Every vendor pitch about AI in finance eventually arrives at the same reassurance – “the model is highly accurate.” And every time, I watch the CFO in the room nod politely and then quietly decide not to let that model anywhere near a payment run without a human checking its work first. That gap, between what the technology can do and what finance is actually willing to let it do unsupervised, is the real story, and almost nobody in the AI finance conversation is naming it correctly.

Here’s the contrarian claim. Finance’s hesitation around AI finance tools isn’t a trust problem with the AI. It’s a trust problem with how AI has been designed and deployed in finance, as a black box asked to make judgment calls in a function where every judgment call has to be explainable to an auditor, a regulator, or a board member six months after the fact. That’s not resistance to change. That’s a rational response to a product category that, in most cases, still hasn’t solved for explainability.

Adoption Is Slowing, Not Accelerating, and That’s the Tell

If AI in finance were purely a change-management problem, you’d expect adoption to climb steadily as familiarity grows. It isn’t doing that. Gartner’s 2025 AI in Finance Survey of 183 CFOs and senior finance leaders found that 59% of finance functions now use AI in some capacity, up only slightly from 58% in 2024, following a sharp jump from 37% in 2023. Momentum didn’t build on itself. It plateaued right after the initial wave of enthusiasm met the reality of production deployment.

Even more telling. Gartner’s 2026 Finance Symposium data found that 84% of finance organizations have implemented or are actively planning generative AI for finance initiatives, yet only 7% report high or very high impact from those efforts. That’s not a technology gap. Widespread implementation with minimal realized impact is what it looks like when a tool gets deployed but never actually trusted enough to be given real decision-making authority – kept running in the background, generating outputs a human reviews and mostly overrides or ignores.

The Real Barrier Isn’t Capability. It’s Explainability.

Ask finance leaders directly why they hesitate, and the answer isn’t “the AI isn’t accurate enough.” Kyriba’s CFO research found that 78% of U.S. financial leaders identify security and privacy concerns as a critical challenge to AI adoption, and when asked to name the single biggest barrier to adopting AI solutions for finance, 31% pointed to security and ethics concerns as their number one issue, ahead of resource allocation (21%) and lack of internal knowledge (17%).

Notice what’s not topping that list is the raw model performance. Finance leaders aren’t primarily worried that the AI will get the math wrong. They’re worried they won’t be able to explain, defend, or reconstruct why the AI reached a given conclusion when someone with authority, such as an auditor, a regulator, or a board member, asks them to. That’s a fundamentally different problem than “the technology isn’t good enough,” and it demands a fundamentally different fix.

This is where most AI finance tools on the market miss the actual design brief. A tool that flags an anomaly without showing its reasoning is asking a controller to sign off on a conclusion they can’t independently verify. In a function built entirely around auditability, where every number needs a defensible trail back to a source document, a policy, and an approver, an unexplainable “trust me” from a model is not a minor UX gap. It’s disqualifying for anything beyond a low-stakes suggestion.

Why This Isn’t a People Problem

It would be easy and convenient for vendors to frame finance’s caution as generational resistance or a skills gap. The data doesn’t support that framing. Gartner found that 67% of finance leaders already using AI report growing more optimistic about it, and optimism scales with maturity. The deeper an organization goes with AI, the more confidence it builds. That’s not the profile of people who distrust the technology on principle. That’s the profile of people who trust it more once they can see, hands-on, exactly how it behaves and where its limits are.

The people problem framing also ignores what finance leaders are actually asking for. It isn’t “make the AI feel more familiar.” It’s structural – governance, explainability, and data readiness. Gartner’s own guidance to CFOs now explicitly centers this, noting that capturing real value from AI impact on finance initiatives requires aligning investment “to business outcomes, supported by strong governance, explainability and data readiness”. That’s an admission, from the analyst firm most invested in tracking this space, that the barrier is a design and governance problem, not a workforce attitude problem.

Where AI in Finance Actually Earns Trust Today

It’s worth being specific about where adoption is working, because the pattern is consistent and instructive. Gartner’s research identifies the leading use cases in finance as knowledge management (49% of organizations using AI in finance), AI-powered AP automation processes (37%), and error and anomaly detection (34%). Notice what these three have in common? They’re all tasks where the AI’s output is a recommendation or a retrieval, checked against a visible, verifiable source, rather than an autonomous financial decision made in a black box.

1. Knowledge management works because the AI’s job is to surface information a human already has the context to evaluate; it’s not asking anyone to trust a judgment call, just a faster search.
2. AI-powered AP automation – extracting invoice data, matching line items to purchase orders, and flagging duplicate submissions works because every output is checkable against the source document sitting right next to it. When an AI AP automation software tool flags a duplicate invoice, the controller can pull up both documents side by side and verify the call in seconds. That’s explainability by design, not by afterthought.
3. Anomaly detection works for the same reason. The AI isn’t asked to decide whether a transaction is fraudulent; it’s asked to surface the transaction for a human to evaluate, with the specific factors that triggered the flag shown alongside it.

Compare that to use cases where adoption stalls, like fully autonomous forecasting adjustments, AI-generated journal entries posted without review, or black-box risk scoring used to auto-approve spend. These are precisely the categories where finance leaders’ 78% security and ethics concern rate translates directly into “we built it, but we don’t actually let it run unsupervised.” The pattern isn’t random. It maps almost perfectly onto how explainable each use case is.

What “Trustworthy AI in Finance” Actually Requires

If the diagnosis is a design problem, the fix is a design fix, not a training program. Four things separate AI in finance and accounting tools that earn real trust from the ones that generate impressive demos and disappointing adoption curves:

1. Show the reasoning, not just the conclusion

An anomaly flag without a visible “why” is a black box. The same flag with the specific policy rule, historical pattern, or data mismatch that triggered it is a tool a controller can actually defend to an auditor.

2. Keep humans in the approval loop for anything irreversible

Recommendation and retrieval tasks can run with high autonomy. Anything that commits money, like a payment, a journal entry, or a policy override, should surface the AI’s recommendation for human confirmation, not execute independently, until the organization has built enough trust through a visible track record to change that.

3. Make the audit trail native, not bolted on

Every AI-assisted decision in finance needs the same documentation discipline as a human-made one – what data it used, what rule or pattern it applied, and who confirmed it. AI in corporate finance tools that treat audit logging as an add-on feature rather than a core design requirement will always struggle in a function this compliance-driven.

4. Start with the use cases where verification is cheap

The reason AP automation and anomaly detection lead AI adoption in finance isn’t a coincidence. Verifying an AI’s invoice-matching decision takes seconds because the source documents are right there. Organizations trying to build AI trust should deliberately start with these low-verification-cost use cases before extending into judgment-heavy territory, letting the track record do the persuading rather than the sales pitch.

The Uncomfortable Truth for AI Vendors

If your AI finance tools pitch leads with model accuracy statistics, you’re answering a question finance leaders mostly aren’t asking. The question they’re actually asking is “when this is wrong, and it will occasionally be wrong, will I be able to figure out why fast enough to fix it before it becomes a material misstatement or an audit finding?” Vendors who can answer that question with a visible, transaction-level audit trail will win finance’s trust faster than vendors chasing another percentage point of model precision.

Finance teams don’t distrust AI because they’re behind the curve. They distrust it because most of the AI built for other departments, like sales, marketing, and customer support, was designed around speed and confidence, and finance is the one function in the enterprise where speed and confidence without explainability is a liability, not a feature. Build for that reality, and adoption stops being the uphill climb it currently is.