In May 2026, KPMG released a research report on AI adoption in finance that contained one number worth circling immediately: 42 percent. That's the share of companies with AI-enabled finance processes that can actually prove their AI works. The other 58 percent? They're using AI to reconcile accounts, forecast cash flow and generate financial statements while their auditors have no way to verify the output.

This gap isn't theoretical. As AI moves into core accounting workflows—reconciliation, close processes, expense categorization, variance analysis—the question shifts from "Is AI useful?" to "Can we audit it?" And right now, most finance teams are failing that second question.

The Adoption-Assurance Gap Is Widening

AI adoption in finance has accelerated. Seventy-five percent of companies now use AI in finance, nearly double the 30 percent rate from two years ago. And 71 percent of finance leaders say AI is meeting or exceeding ROI expectations, especially for decision quality, forecast accuracy and responsiveness.

But adoption and assurance are not the same thing.

According to KPMG's research, only 42 percent of organizations are "fully assurance-ready" for their AI. This means they can produce audit evidence that the AI works as designed, explain how the AI made a specific decision, track when and why it fails or deviates from normal patterns.

Organization Type AI Adoption Rate Assurance-Ready Error Reduction Scale Confidence
Assurance-Ready (42%) 75% use AI Full audit trail + explainability 33% error reduction 42% confident in scaling
Not Assurance-Ready (58%) 75% use AI Limited visibility into decisions 6% error reduction 14% confident in scaling

The outcomes tell the story. Assurance-ready organizations see three to six times higher error reduction compared to their peers. They catch mistakes in reconciliation, expense coding and variance analysis at rates the other group can't touch. They're confident enough in their controls to keep expanding AI use without fear of audit surprises.

Why Can't Most Firms Explain Their AI Decisions?

The barrier isn't intention. Finance leaders aren't dragging their feet on purpose. The problem is structural.

First: data quality. Thirty-six percent of firms cite this as their biggest barrier. If cost centers are mislabeled, transaction dates are inconsistent, or account mappings are outdated, the AI learns those patterns. When your auditor asks why the model coded something a certain way, the honest answer is often "because your historical data was messy." That's not assurance-ready.

Second: visibility. Most organizations lack basic transparency into where and how their AI operates. Only 29 percent of firms track where AI fails. That means seven in ten companies don't know when their AI is making mistakes or operating outside its training parameters. You can't audit what you can't see.

Third: documentation. AI runs in the background. It processes thousands of transactions daily. You see outputs but not decisions. An internal auditor can't easily trace back and ask, "Why was this journal entry suggested?" The audit trail either doesn't exist or is in a format auditors don't know how to evaluate.

What This Means for Your Practice

The assurance readiness gap isn't a technology problem you can solve by upgrading your software. It's a control problem. Your finance software vendors have shipped AI features, but they haven't necessarily shipped the documentation, explainability and tracking that auditors require. That gap is yours to close.

Three Questions Your Auditors Will Ask

Before your next audit, prepare answers to these three questions. If you can't answer them clearly, you have a control gap to address.

One: How does the AI decide? Not "what does it do," but how. Your external auditors will want to understand the model logic. Not the technical details, but the business logic. "It reviews historical patterns and flags outliers" is not enough. "It identifies transactions outside the range of the account's standard deviations, applies a company rule that we've documented and flags anything above the threshold we've set" is defensible.

Two: How do you test it? Assurance-ready firms have validation processes. They run the AI model on a test dataset with known results, confirm it produces correct outputs and retrain or adjust the model when accuracy drifts. You should be able to show your auditor: "Here's how often we validate. Here's when it last failed. Here's how we fixed it."

Three: What happens when it fails? Every model fails sometimes. The question is whether you know when, and whether you have a documented response. If the reconciliation AI suggests an adjustment and it's wrong, does it loop back to a human reviewer? Does that exception get logged? Can you trace it? If you can't answer this one, your AI decisions aren't defensible under audit.