- Home
- The Thinking Wire
- AI Spend Came In at 2.5x the Plan. Give It Its Own Line in the Board Plan.
AI Spend Came In at 2.5x the Plan. Give It Its Own Line in the Board Plan.
Enterprises in Menlo Ventures’ 2026 survey planned, on average, to spend 1.9 times last year’s AI level. They reported spending 4.7 times. That is roughly 2.5x the plan, according to Menlo Ventures’ 2026 survey of nearly 400 US enterprise decision-makers. The median is calmer, 2.2x actual against a 1.8x expectation, and that median actual still sits about a fifth above the median expectation.
A forecast missed by a fifth is a bad year. A forecast missed by 2.5x on average suggests the forecast was never tied to whatever drives the spend. In any other cost line, a miss that size would prompt a review of how the plan was built.
Our reading is that the overrun is a planning-control failure before it is a cost problem. The fix is structural and fairly boring: give AI its own line in the board plan, attach a unit and a benefit date to it, and write down in advance what will trigger a re-plan.
What the survey measured, and what it did not
The Menlo figures come from a survey fielded by Morning Consult between September 8 and 26, 2026. Respondents reported their own spending and their own expectations, so the 4.7x and the 1.9x are self-reported numbers from the same people. That makes the ratio between them more meaningful than either number on its own. Nobody had to guess what someone else planned.
Menlo also publishes market estimates: enterprise AI deployment spend reached $177B by September 2026, up from $33B a year earlier, and frontier labs now hold 69% of those deployment dollars, up from 50%. Two caveats apply. The dollar figures are Menlo’s own estimates, and the authors disclose portfolio stakes in companies that sell into this market, including Anthropic and Fireworks. Read the market size as a directional signal. The planned-versus-actual ratio is the sturdier finding, because it does not depend on Menlo’s model of the market.
The direction is clear enough. Spend grew roughly fivefold in a year by Menlo’s estimate, and a growing share of it flows to frontier labs (69% by Menlo’s estimate) whose prices the buyer does not set.
Why the plan broke
Annual plans assume the drivers of a cost line are known and move slowly. Headcount and rent behave that way. AI spend, as I read it, has three properties that break the assumption.
Usage is decided by individuals, often engineers, at the moment of work. Nobody approves each request. Prices and models change inside the planning year. And the benefit, when it comes, shows up in other lines: fewer hires, faster delivery, lower support cost. The spend sits in one place and the return sits somewhere else, so the line looks like pure cost growth until someone connects the two.
When a cost line has no unit, no owner and no stated benefit, the only number the plan can carry is a growth multiple. The survey shows how well growth multiples held up.
The board-plan line
OnlyCFO’s annual-plan guide for the AI era is direct about where this belongs: “AI spend is (or will likely soon be) your biggest expense, so of course you need to break it down to the board.” The guide splits that spend into two categories, COGS and opex. The distinction matters. As I read the split, AI inside the product you sell scales with customers and belongs in gross margin. AI that your own staff use to work scales with headcount and belongs in operating expense. Mixing them hides both the margin story and the productivity story.
The second piece is the comparison the guide asks the plan to show: “How are we spending on AI relative to headcount? What measurable benefit are we getting (lower costs, avoided hiring, greater output, etc) and when does it show up in the plan?”
That question has two halves, and the second half carries most of the weight. A benefit with no date is a hope. A benefit with a date can be checked against actuals in the quarter it was promised. If the plan says AI tooling will avoid four engineering hires in the second half, the hiring plan for the second half should already show four fewer hires. When it does not, the board can see that the benefit is unclaimed.
Put together, a usable AI line in the board plan has four fields:
- Category. COGS or opex, never a blend.
- Unit. What one unit of spend buys: per customer, per engineer, per completed task. Earlier pieces here covered unit cost per completed task and the billing unit as a control surface.
- Benefit and its date. Which other line moves, by how much, and in which quarter.
- Re-planning trigger. The variance at which the plan is reopened rather than defended.
A budget limit as a planning control
Revolut offers one concrete account of what happens when the unit is missing. The Register quotes David Abitbol, Revolut’s operations lead: “We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits.”
These figures come from one speaker at a conference, reported by The Register and not independently verified. The pattern is still useful. A tenfold spread in monthly spend per engineer, with little return to show for it, is exactly the signal a per-unit view surfaces and a total-spend view hides. Revolut launched with strict budget limits and moved to self-hosted models. By the same report, self-hosting GLM cost “about a tenth” of Codex or Claude Code, and AI coding inference costs fell “by roughly 90 percent.” The report also notes that GLM has not replaced the frontier models entirely.
The line I find most relevant for a CFO is the third one: “The billing is flat, we don’t have any bill shock at the end of the month, and the frontier is the exception.” A flat bill is plannable. Frontier usage becomes a choice someone makes and can be asked about, which is what a plan needs. The economics of moving off flat-fee pricing and of budgets that treat tokens as unlimited are covered elsewhere. The point here is narrower: by Abitbol’s account, the move to capped and then self-hosted usage turned an open-ended line into a flat one, and that is what made it plannable.
Write the re-planning trigger before you need it
OnlyCFO is also candid about the limits of the exercise: “AI is moving too fast to strictly hold to an annual plan for 12 months. Every company must be willing to tear it up completely if the underlying assumptions break.”
Willingness needs a mechanism behind it. A team that tears up the plan only when someone feels it has broken will do so late, after the overrun is already in the actuals. The trigger needs to be written into the plan when it is approved, with numbers. Three forms are practical:
- Variance trigger. Actual AI spend exceeds plan by a stated percentage for a stated period.
- Unit trigger. Cost per unit (per engineer, per customer, per task) moves outside a stated band, even if the total is on plan.
- Benefit trigger. The promised benefit has not appeared in its target line by the quarter the plan named.
The second and third matter as much as the first. A total can stay on plan while unit cost doubles and usage falls, and a total can run over while the benefit arrives early. Each trigger asks a different question, and a board that only watches the total will miss two of them.
When a trigger fires, the response is a re-plan with the owner of the line in the room: revise the assumption, change the cap or change the vendor mix. The outcome to avoid is the one the survey points to: actuals averaging 2.5x the plan by year end.
Do this now
Before the next budget cycle closes, take last year’s AI spend and split it into COGS and opex. For each half, name the unit and the owner. Then write one sentence per half that says which other line in the plan the spend is supposed to move, by how much, and by which quarter.
Add the three triggers to the plan document with numbers attached, and put the AI line on the board pack as its own row. If the split, the unit or the benefit date cannot be written down, that is the finding to bring to the board. A line that cannot be described cannot be planned, and the survey shows what happens when AI spend outruns its plan.
This analysis synthesizes 2026: The State of Enterprise AI (Menlo Ventures, November 2026), Presenting an Annual Plan (AI-Era Edition) (OnlyCFO, November 2026), and British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (The Register, November 2026).
Victorino Group helps finance and engineering leaders turn AI spend into a planned, owned line with units, benefit dates and re-planning triggers. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation