The finance AI conversation shifted at Sage Future this spring. Not "can AI do the job?" but "can you explain what it did?"
That's a meaningful upgrade. According to Ncontracts research, 72% of financial institutions report that their boards now hold leadership accountable for AI-driven decisions. ERP Today tracks CFO accountability similarly, putting it near 70%. The message is clear: signing off on an AI-generated output means you own it.
Sage built their response to this into the product. Their trust framework runs on three pillars: Confidence (outputs are grounded in real data, not invented), Control (AI suggests, it doesn't post), and Accountability (every action is fully auditable and traceable).
Aaron Harris, Sage's CEO, put it bluntly at the conference: "If I can't figure out why it did what it did, it's not functional. It's a paperweight."
He's right. Explainable AI is the correct direction. Finance teams deserve to see the reasoning behind every recommendation, and regulators and boards are starting to demand it. The "glass box" framing is more than a product feature. It's a shift in how vendors think about their responsibility to the people using their software.
Here's the problem: glass box AI shows you exactly where its reasoning came from. What it can't do is tell you whether the data it started with was accurate.
Transparency at the AI layer doesn't fix trust problems upstream. An auditable AI will confidently trace an output straight back to a corrupted source. That's not accountability. That's a paper trail leading to a dead end.
The integration layer, the pipes that move data between your systems before the AI ever sees it, is where most organizations have their real exposure. And most of those pipes weren't built for the accountability standard the AI layer now requires.
Consider a few scenarios that are more common than they should be in mid-market finance:
A Salesforce-to-Intacct connection running on a nightly export. Same-day changes don't make it into the books until morning. The AI closes the month with yesterday's data.
A Stripe reconciliation that handles straightforward transactions cleanly but drops edge cases around subscription upgrades mid-cycle. The discrepancy is small enough to miss in a manual review. The AI flags it as reconciled.
A manual field mapping that someone updated in the source system six months ago and nobody caught. The connection kept running. The data kept flowing. The values stopped meaning what anyone thought they meant.
In each case: your AI is auditable. Your data pipeline wasn't.
The standard for integration has to rise to match the AI layer sitting on top of it. That means a few things in practice:
These aren't exotic requirements. They're what accountability actually looks like when AI is interpreting the output.
Whether you're evaluating a new connection or auditing an existing one, these are worth asking:
If the answers are vague, that's information.
Glass box AI is a meaningful development for finance. Sage is building the right things, and the accountability conversation at Sage Future reflected a real shift in how the industry is thinking about AI governance.
But explainability at the AI layer only gets you so far. If the data feeding your AI isn't trustworthy, auditable, and structurally sound, you'll have a very clear view of a problem you can't fully explain.
The audit trail has to start at the source: the connections between your systems, the pipelines that carry your data, and the standards those pipelines are held to. If that layer can't show its work, your AI can't either.
If you're thinking through what integration accountability looks like for your organization, we'd be glad to talk through it.