Curated research

What finance teams actually think about AI

Finance teams are not sceptical about whether AI works. They are sceptical about letting it act unsupervised — and that single distinction explains almost everything else in the data. Each figure below carries its publisher, its year and a link to the document it came from.

Last updated:

Sources cited: 13

The headline numbers

The eight figures worth carrying into a meeting. Each links to the section that gives it context, and to the document it came from.

Headline statistics on what finance teams think about AI, with source and year
FigureWhat it measuresSource
2.7%Finance staff who would let an AI agent make decisions including judgment callsDeloitte, 2025
59.7%Who would let one act, but only inside a defined framework, with judgment left to peopleDeloitte, 2025
71%Finance leaders who would reject a 99%-accurate AI tool that cannot explain its answersSage, 2026
59%Finance functions using AI in 2025 — against 58% in 2024, after 37% in 2023Gartner, 2025
91%Finance teams running AI who reported only low or moderate impact at firstGartner, 2025
36%Who name data quality as both their biggest barrier and their biggest opportunityKPMG, 2026
+5% / −6%Projected US employment change for accountants and auditors versus bookkeeping clerksUS Bureau of Labor Statistics, 2026
46%People worldwide willing to trust AI systems — a minority, and fallingKPMG & University of Melbourne, 2025

On this page

  1. 1.Why is AI the top finance priority but still not in production?
  2. 2.Would a finance team let AI post an entry without a human checking?
  3. 3.Will AI replace accountants, or just the data entry?
  4. 4.Which finance jobs do teams actually hand to AI first?
  5. 5.What actually blocks AI adoption in a finance team?
  6. 6.Why is AI held to a higher accuracy standard than people?

Why is AI the top finance priority but still not in production?

This is the widest gap in the data. Finance leaders expect AI to become standard equipment within a few years, and say so in large numbers. The share that has actually put it into daily use is far smaller — and in the most recent measurement, it barely moved at all.

AI use in the finance function, by year

Gartner's own series, asked the same way three years running. The jump happened in 2024. 2025 was flat.

2023
37%
2024
58%
2025
59%

What we take from this

Read the two adoption numbers together rather than picking the flattering one. Gartner asks a narrow question of a small, senior sample and gets 59%. KPMG asks a broader one of a thousand finance leaders and gets 75%. Both can be true, because "using AI" stretches from a chatbot somebody opened once to an extraction pipeline that feeds the ledger nightly. Anyone quoting a single adoption figure without saying which question it answers is not telling you much.

The number that is hard to argue with is the 91% reporting low or moderate impact at first. That is not a technology failure. It is what happens when a pilot runs on a sample of documents, in a sandbox, without the integration work that would let it touch the real process. The pilot succeeds and nothing changes, because nothing was ever wired in.

Our own read, from building these systems: the projects that stall are almost never stopped by model quality. They stop at the boundary — the ERP that will not accept a write, the approval step nobody owns, the exception queue with no home. That work is unglamorous, it is most of the job, and it is what separates a demo from a deployment.

Would a finance team let AI post an entry without a human checking?

Almost never — and the figure is starker than the general debate about AI at work would suggest. When Deloitte asked finance and accounting professionals how much decision-making authority they would hand an AI agent, the share willing to let it exercise judgment on its own did not reach three in a hundred.

How much decision authority finance staff will give an AI agent

Deloitte, 2025. The middle bar is the real finding: most people are not refusing AI, they are refusing unsupervised AI.

Only within a defined framework
59.7%
Not in any capacity
19.9%
Always, including judgment calls
2.7%

What we take from this

This is the section that explains every other section on the page. The bottleneck in finance AI is not capability, and has not been for a while. It is that the people who sign the accounts will not put their name to a system that acts without them.

Look at the shape of the answer, though, because it is more useful than the headline. Only one in five refuses AI outright. Six in ten will accept it inside a defined framework. That is not resistance — it is a specification. It tells you exactly what to build: something that proposes, evidences the proposal, and then waits.

Every build we ship works that way, because it is the only shape finance teams accept. AI drafts the collections email and a person releases it. AI extracts the invoice and a reconciliation gate checks the arithmetic before anything reaches the ERP. The autonomy people actually want handed over is autonomy over the reading and the keying, never over the decision.

Will AI replace accountants, or just the data entry?

The fear is real and widely reported. The labour statistics point somewhere more specific: the routine keying role is shrinking, the qualified accounting role is growing faster than the economy, and the official explanation for the shrinkage describes work being reshaped rather than removed.

Projected US employment change: two accounting occupations

Same profession, opposite directions. Bars show magnitude; the clerks figure is a decline. Source: US Bureau of Labor Statistics.

Accountants & auditors (2024–34)
+5%
All occupations, average
+3%
Bookkeeping & accounting clerks (2025–35)
−6%

What we take from this

The two BLS lines are the most useful pair of numbers on this page, because they come from the same agency, use the same method, and disagree with each other in exactly the way the headlines do not. Qualified accountants: growing faster than the economy. Clerks doing the keying: shrinking.

