as AI Pricing Breaks Traditional SaaS Billing Model," and followed it in April with a sharper framing of the internal dynamic: engineering and product teams optimizing for capability with no unit-economics guardrail. TheStreet picked up the same thread this week under the heading that AI agents are creating a new problem for enterprise software. The common thread across all of it is not that software is getting more expensive - it is that software is getting less predictable.
That distinction matters more than the absolute number. A finance organization can absorb a 20% price increase; it plans for it, negotiates against it, and books it. What it cannot easily absorb is a line item whose monthly value depends on how many autonomous agents a product team spun up last Tuesday and how verbose those agents were. Per-seat pricing had one enormous virtue from a controllership perspective: the cost driver was a number finance already tracked. Headcount is governed, approved and visible. Token consumption is none of those things.
ERP Today reported in April that 88% of CFOs say cloud spend is rising, framing it as a cloud cost squeeze. Treat that figure with appropriate caution - the sponsor and sample behind it warrant verification before anyone puts it in a board deck - but the direction is corroborated elsewhere. Gartner's February release on 2026 CFO budget plans showed technology and AI among the priority categories for incremental spend, and CFO Dive reported that Gartner subsequently raised its 2026 IT spending forecast on AI demand. Budgets are being approved into a market where the pricing mechanism itself is unstable.
Three things break in the finance process
The first is forecasting. A consumption line cannot be straight-lined, which means the monthly accrual stops being a copy-paste exercise and becomes an estimate with real variance. If your close process currently books software on a twelfth of the annual contract value, that method will produce true-ups you have to explain. Reforecast cadence has to tighten around the volatile portion of the spend, and the FP&A team needs at least a rough consumption model - expected agent runs, average cost per run, seasonality of usage - rather than a contract value divided by twelve.
The second is contracting. Renewal math built on headcount growth is the wrong tool. The negotiation now turns on commitment tiers, overage rates, rollover of unused commitment, price protection on the underlying model costs, and whether the vendor passes through its own compute deflation or pockets it. Committing to a volume tier for a discount is only rational if you can forecast the volume - which brings you back to the first problem. Many finance teams will find themselves buying commitment they cannot yet justify, simply to avoid list-price overage.
The third, and the least discussed, is accountability. Nobody owns a token budget the way a department owns headcount. Seats had an implicit throttle: a manager had to hire someone before the cost appeared. Consumption has no equivalent gate, so the spend lands centrally, in IT or in a shared services cost center, disconnected from the business unit generating it. Until consumption is allocated back to the team that triggered it - with a visible budget and a named owner - there is no behavioral brake at all.
The quiet accounting consequence
There is also a capitalization question worth raising early, before it surfaces as a variance explanation. A traditional software implementation involved meaningful configuration and development cost, a portion of which was capitalized and amortized. Consumption fees for AI inference are almost certainly period operating expense. The practical effect is that a shift from build-and-capitalize to consume-and-expense moves cost above the EBITDA line, even where the total cash outlay is flat or lower.
That is not a reason to avoid consumption pricing, but it is a reason to flag it to the audit committee and to anyone modeling margin on a multi-year basis. If your operating margin guidance assumes a technology cost profile that included capitalized implementation spend, the substitution of usage fees will show up as margin compression that has nothing to do with commercial performance. Get ahead of the narrative.
What the vendors are offering back
The market is responding, if unevenly. PYMNTS reported in July on agentic payment infrastructure pitched explicitly as a way to get "usage pricing with subscription discipline" - effectively, spending controls, pre-authorization limits and per-agent budgets applied at the transaction layer. Whether that is a genuine control or a way to make consumption pricing palatable enough to accept is a fair question, but the existence of the pitch confirms that vendors have heard the complaint.
CFOs should press for three things in any consumption contract: real-time usage visibility that finance can access directly rather than through a monthly invoice; hard caps or alert thresholds that can be set by cost center, not just at the tenant level; and clarity on what happens to the rate when the vendor's own model costs fall. The third is where most of the long-term value sits. Underlying inference costs have been declining steadily; a contract that locks a per-token rate for three years may look prudent today and expensive by 2028.
A practical starting position
The immediate work is unglamorous. Inventory every contract that has a consumption component, including ones where the AI feature is currently bundled and free - those are the next repricing events. Separate the fixed platform fee from the variable usage fee in the budget, and forecast them differently. Assign an owner and a monthly ceiling to each material usage pool, and make the owner explain overages in the same way they would explain an unbudgeted hire.
Then decide, deliberately, how much variability you are willing to carry. Some consumption exposure is worth it: paying only for what you use is genuinely cheaper than paying for idle seats, and for workloads with unpredictable demand it transfers risk to the vendor. But the trade is predictability, and predictability has value in a public-company forecast. The organizations that manage this well will be the ones that treat software the way treasury already treats commodity exposure - as a variable input to be measured, hedged where sensible, and owned by someone specific.
Key takeaways
- Stop straight-lining the software accrual. Split contracts into fixed platform fees and variable usage, and forecast the variable portion with a consumption model rather than contract value divided by twelve.
- Renewal negotiations now hinge on commitment tiers, overage rates, rollover terms and price protection as underlying inference costs fall - not on projected headcount.
- Consumption has no natural throttle. Allocate usage to the business unit that generates it, with a named owner and a monthly ceiling, or the cost lands centrally with no brake.
- Usage fees are period opex where implementation spend was often capitalized; flag the EBITDA effect before it appears as unexplained margin compression.
- Demand real-time usage visibility and per-cost-center caps in the contract itself - a monthly invoice is not a control.
