Most Finance Teams Have Adopted AI. Few Can Show a Return.

Two things are true about the same market this year: most finance teams say they use AI, and few can prove it paid off.

Adoption. Gartner's survey of CFOs and senior finance leaders puts AI use in the finance function at 59% (n=183, fielded May–June 2025). KPMG found over 75% of organisations using AI in financial planning, reporting and commercial analysis (n=1,013, revenue ≥US$250M, fielded March 2026). Some vendor surveys go as high as 97%.

Return. Gartner: 7% of CFOs report a strong impact from AI investment. RGP: 14% of 200 US CFOs report a clear, measurable return. CFO.com: 28% see measurable financial impact.

Those three ROI figures disagree with each other by a factor of four. They still all land in the same place, and it is a long way from 59%.

We sell AI agents to finance and accounting teams. It would be more convenient for us if this gap didn't exist. But it's the single most consistent pattern in the 2026 data, and we'd rather address it than pretend it isn't there.

The gap isn't a measurement problem

The easy explanation is that AI value is real but hard to measure: soft gains in speed and quality that don't show up in a P&L.

There's something to that. But the production data suggests a simpler explanation: most finance teams have deployed AI against work that wasn't expensive in the first place.

Gartner's November 2025 read on AI use cases running in production:

The top use case is knowledge management: search, summarisation, answering questions about documents. It is useful. It is also almost impossible to tie to a number, because it doesn't remove a process. It makes the existing one a little faster for everyone.

Meanwhile the clerical volume (payables, receivables, reconciliation, close preparation) is where the hours and the payroll sit. AP automation is the highest-penetration operational use case at 37%, and it is the one most likely to produce a defensible number.

Adoption has already flattened

This is not an early market still climbing.

Gartner's figure was 37% in 2023, 58% in 2024, and 59% in 2025. Adoption roughly doubled in one year and then stopped.

A plateau at 59% with 7% reporting strong impact describes a market where most teams have tried AI, found it interesting, and not yet found the thing that changes their numbers. That is a different situation from a market that hasn't started. It means the next wave of buying will be more sceptical, not less.

KPMG sharpens the picture: 71% of organisations say AI meets or exceeds ROI expectations, but only 23% say it exceeds them. Most deployments are landing at "about what we expected," which in practice often means "we didn't set a target."

Three things separate the teams that can show a return

They deployed against payroll, not against convenience. The pool AI is competing for in accounting is labour, not software. US accounting labour (accounting services plus payroll and bookkeeping) is roughly $239B in 2026, against a global accounting software market of about $23B. A tool that makes existing staff 10% faster produces a number no one can find. A tool that completes work those staff used to do produces one anyone can.

They set the baseline before they started. Close duration in days. Invoices processed per FTE per month. Hours per client per month. Days sales outstanding. If the baseline was never recorded, the improvement is unprovable even when it's real. That's one mechanical reason measured-return figures look so low.

They didn't confuse a pilot with a deployment. KPMG found agentic-AI deployers outperforming non-deployers by roughly 32 percentage points on average. But a pilot that runs alongside the existing process, with the existing headcount, doing the same work twice, cannot produce a return by construction. It can only produce a demo.

The autonomy ceiling

There is a ceiling on near-term returns the market doesn't discuss much. Grant Thornton's 2026 survey (n=950, fielded February–March 2026) found only 5% of organisations allow agents to execute high-stakes decisions without human review, with 60% capping agents at moderate-risk automation. In accounting, only 19% of professionals trust AI output enough to use it with limited review, and 52% disagree that they would (Financial Cents, n=486, fielded July–August 2026).

We could not identify a single vendor shipping fully autonomous journal-entry posting to a general ledger. Every product we examined, across incumbents and AI-natives alike, gates posting behind a human.

That is the right design. It also means the realistic 2026 return comes from compressing preparation, not from eliminating review. The gain is that a person starts at review instead of starting at a blank workpaper. Any vendor promising more than that in a regulated close is describing a product that doesn't exist yet.

Before you buy anything, ask this

Skip "does this use AI." Everything uses AI now.

"Which specific hours does this remove, and how will I know?"

If the answer involves faster search, better summaries, or improved visibility, it may still be worth buying. But it will land in the 71% that met expectations, not the 23% that exceeded them, and you won't be able to defend it in a budget review.

If the answer is a count (invoices coded, reconciliations prepared, days off the close), you have something measurable. Record the baseline first.

---

We'd rather have this conversation before a pilot than during one. If you want to scope what's measurable in your close, we'll help you set the baseline whether or not you buy anything.