Finance Teams Spend 13 Hours a Week Fixing AI Outputs

What is the "glass box" model in AI finance — and why does it matter now?

Glass box AI shows every output, assumption and data source in real time — letting finance teams act on results without reconstructing the logic afterward.

The term comes out of Sage Future 2026, the annual Sage finance and technology conference held April 28–30 in San Francisco. Sage CEO Steve Hare used it to describe what his company believes trusted AI in finance must look like: "Finance does not run on answers alone — it runs on answers you can explain. If you cannot show how a number was produced, you cannot use it."

The contrast is the black box model that currently dominates AI tools in finance. Black box AI produces outputs without making its reasoning visible. Finance teams consume those outputs, then spend significant time asking: Where did this number come from? Is this assumption correct? Can I defend this to my auditor? That reconstruction work is not a temporary onboarding cost — it's an ongoing operational overhead that compounds every time AI touches a financial process.

What does the IDC research actually say — and what does 12.9 hours a week mean for your firm?

Finance professionals are spending an average of 12.9 hours per week on AI validation work — not using AI, but checking whether AI outputs are correct enough to act on.

That figure comes from research Sage commissioned from IDC and announced at Sage Future 2026 on April 28, 2026. The full finding: 71% of finance leaders would reject an AI system if it cannot explain its outputs, even when the outputs are highly accurate. The trust issue isn't about AI capability — it's about visibility into how the capability arrived at its conclusion.

The 12.9-hour figure is the operational consequence. In a 40-hour workweek, that's roughly a third of a finance professional's time spent on work that AI was supposed to eliminate. And the research quantified the leakage precisely: 26% of AI time savings are consumed by verification, explanation and reconstruction work. So the net productivity gain from AI adoption in finance is currently being offset — in measurable terms — by the trust cost of systems that don't explain themselves.

How do you tell a glass box AI from a black box AI?

The simplest test: ask the system to show you its work. A glass box AI can. A black box produces a result and leaves you to verify it manually.

Sage's conference session made this concrete with a real-world case. An AI agent handling financial data performed well initially — then began making subtle errors. It misclassified entries, removed data and introduced inconsistencies that only became apparent under review. The outputs looked credible. Nothing flagged automatically. Human judgment caught the problem. That's a black box failure: the system's reasoning wasn't visible enough for anyone to see the drift happening in real time.

In a glass box model, that same scenario looks different. The misclassification would be traceable — you could see which data source the agent was drawing from and which assumption it had applied. The correction is targeted, not a full re-audit. More importantly, the error doesn't compound across multiple periods before someone catches it.

The practical question for evaluating any AI finance tool is whether it supports auditability at the point of use — not as an optional export or a support ticket, but as a default behavior. Can you click through to the underlying transaction? Can you see which rule was applied and why? If the answer is no, the tool is a black box regardless of how its marketing describes it.

What is the Sage and PwC "Beyond the Black Box" initiative?

Sage and PwC jointly committed at Sage Future 2026 to making AI explainability a standard in finance software, combining product development with governance frameworks.

The initiative is called "Beyond the Black Box." Sage is the software provider; PwC is the partner bringing deployment experience and risk management structure. PwC's involvement is notable specifically because it's already operating AI at scale internally: 86% of PwC employees are actively using AI tools, with more than 240,000 Microsoft Copilot licenses deployed and over 4,000 custom GPTs built and reused across the firm. That's not a firm talking about AI governance theoretically — it's a firm that has encountered the black box problem in practice and built internal governance around it.

For finance teams, the initiative is less a product announcement than a signal: the explainability gap in AI tools is large enough that a major accounting software vendor and one of the Big Four are jointly committing to address it. Sage's framing — calling it the "trust cost of AI" — gives practitioners language for a problem they've been absorbing without being able to name it clearly.

What should you demand from AI finance tools before your next contract renewal?

Demand three things before any AI finance tool renewal: visible reasoning, traceable source data and a clear way to find errors before they compound.

The first question tests the glass box requirement directly. Vendors who answer with "you can export a summary report" are describing a post-hoc audit tool, not an explainable AI system. The distinction matters because auditability after the fact still requires significant reconstruction work — that's the 12.9-hour problem. What you need is transparency at the point of use, before you act on the output.

The second question targets data lineage. For any AI-generated financial output, you should be able to trace the number back to the source transaction or data source without a support ticket. If the vendor can't demonstrate this during a sales conversation, the tool doesn't have it.

The third question is about error detection. The Sage conference case study — the agent that gradually introduced misclassifications — is not an edge case. It's a predictable failure mode for any AI system running in a production finance environment. Ask the vendor how the system surfaces its own errors. If the answer is "you'd catch it in review," the burden of error detection has been transferred back to your team. That's not AI reducing overhead — that's AI creating a new category of oversight work.