The CFO Now Owns the AI Budget: What Survives 2027

For two years, AI budgets were waved through. A team asked for a few licences and a pilot, someone senior said "we can't be left behind," and the money appeared. 2027 planning looks nothing like that. Finance has read the returns, and the returns are thin: McKinsey's 2026 survey found the share of companies reporting any EBIT impact from AI stuck at roughly 37% for a second year running, and a KPMG survey put the share of leaders claiming established AI ROI at about 7%. Those numbers have a cost now. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and the quiet work of deciding which ones has already started inside most finance teams.

So the question for anyone who owns an AI initiative is no longer "what can we build." It is "can I defend this line when the CFO opens the spreadsheet in Q1." This is a plain look at what that review is testing, what tends to get cut first, what survives, and how to move your project into the second group before the cut happens.

The budget mood has flipped

The shift is about who signs. Through 2024 and 2025, AI buying lived with engineering and whichever department was most excited. In 2027 it moves toward finance, and finance buys differently. It wants a named owner, a baseline, and a number, not a demo and a vibe.

You can see the mechanics changing in public. Companies are putting hard per-user caps on agentic tools, Uber's reported cap of 1,500 dollars per employee per month on coding agents being the example everyone cites. Forrester has estimated that enterprises would defer around a quarter of planned AI spend rather than approve it on faith. None of this means AI is being abandoned. It means the free pass expired.

What finance is actually asking now

The 2027 review is not hostile, but it is specific. Expect some version of these questions on every AI line:

• What business number did this move, and by how much? Hours saved, cost per case, cycle time, error rate. A screenshot of happy users is not an answer.
• What is the fully loaded cost? Not the licence. The licence plus the model usage plus the integration work plus the person maintaining it. Orchestrated agent workflows can cost many times a simple 2023-era query, and that recurring usage bill is where budgets quietly blow out.
• Who owns the outcome? A line item without an accountable name is the easiest thing in the world to cut.
• What happens if we stop funding it? If nothing breaks, you have just argued for your own cancellation.

What gets cut first

Across the projects that do not survive 2027 planning, the pattern is consistent, and it has little to do with how clever the technology is.

1. Horizontal "AI platforms" with no first use case. The in-house effort to build general-purpose tooling that something, someday, will use. These burned the most and shipped the least. We wrote about this failure mode in detail in why 95% of enterprise AI projects fail.
2. Seat sprawl nobody uses. Hundreds of licences bought in a wave of enthusiasm, with single-digit weekly active use. Finance can pull a usage report, and it will.
3. Duplicate tools. Three teams paying for three overlapping assistants. Consolidation is the first lever a CFO reaches for, and it is painless to justify.
4. Pilots that demo well and never shipped. A year in, still a pilot. The clearest signal that the value was never real enough to finish.

What survives the cut

The projects that keep their funding share a shape. They are narrow, owned, wired into a real process, and attached to a number someone agreed to in advance.

• One process, one owner, one metric. Invoice matching that cut processing time by a measured amount. First-line support triage with a resolution rate you can read off a dashboard. Scope you can point at.
• Connected to real systems, not a generic chatbot. The work that survives reads your order data, your tickets, your rules, and does something a human used to do by hand. That integration is also what makes it hard to rip out.
• A running cost that is smaller than the saving. Not break-even someday. A monthly bill that is visibly below the monthly value, with both tracked.
• Evidence, not anecdote. A before and after, in the units the business already uses. If you have that, the review is short.

How to make your project fundable before Q1

If you own an AI initiative and the review is coming, there is still time to move it into the survivor group. The work is unglamorous and almost entirely non-technical.

• Pick the single process it most clearly improves and drop the rest of the scope, at least for the funding case. A narrow win that is measured beats a broad vision that is not.
• Reconstruct the baseline now, even retroactively. What did this task cost in hours or errors before, and what does it cost today? Get it into the same numbers the finance team uses.
• Add up the true run cost, model usage included, and compare it honestly to the saving. If it is underwater, say so and fix the scope before someone else finds out. Our guides on measuring AI ROI and the total cost of owning an AI agent walk through the exact lines to count.
• Name an owner who will stand behind the number in the room.
• Decide build, buy, or partner on evidence, not habit. A generic tool is fine where your process is standard; the moment your advantage depends on your own data and rules, an off-the-shelf seat hits a wall. The honest version of that trade-off is in build vs buy for custom software.

The teams that sail through 2027 planning are not the ones that spent the most or picked the best model. They are the ones that did this scoping work early, so that when finance opened the spreadsheet, the line defended itself.

If you have an AI project that keeps demoing well and stalling before it can prove a number, tell us which process you are trying to fix. Getting past the CFO almost always starts with narrowing the scope to one measurable win, not asking for a bigger budget.