And the agency's own stated reason is not "replaced by software". It is that the same people move into analysis and advisory work once the routine entry stops. That is a description of a job changing, not ending — which matches what we see on site. Nobody has been let go because an extraction pipeline went live. The month-end just stopped eating the first week of it.

Where the fear is well-founded is in roles defined entirely by rekeying. If a job is transcription from a PDF into an ERP and nothing else, the honest answer is that the job is genuinely exposed. The realistic response is the one the BLS is already describing: move the person up the chain into exception handling, controls and analysis, which is work that needs somebody who understands the ledger.

Which finance jobs do teams actually hand to AI first?

There is a clear ordering, and it tracks the trust data almost exactly. The work that gets automated first is the work where a mistake is visible, cheap and caught immediately. The work that stays human is the work where a mistake is expensive and quiet.

Most-adopted AI use cases in finance functions

Gartner, 2025. Three use cases were adopted by more than a third of finance organisations already running AI.

Knowledge management
49%
Accounts payable automation
37%
Error & anomaly detection
34%

What we take from this

Accounts payable sits second for a reason that has nothing to do with how interesting it is. An invoice has a right answer printed on the face of it. If the extraction is wrong, the arithmetic fails, the purchase-order match fails, and the exception surfaces before anything posts. It is the safest possible place to put a model, because the model's work can be checked by machine rather than by trust.

Anomaly detection ranks third for the same reason inverted: the AI never decides anything, it only raises a hand. A false positive costs a minute of somebody's attention. That is a risk profile a controller can sign off without a governance committee.

What sits far down the list — anything requiring judgment about a customer relationship, a disputed charge, a write-off — is exactly what the trust data predicts. Teams are not ranking use cases by how much value is available. They are ranking them by how quickly a mistake becomes visible. Start where the feedback loop is tightest and you get the adoption; start where the prize is biggest and you get a stalled pilot.

What actually blocks AI adoption in a finance team?

Not the model. Across every survey we could verify, the blockers cluster in the same three places: the state of the data, the systems the data lives in, and whether anyone could prove afterwards what the AI did.

Assurance-readiness changes the outcome

KPMG, 2026. Organisations able to evidence what their AI did report materially better results than those that cannot.

Error reduction — assurance-ready
33%
Error reduction — not assurance-ready
6%
Confident scaling AI — assurance-ready
42%
Confident scaling AI — not assurance-ready
14%

What we take from this

The assurance finding deserves more attention than it has had. It is not saying that careful organisations are pleasanter to deal with. It is saying that the ability to evidence what the AI did is what separates deployments that actually reduce errors from ones that do not — a five-fold difference on KPMG's own numbers.

That fits the pilot problem in the first section exactly. A pilot proves a model can read an invoice. Production requires that you can show an auditor, eleven months later, which version of which document produced which posting, and what the system was told at the time. Almost none of the effort in a pilot goes there. Almost all of the effort in a real deployment does.

So the readiness question for a finance team deciding where to start is not "is the AI good enough yet". It is "is our data reachable, and could we prove what happened". Both are answerable in an afternoon, before committing budget to anything — which is the entire point of the assessment linked below.

Why is AI held to a higher accuracy standard than people?

Because it is, measurably, and the research explaining why predates the current wave by a decade. People forgive a human error and generalise from a machine one. In finance this surfaces as a demand not merely for accuracy, but for accuracy that can show its working.

Acceptable error rate: AI against a human doing the same task

Lenskjold et al., BJR Open, 2023 (n = 44). Respondents allowed a human roughly two-thirds more error than they allowed the machine.

Acceptable from a human
11.3%
Acceptable from AI
6.8%

What we take from this

The Sage finding is the one to keep. A finance leader will turn down a tool that is right ninety-nine times in a hundred if it cannot show its working. That means explainability is not a feature competing with accuracy for space on the roadmap. It is a gate standing in front of accuracy: a system that cannot be interrogated never gets to compete on being right.

There is a sting in the KPMG data worth sitting with, though. The same population that demands machine-grade accuracy also reports that two-thirds of employees use AI output without checking it, and that more than half have made mistakes as a result. The stated standard and the actual behaviour are not the same standard. People want the machine to be perfect and they are not, personally, verifying that it is.

That gap is the argument for building the check into the system rather than into the policy. Asking people to reliably verify AI output does not work — the numbers above are what that looks like in practice. Making the arithmetic reconcile before anything can post does work, because it does not depend on anyone remembering to care at half past five on a Friday.

How this page was put together

Every figure was read out of the publisher's own document — the report, the press release, the paper, the statistical handbook — and not out of an article quoting it. Each line carries the publisher, the edition and a link, so you can check any of it in a click.

