The CFO's AI Budget: What to Fund, Pilot, or Ignore

The ask lands on your desk in one of two shapes. Either it's a single line item with an ambitious number and a return you can't model — a platform, a "transformation," a vision on a slide. Or it's a dozen small requests scattered across departments, each with a champion, none with a baseline.

Both put you in the same corner. Say yes and you're underwriting something you can't defend to the board. Say no and you're the person who was against the future while a competitor moved. That's a false choice, and you don't have to accept it.

The question you're actually being asked

"Should we do AI" is not a CFO's question. It's a conference-panel question. Yours is narrower and far more useful: given finite capital and finite attention, what do I fund, what do I pilot cheaply, and what do I ignore this year?

That's not a technology problem. It's a capital-allocation problem, and you already run a discipline for allocating capital under uncertainty. You don't fund every request at the number attached to it. You sort. A CFO's AI strategy, stripped of the noise, is exactly that sorting discipline — nothing more exotic than what you already do with every other line that crosses your desk.

So stop grading proposals on how impressive the technology sounds. Grade them on evidence, ownership, and the cost of finding out. For any AI ask, three questions in order:

1. Is the win proven and expressible as reclaimed capacity you can redeploy? If yes, fund it.
2. If it isn't proven, is it uncertain but cheap to test? If yes, pilot it — with a kill criterion and a time box.
3. Neither — no owner, no operational tie? Ignore it, and drop it on a watchlist.

Fund: proven internal leverage

Fund the wins that show up as reclaimed capacity on work you already pay for — not new revenue you're hoping for, but hours returned on a process you can name.

The pattern is boring and reliable: first-draft work, summarizing, reconciliations, the analysis that used to take an analyst a full day and now takes an hour to review instead of write. These aren't moonshots. They're the operational middle of the organization, and the return is legible. You can point at the process, the hours, the loaded cost, and where the freed time goes.

Picture a controller's team that loses three days every month assembling the board pack — pulling the numbers, formatting, drafting the commentary. Move the first draft to AI and a reviewer and you don't save a rounding error; you hand a senior person most of a week back, every month, on work that was never the best use of them. That is a fundable case, because you can name all four parts before you spend a dollar.

A fundable case has four parts: the process, the hours it consumes today, the cost of those hours, and what the reclaimed capacity gets redeployed to. If a proposal can't fill in those four blanks, it isn't ready to fund — it's ready to pilot. Fund where the return is capacity you can see and reassign, not capacity you've been promised.

"Proven" has a low bar, but it is a bar. It means you've watched it work in a small test, or it's a pattern so well-worn that the risk is execution, not possibility. "Promised" means a vendor demo and your own optimism. Fund the first. Pilot the second.

Pilot: a cheap bet with a kill switch

Most of what's worth doing is uncertain, and uncertainty isn't a reason to say no. It's a reason to buy information before you buy commitment. That is all a pilot is: a small, time-boxed purchase of evidence.

Two conditions make a pilot real, and without both you don't have a pilot — you have a slow yes. First, a kill criterion agreed before you start: the specific result that means you stop. Second, a time box and a spend cap — this many weeks, this much money, then a decision either way.

A pilot without a kill criterion isn't a pilot. It's a subscription you forgot to cancel.

Make the criterion specific enough to be uncomfortable. "The support team drafts replies with it for three weeks; if median handle time doesn't fall and quality holds, we stop and don't renew." That one sentence is worth more than a business case, because it commits you to a decision instead of a hope.

Define what "it worked" looks like in a number you'll actually believe, and what "it didn't" looks like, on the same page, before anyone logs in. The organizations that waste the most money on AI aren't the ones that bet wrong. They're the ones running "pilots" that never end and never decide — perpetual experiments no one has the standing to kill.

Ignore: hype without an owner

Some of it you should simply set down. Not fight, not route to a steering committee — set down, on purpose, and move on.

The test is two questions. Is there a named person in this building who will own it? Does it attach to a process you actually run? If the answer to either is no, it's not a decision you need to make this year. Ignoring isn't a verdict on the technology. It's a verdict on the timing and the ownership: this doesn't have a home yet.

Keep a short watchlist so "ignore" doesn't decay into "forget." Note the thing, note what would have to change for it to earn an owner, and revisit quarterly. Most of what's loudest right now will still be looking for a home next quarter. Some of it will have found one — and then it graduates to a pilot. Active deferral is a position you can defend to a board. "We got swept up" is not.

Funding it without a moonshot budget

None of this requires a new budget line with a frightening number on it. Both of the classic failure modes come from believing it does.

The first is over-funding a vision: one big check written on the strength of a deck, no evidence underneath it, and by the time it underdelivers you're too far in to stop. The second is the mirror image — starving every experiment, funding nothing, learning nothing, while the cost of that caution shows up quietly as the competitor who ran the small bets you wouldn't.

The triage protects you from both. Fund concentrates real money where there's evidence. Pilot spreads small, capped bets where there isn't. Ignore keeps attention off the noise. Two moves cover almost all of it:

• Redeploy reclaimed hours. The first fundable wins return capacity. That capacity — not fresh cash — is what pays for the next round of pilots. The organization funds its own learning out of the time it just got back.
• Buy seats before platforms. Rent before you build. A handful of licenses and a two-week pilot will teach you what a six-figure platform contract only assumes. Earn the platform with pilot results; don't lead with it.

And set expectations honestly on time. Meaningful payback on serious AI investment tends to run in years, not quarters — closer to how you'd model a capability build than a software subscription. Budget for that reality and you'll fund the few things that compound instead of the many that demo well and fade.

Run the filter this week

You don't need a strategy offsite. Take the AI asks currently on your desk and sort them into three columns: fund, pilot, ignore. Force every item into one. The ones with no owner and no operational tie sort themselves into the third column faster than you'd expect.

For the fund column, write the four blanks — process, hours, cost, redeployment. For the pilot column, write the kill criterion and the time box. Do that and you've replaced a vague, anxious "AI strategy" with a page a board can read in two minutes.

If you'd rather run that triage against your own numbers than a generic template, The Operational Intelligence Method interviews you about your operation — no system access, no data exports — and produces two of its eight documents that this article is really about: an ROI Snapshot that states reclaimed capacity as conservative ranges, and a Governance Pack that gives every bet an owner and a boundary.

The point was never to spend more on AI, or less. It's to spend deliberately — architected, not automated. Fund the proven, pilot the uncertain, ignore the ownerless, and let the evidence tell you when a pilot has earned its promotion.