What CFOs Are Actually Measuring When They Approve AI Spend in 2026

11 hours ago
3 min read

Roland Berger's ASEAN GBS data shows forty percent of organisations with an AI vision have reached implementation. That sounds like progress until you separate two things that get treated as one: an implemented model and a proven one are not the same milestone. This is the real CFO AI spend measurement problem in 2026: most Finance functions can tell you the tool is live. Far fewer can tell you, in a number the board will accept, what it actually changed.

That gap shows up the same way in almost every Finance function we look at. The tool gets built, it goes live, and the measurement framework that was supposed to prove it worked gets treated as an afterthought, something to sort out once the dashboards exist.

Knowing what a board wants to see is one problem. Having actually built the system that produces it, before the tool goes live rather than after, is a separate one, and it is the one most Finance functions get wrong.

The CFO's AI spend measurement problem in 2026

The instinct is to build the measurement framework after the tool is already running, treating it as a reporting exercise bolted onto a deployment that is already underway. That ordering is the forty percent problem in practice. A framework designed after the fact tends to measure what the tool makes easy to measure, hours saved, tickets closed, rather than what the board actually asked for, which is P&L impact, working capital movement, and risk exposure.

The fix starts earlier than most CFOs expect. Before any agentic AI build is approved, the finance function needs three things settled: which specific line on the operating P&L this is meant to move, who signs off on what counts as a valid before-and-after comparison, and what the review cadence looks like once the tool is live. The tool does not need to exist yet for any of that to happen, but it becomes far harder to retrofit once it does.

What the functions that get this right do differently

The functions we've seen do this well treat measurement design as part of the build, not a follow-up task. The KPI gets written into the project charter before a vendor is selected, and the finance team that owns the metric is in the room for the technical scoping, not handed a dashboard after go-live and asked to make sense of it.

The functions that struggle tend to share one habit. They can describe the AI tool in detail and its measurement plan in a sentence. That imbalance is visible from the outside within a single conversation, and it is usually the first thing that stalls board sign-off on the next phase of spend.

The test that actually tells you where you stand

None of this requires new technology or a bigger budget. It requires the measurement conversation to happen before the procurement conversation, not after it. A CFO who can name the specific metric, the owner, and the review date before a single line of the contract is signed has already closed the forty percent gap for that project, regardless of what the tool itself turns out to do.

If your Finance function has an AI deployment underway right now, the honest test is simple. Ask whoever owns it to name the one number the board will see in six months. If the answer takes longer than the description of the tool itself, the measurement framework was assumed into existence rather than actually built.

Implementation is no longer the hard part. Proof is.

AGOS Asia is an AI-first digital GBS transformation partner operating across ASEAN. We work with finance and shared services leaders to rethink, redesign, and rewire their operating models for the AI era. AGOS GBS Summit 2026, the ninth edition, takes place on 10 September at Sheraton Petaling Jaya