Statistics are tagged by where they come from. Official statistics, academic work and independent research are stronger evidence than a vendor-commissioned survey. Vendor research is included where the sample is large and the method is disclosed, but it is labelled so you can weigh it accordingly — one figure on this page carries that tag.

Where two credible surveys disagree — as Gartner and KPMG do on how many finance functions actually use AI — both are printed with the disagreement explained, rather than picking the one that suits the argument. And where a widely-repeated statistic could not be traced to a document we could open, it was left out: that cost this page the much-quoted S&P Global figures on abandoned AI projects, which sit behind a subscription with only trade-press quotes retrievable.

Two sources are older than the five-year bar this page normally applies. Both are kept deliberately: the 2015 algorithm-aversion paper, still the reference for why a machine error is punished harder than a human one, and KPMG's 2023 measurement of whether financial reporting leaders expect AI to take their jobs, which nothing since has asked as directly.

Reviewed annually. Gartner's AI in Finance survey lands each November, KPMG's Global AI in Finance each May and the BLS projections each year, so this page is re-checked against the newest editions every August.

What "inside a defined framework" looks like in practice

The trust data describes a specification rather than a refusal: AI does the reading, the matching and the drafting; a person keeps the decision. These are builds running that way in production.

1,500 Supplier Invoices a Month, Matched Automatically

AP extraction and PO matching behind a reconciliation gate — the accuracy check runs before anything reaches the ERP.

1,080 Overdue Invoices, Chased Automatically by AI

Every collections email drafted by AI and released by a person: the defined framework 59.7% of finance staff say they want.

$11.8M in Receivables, Collected Systematically

Daily risk scoring that flags and escalates, but never decides — anomaly detection as finance teams actually accept it.

Related research

This page covers what finance teams think. Its sibling covers what the manual work they are trying to escape actually costs.

Manual data entry in finance: the statistics

Everything citable about what manual finance work costs — hours, euros, errors, close days and people. Each figure below carries its publisher, its year and a link to the document it came from, so you can check it rather than take our word for it.

Questions people ask

What percentage of finance teams use AI?

59% of finance leaders reported using AI in the finance function in 2025, against 58% in 2024 and 37% in 2023 (Gartner, 2025 AI in Finance Survey, 183 CFOs and senior finance leaders). A broader KPMG survey of 1,013 finance leaders published in May 2026 put active AI use across the finance function at 75%, up from 30% in 2024. The two differ because they ask different questions of different samples — quote whichever you use with the question attached.

Would finance teams let AI post a journal entry or pay an invoice without human review?

Very few would. In a Deloitte poll of more than 3,300 finance and accounting professionals taken in January 2025, just 2.7% said they trust AI agents to always make decisions including judgment calls. 59.7% trust AI agents to make decisions only within a defined framework, with judgment calls left to people, and 19.9% do not trust the technology to make decisions in any capacity. Trust was also the single leading barrier to adoption, named by 21.3%.

Will AI replace accountants?

The labour data points to reshaping rather than replacement, and it splits sharply by role. The US Bureau of Labor Statistics projects employment of accountants and auditors to grow 5% from 2024 to 2034, faster than the 3% average for all occupations, while bookkeeping, accounting and auditing clerks are projected to decline 6% from 2025 to 2035. The BLS attributes that decline to automation of routine tasks moving clerks toward analytical and advisory work rather than eliminating the role outright.

Which finance processes do teams automate with AI first?

Three use cases were adopted by more than a third of finance organisations already running AI: knowledge management (49%), accounts payable process automation (37%) and error and anomaly detection (34%). Source: Gartner, 2025. The ordering tracks how quickly a mistake becomes visible — an invoice has a right answer printed on it and an anomaly alert only ever raises a hand, so both can be checked mechanically.

What is the biggest barrier to AI adoption in finance?

Data and trust, not the technology. 36% of finance leaders cite data quality as both their biggest barrier and their biggest opportunity (KPMG, 2026). Gartner names data literacy and technical skills alongside inadequate data quality and availability as the largest obstacles. In Deloitte's 2025 poll, trust in agentic AI was the leading barrier at 21.3%, ahead of integration with existing systems at 20.1% and a shortage of skilled people at 13.5%.

Is AI held to a higher accuracy standard than humans in finance?

Yes, measurably. 71% of finance leaders would reject an AI tool that was 99% accurate if it could not explain its answers (Sage, 2026, survey of more than 2,000 senior finance decision-makers). A 2023 study in BJR Open found respondents set an acceptable error rate of 6.8% for AI against 11.3% for a human doing the same task. The underlying effect, algorithm aversion, was documented by Dietvorst, Simmons and Massey in 2015: people abandon an algorithm after seeing it err, even when the human alternative performs worse.

The bottleneck is trust, not technology

If the numbers above describe your team — convinced by the idea, unwilling to let anything post unsupervised, unsure whether your systems could even evidence what happened — that is the normal starting position, and a solvable one. The assessment is free and tells you where your data and process actually stand.

Free AI Readiness AssessmentHow we control what AI can do