The AI line item is up forty percent over last year. Your CFO has noticed. The board has noticed. You have a budget meeting in two weeks and a paragraph drafted that says, more or less, “we are being responsible stewards of this investment and have committed to holding spend flat through Q4.”
That paragraph will lose you the meeting. Not because the CFO disagrees. Because the moment you stake the defense on the total, you have agreed to a frame in which any future increase is a failure. AI spend is going up. Next quarter, the quarter after that, and through 2028 at minimum. You cannot defend a rising number by promising it will stop rising.
The exec who wins this meeting defends a different thing.
The wrong defense
Most AI budget defenses I see are some flavor of restraint theater. The headcount-normalized number is held up as evidence of discipline. The vendor consolidation slide is presented as savings. The “we evaluated five tools and chose two” story is dressed up as rigor. Each of these reads, to a CFO who has sat through twenty quarterly reviews, as a budget owner trying to look responsible about a spend they don’t actually understand.
The problem isn’t that restraint is bad. The problem is that restraint is the wrong unit of defense for a category that is structurally growing. SaaS spend was defensible on per-seat math because seats were a stable proxy for use. AI spend isn’t. A seat that consumes ten million tokens a month and a seat that consumes zero look identical on the procurement contract and produce wildly different outputs. Defending the seat count tells the CFO nothing they care about.
Worse, the flat-spend pledge concedes the strategic point. If your competitors are letting their AI line item grow with their leverage, and yours is held flat by executive fiat, you are not being responsible. You are being slow. The CFO won’t say that out loud in the meeting. They will think it the next time someone asks why a peer company is shipping faster.
Why the number is going up
The structural reason is pricing. The flat-seat era ended in April when Anthropic dropped the $200 enterprise SKU and replaced it with $20 plus consumption. OpenAI’s Workspace Agents is credit-based on the same logic. Google’s Gemini Enterprise Agent Platform is consumption-priced. Microsoft is the holdout, and Microsoft’s product is the one your people open least. The vendors who are winning have moved to a pricing model that bills you in proportion to the leverage you extract. That model produces growing line items by design. (I walked through the mechanics in Evaluating Spend.)
The behavioral reason is leverage compounding. The person on your team who used Claude to draft client emails last year is, this year, running an automated pipeline that generates the first draft of every proposal in the firm. Their consumption went up by an order of magnitude. So did their output. The right reaction to that pattern is not to cap their token budget. It is to find the next person who can do the same thing, and the next.
The competitive reason is that you do not get to opt out. Your peer companies are buying the same tools, hiring people who already know how to use them, and putting agent platforms into production whether they’re mature or not. The cost of falling behind on AI leverage in 2026 is not the cost of the seats you didn’t buy. It is the cost of the contracts your competitors won because their proposal team is six people doing the work of twenty.
A leveraged organization should see AI spend grow faster than headcount for the next several years. If your AI line item is flat year over year, the most likely explanation is that nothing is changing in how your people work. That is the bad version of restraint.
Defensible means legible, not low
Here is the reframe. Defensibility is not about how much you spend. It is about whether you can point at every dollar, name who is using it, and describe what it produced. A $50,000-a-month AI budget where every dollar is traceable to a named person and a named workflow is more defensible than a $20,000-a-month budget that nobody can explain. The CFO’s job is not to minimize spend. It is to be able to answer the auditor, the board, and the bank when they ask what the company is doing with its money. Give them an instrument they can point at.
Four numbers do most of the work.
Cost per active user, not cost per licensed seat. Total spend on a tool divided by the number of seats with meaningful weekly use. If your headline per-seat number and your per-active number are within twenty percent of each other, you have a healthy deployment. If your per-active is two or three times your per-seat, you are funding shelfware and the audit will say so.
Concentration ratio. What percentage of your AI dollars are landing on the top decile of users by output? In a healthy deployment, the answer is high — sixty, seventy, eighty percent. That sounds bad until you remember that the top decile is producing the leverage. A flat per-employee distribution is the failure mode, not the goal. Spend should follow output.
Captured shadow AI. What you found on expense reports last quarter, and what fraction of it is now on a governed plan. This number going up is good. It means people who were already paying out of pocket are now under your security perimeter, on a tool you can audit. The fact that it ever existed is the leadership signal — your people went around procurement because procurement was selling them the wrong thing.
A named workflow per dollar bracket. For every five thousand dollars of monthly AI spend, you should be able to name one workflow that observably changed. Not “improved productivity.” A specific thing: the proposal pipeline that used to take four people now takes one. The customer escalation queue that used to clear in two days now clears in four hours. If you cannot name the workflow, the dollars are not producing leverage and the bracket should be reallocated.
These four numbers do not justify the total. They justify the shape. The CFO conversation moves from “is this number too big” to “is this number well-allocated,” which is a question you can answer.
The script
When the CFO asks why AI is up forty percent, this is the paragraph. Edit lightly.
AI spend is up because the tools are now priced in proportion to use, and our use is up because the people producing the most leverage are using them most. Cost per active user is down twelve percent year over year even as total spend grew, because we shifted seats from idle deployments to the people who actually open the tools. Seventy percent of the spend is concentrated on roles where we can name the workflow that changed. We expect this line to grow at fifteen to twenty-five percent a year through 2028 as the agent platforms mature and we move more workflows under their oversight. We audit it quarterly and reallocate. The number is supposed to go up. The discipline is in the shape, not the cap.
That paragraph closes the meeting. It puts a number on the table, signals you understand the difference between cost and allocation, sets an expectation for future growth, and commits to the audit as the governance mechanism. It does not promise restraint, because restraint isn’t the value being delivered. Leverage is.
The next several years
Using AI well is a leadership skill, not a procurement decision. (Tim Ariyeh has a piece on the management framing that’s worth reading alongside this one.) The orgs that hold AI spend flat through 2028 will be the ones whose proposal teams are still six people, whose engineering throughput is still measured in last year’s units, and whose leadership team is still arguing about pilot programs while their competitors ship. The orgs that let the spend grow, and who can defend the shape of it line by line, will be the ones doing the work.
Defending a rising AI line is the leadership work this cycle. Stop trying to flatten the number. Start being able to point at every dollar.