Most 2027 AI budgets being written this quarter will be built the way 2026's were: last year's number, plus a percentage, plus whatever the vendors pitched well. That approach is about to meet a much harder audience.
The money is still growing — AI spending reaches roughly $297 billion by 2027, inside global IT spend of $6.37 trillion. But the scrutiny grew faster. 98% of CIOs report that board pressure to demonstrate measurable AI ROI has risen since 2024, and 85% now have their own compensation tied to AI outcomes. As we covered when the mid-2026 reckoning arrived quietly, the cull was of projects rather than people — but it was real, and it isn't finished.
So here is what separates a 2027 AI budget that survives the CFO from one that gets halved in February.
1. Budget in unit economics, not totals
"£400,000 on AI" is an unanswerable number. It invites one question — can we spend less? — and has no defence.
"£1.80 per invoice processed, down from £4.20" is a completely different conversation. It's a business case, it scales predictably, and it tells you when to stop.
For every initiative, agree one primary business metric and capture its baseline before you build: cost per case, hours removed, cycle time, exception rate. This is the single highest-leverage habit in AI budgeting, and the one most often skipped, because doing it properly means admitting you don't currently know the number. (Our cost calculator is a quick, conservative way to put a defensible figure on a manual process before you commit a penny to automating it.)
Then track the fully-loaded cost against that metric — inference plus retrieval infrastructure, evaluation, human review and rework. The model API is frequently the small line item.
2. Put a kill list in the budget
The strongest FinOps move available is stopping the project that never moved its number. Nearly half of organisations have already delayed, stopped or scaled back AI projects on budget grounds.
That's going to happen to some of your initiatives regardless. The choice is whether you decide which, on evidence, in October — or whether finance decides for you, on instinct, in March. Name the review point and the threshold now: if this hasn't moved its metric by Q2, it stops and the budget moves to the one that did. Boards fund teams that do this. They defund teams that can't tell them which project is failing.
3. Add a capacity line — this is new
Most AI plans quietly assume compute will be available, cheap and instant. Gartner expects 40% of existing AI data centres to be operationally constrained by power by 2027, with AI data centres drawing roughly 500 TWh/year, and 83% of operators expecting workloads to exceed their capacity within two years.
That makes capacity a planning assumption you should state explicitly rather than inherit. It's also the strongest commercial argument yet for right-sizing your models: using the smallest model that does the job has stopped being a cost optimisation and become a capacity strategy.
4. Fund the tech debt, honestly
Legacy systems are the barrier CIOs cite most often as blocking AI adoption. You cannot automate a process you cannot integrate with, and no amount of model quality fixes an estate that can't expose its data.
The trick is framing. Presented as "maintenance", modernisation loses every budget round. Presented as tech equity — the thing that unlocks the AI roadmap and shows up on the balance sheet — it wins them. It's the same work; one framing is fundable.
The bottom line
A 2027 AI budget that holds up has four things most don't: unit economics with a baseline, an explicit kill list, a stated capacity assumption, and funded tech debt. Build it that way and the board conversation shifts from "can we spend less on AI?" to "where should we spend more?" — which is the only version of that conversation worth having. Turning AI investment into a defensible, board-ready number is exactly what our Executive Advisory work does.
Want a clear-eyed view of what your AI is actually returning before you set next year's number? Let's talk.