38 percent AI growth inside a 2 percent budget

On one service and support budget sheet the total line grew 2 percent and the AI line grew 38 percent. Gartner surveyed 199 service and support leaders between April and May this year and found exactly that combination. The two numbers are not in conflict, because the AI line is still small enough that its fast growth fits inside an envelope that barely moves.

Do the arithmetic and the size becomes obvious. If AI is 2 percent of the envelope and the rest grows at 2 percent, total growth is 2.7 percent, which matches the 2 percent on the sheet. The 38 percent does not hurt today because the denominator has not been bitten yet.

What actually changed is not the growth rate. It is where the money sits. The same AI program came out of an innovation budget last year and comes out of a departmental budget this year. Move the line, and the people who sign, the people who review, and the people who get asked when something goes wrong are all different people.

Zero sum inside a fixed envelope

Innovation money and departmental money behave differently. The first is usually a central pool allocated by quarter, and review asks whether the idea is worth trying, with attention on how much failure is affordable. The second is an annual fixed amount, and review asks what the unit cost becomes once the money is spent. When a program moves from the first to the second, proof of capability stops being a passport. Unit cost is.

A fixed envelope means every addition corresponds to a subtraction. That correspondence usually goes unrecorded in the procurement process: which line funds the AI line, and who signed it, are not columns on the form. In the same body of Gartner research from March, more than half of IT leaders worry about cost overruns while only 44 percent of organizations have put financial guardrails or AI FinOps practices in place. Those two facts are one fact seen twice. The money is growing and the rules for it are not.

No rule does not mean no behavior. Money takes the easiest path: whichever approval moves fastest, or whichever bill folds into an existing cloud contract and stops being reviewed on its own. By the time finance notices one curve steepening at quarter end, the only action left is a cap.

Consumption pricing turns success into cost

Software sold by seat has a cost that arrives in steps and can be forecast. AI sold by consumption inverts that: the more it is used, the more the bill grows. One consequence rarely reaches the business case: under this pricing, the more successful a workflow is, the larger the bill it hands the department.

So the cap arrives first at the most successful place. The business side asks for more seats and more access so more people can use it. Finance sees the monthly curve steepen and reaches for an account level spend cap. A cap on the account cuts the calls with the highest usage, and the highest usage is usually not waste. It is the long context, multi turn, high value work. On a common distribution, where 15 percent of units consume four times the tokens of the rest, those units account for 41 percent of one account token bill. Trigger the cap and that 41 percent goes first.

That ordering belongs on the risk register. Cost control actions get sorted by how easy they are to execute rather than by value, so the most valuable use is cut first and what remains is low volume and low value. The bill falls. So does the return.

A ninety percent price cut does not fix the ledger

Inference prices did fall sharply over two years, and many teams put their savings case there. Cut the unit price by ninety percent, grow usage tenfold, and the bill still rises. The harder part is that usage grows because the product got better: more people, longer sessions, longer contexts, more retries.

The real lever is visibility, not price. What one workflow spent last month is a number most organizations cannot produce. Without it, cost cutting can only happen along the easiest axis, which is accounts, models and quotas, and never along value. The problem with a cross cutting axis is that it does not know what it cut.

The savings were booked first

There is also a timing gap. In the Gartner CFO survey from February, planned growth for the HR budget fell from 2.4 percent to 0.7 percent, and the stated reasons were reduced hiring and AI efficiency gains. Read that carefully. The AI savings have already been booked into somebody else budget as an accomplished fact, while the AI cost is still settling monthly with no owner.

Savings booked first, costs arriving later, and every report in between looks fine. The reconciliation happens at the next budget round, when both sides ask for delivery at once: one side has to show the promised saving happened, the other has to explain why the bill is climbing.

What one workflow actually costs

Take a concrete process. Eight thousand units a month at 4.5 dollars fully loaded labor per unit. After AI, 45 percent of units run without a person, 55 percent still go to staff, model cost is 0.62 dollars per call, and every automated unit still gets 25 seconds of human review at 48 dollars an hour.

The ledger reads: 4,960 dollars of model spend, 1,200 dollars of review, 19,800 dollars of remaining labor, 25,960 dollars in total against 36,000 dollars before. That is 10,040 dollars saved, a 27.9 percent reduction, and unit cost falls from 4.5 to 3.25 dollars. Those parameters are mine, set to show structure, and you can recompute with your own.

Two features of that ledger matter. Model spend is 19 percent of the total, review is 5 percent, and labor is still 76 percent. The cost of this workflow sits mostly outside the AI line, so cutting the AI line saves little and costs a lot: remove the 4,960 dollars of model spend and the 10,040 dollars of savings go with it. Second, unit cost is sensitive to the automation ratio. Drop from 45 to 35 percent and labor rises 1,800 dollars, savings fall by nearly a fifth, and failed automation usually adds rework on top.

Track cost by workflow, not by account

A cost ledger has to answer four questions: which workflow is this, what is the completed unit, what is the fully loaded cost per unit, and whose budget does it hit. Fill all four and cost cutting has a target. Leave one blank and the only available action is an account level cap.

The completed unit is the column most often filled with call counts. Call counts are the supplier unit, not the buyer unit. The buyer unit is work finished: a claims file processed, a ticket closed, a reconciliation completed. Use call counts and the optimization target silently becomes fewer calls, and the cheapest way to fewer calls is for the agent to do less, which is the lowest value direction available.

The owner column is the hardest, because it wants a role, not a department. Inside the business unit, the owner has a reason to push unit cost down. Inside a shared service center billed per item, pushing cost down pushes revenue down with it. Inside IT, it is platform cost and the metrics are availability, not unit cost. Three homes, three behaviors. Any of them works, but one has to be chosen and written into the budget document, or the line becomes the thing everyone assumes someone else is watching.

Put the cap where reversibility is

Spend caps are not the problem. The placement is. An account level cap is the last resort, not the first line. The first line belongs on the workflow, aligned with reversibility: read only and replayable workflows can take a loose cap and queue when they exceed it; workflows with external side effects need a tight cap that stops and notifies; irreversible workflows need a confirmation gate rather than a cap, because throttling an irreversible action only delays it.

Placed that way, cost pressure lands where it can be absorbed. The cost is that someone has to classify workflows by side effect, which does not automate, only gets done one at a time, and has to be maintained as processes change. It looks like governance work. It is accounting work. Misclassify and the bill explodes in the wrong place.

The cross department mismatch

Attribution has a cross department version too. A support agent that explains the return policy cuts ticket volume in support, while the benefit lands in sales margin through a lower return rate. The bill sits in the support cost center and the gain sits in sales margin. In that mismatch the most rational move for the support leader is to suppress usage, even when the company level ledger is positive.

The reverse happens as well. A quoting assistant on the sales side blocks misconfigured orders before they are placed, saving fulfillment and after sales rework, while the bill lands in the sales expense line. Neither department can make its own numbers work, only the company level number is positive, and nobody does the monthly roll up because a monthly roll up appears in nobody performance review.

The answer is not asking someone to be generous. It is writing the beneficiary into the allocation rule: whichever department sees lower ticket volume or lower rework funds part of the line. Set that rule before procurement, because the conversation before signing is about structure and the conversation at renewal is about blame.

Write the cap into the contract

One more placement for the cap gets ignored: the contract. Most AI purchases are signed as a unit price plus a committed volume, with no overage rate and no ceiling clause. Usage above the commitment bills at the same unit price, so finance sets a hard cap in the console, and a hard cap firing means an outage, usually in whichever workflow was performing best that month.

A better shape has two tiers: one price inside the committed volume, a higher price above it, and a right to adjust the commitment once a quarter. Cloud procurement has worked this way for years and model calls mostly have not. The buyer trades an outage for a price increase, and the seller avoids having growth cut off with a hard edge.

The cost attribution endpoint belongs in the same contract. Breaking the bill down to the workflow means the supplier changes metering and the buyer passes a workflow identifier, so both sides change and the ask only works at signing. At renewal it is six months late.

There is a rhythm problem as well. Budgets are annual and usage moves by month and by project. A large customer goes live mid year, usage doubles, and the envelope does not double with it. That leaves two options: an in year top up, or pushing usage into the next fiscal year. The second is technically possible and usually unacceptable commercially, so the top up becomes the default path, and a top up faces harder scrutiny than the original budget because it has to explain why the first forecast was wrong.

Our judgment

The binding constraint on AI scale in 2027 will not be model capability. It will be the allocation rule inside a fixed envelope. Consumption pricing turned success into cost while cost attribution stayed at the account level, so the most successful workflow hits the cap first. Falling unit prices will not remove this, because usage grows faster than price, and growth in usage is itself evidence that the product is being used.

Three moves need no new tooling. First, attribute AI cost by workflow, with completed work as the unit instead of call counts. Second, set caps by side effect class per workflow and keep the account cap as a backstop. Third, add one line to the approval document naming which budget line funds the AI line and who signed it. All three finish inside a month, and the next budget argument moves from impressions to numbers.

Two predictions that can be checked. First, AI products will ship configurable usage caps priced per workflow rather than per tenant next year, because procurement will use that as a lever. Second, suppliers will be asked for cost attribution endpoints that break the bill down to the workflow, and the ones that cannot will drop out at the comparison stage.

Sources

AI spending by service and support leaders up 38 percent while the overall function budget rose 2 percent, sample of 199 leaders surveyed April to May 2026: Gartner press release, 26 August 2026, primary. gartner.com 2026-08-26 press release

Only 44 percent of organizations have adopted financial guardrails or AI FinOps practices, and more than half of IT leaders worry about overruns: Gartner Data and Analytics Summit, March 2026, reported by MarketScale, secondary. marketscale.com summary of Gartner 2026 research

HR budget growth falling from 2.4 to 0.7 percent on reduced hiring and AI efficiency gains: Gartner CFO budget survey press release, 10 February 2026, more than 300 CFOs and finance leaders surveyed in October 2025, primary. gartner.com 2026-02-10 press release

Worldwide AI spending forecast at 2.52 trillion dollars in 2026, up 44 percent, led by AI infrastructure: Gartner press release, 15 January 2026, primary. gartner.com 2026-01-15 press release

The 41 percent share assumes 15 percent of units consume four times the tokens of the rest, computed as 0.15 times 4 divided by the sum of 0.15 times 4 and 0.85. Workflow parameters in the worked example are eight thousand units, 4.5 dollars labor per unit, 45 percent automation, 0.62 dollars model cost per call, 25 seconds of review, and 48 dollars per review hour. All are my assumptions for showing structure, not industry averages.