Two numbers from a new executive survey tell you almost everything you need to know about where AI in business reporting stands right now. Eighty-four percent of executives say they’re at least somewhat confident in AI output accuracy without a human reviewing it first. Eleven percent say their underlying data is actually good enough for AI to work with. That 73-point gap isn’t a rounding error — it’s the reason a growing number of AI mistakes are showing up somewhere they really shouldn’t: in front of boards and outside investors.
Workiva’s 2026 Midyear Executive Benchmark Survey, which polled 2,272 finance, risk and sustainability professionals — including 847 C-level executives — found that 26% of organizations, roughly one in four, had internal audits turn up AI errors that had already reached external audiences or board members before anyone caught them. Separately, 89% of the 367 institutional investors surveyed said they’re concerned about AI accuracy in corporate disclosures. The people closest to the output trust it the most. The people reading it are the least convinced.
The Gap Isn’t About the Model — It’s About the Data Feeding It
It’s tempting to read “AI made a mistake in a board report” as a model problem — pick a better LLM, add a review step, move on. The Workiva numbers point somewhere less comfortable: only 11% of executives believe their data quality is sufficient for AI use at all, 27% say poor data quality has significantly blocked AI deployment in key workflows, and 71% say it has at least moderately affected AI use specifically in financial and sustainability reporting. The model isn’t guessing wrong in a vacuum — it’s confidently summarizing numbers that were already stale, duplicated, or missing context before the AI ever touched them.
A separate study backs this up from a different angle. “Scaling AI Agents with Trustworthy Data,” a report from Google Cloud and MIT Technology Review Insights based on a survey of 300 data and technology executives plus interviews at companies including HCA Healthcare, Shopify and Deutsche Telekom, found that organizations give their AI systems access to only 45% of their data on average — and that gap in access tracks almost exactly with a gap in trust. Among the “data laggards” in the study, who grant AI access to 30% or less of enterprise data, only 22% trust the accuracy of their AI agents’ decisions. Among the “data leaders,” who open up more than 70% of their data with proper context and governance, 100% describe their AI as consistently or mostly accurate. Same technology, wildly different outcome, and the difference is entirely upstream of the model.
Why This Is an Analyst’s Problem, Not Just an IT Problem
Barbara Larson, Workiva’s CFO, put it bluntly: “Confidence in AI without control over data quality is a liability, not a strategy.” That liability lands squarely on whoever signs off on the report before it leaves the building — which, in most organizations, is an analyst or a finance team, not a data engineer. If you’re the one pulling the AI-assisted summary into a board deck or an investor update, you’re the last line of defense between a plausible-sounding number and a document that outlives your job title.
The practical risk isn’t that AI hallucinates wild, obviously-wrong figures — most review processes would catch those. It’s the quietly-wrong ones: a metric computed from a table that’s missing the last week of data, a sustainability figure that double-counts a subsidiary, a revenue number that didn’t account for a currency restatement. Those pass a casual read because they look exactly like something a careful analyst would have written. That’s precisely why 26% of them are getting through.
A Practical Checklist Before You Trust the Number
None of this means pulling AI out of the reporting workflow — both surveys found executives are moving in the opposite direction, toward more agentic use, not less. It means treating the data behind the number with the same scrutiny as the number itself. Before an AI-assisted figure goes into anything that leaves the building, it’s worth running through a short list:
- Trace the number back to its source table or query — can you point to exactly where it came from, or only to what the AI said it came from?
- Check the freshness of the underlying data, not just the output’s timestamp — a same-day summary can still be built on last week’s extract.
- Confirm business context carried over correctly — subsidiary structures, currency conversions, and one-off adjustments are where AI-generated summaries most often go quietly wrong.
- Ask whether the same number, pulled independently, matches — the cheapest sanity check is still a second, boring calculation.
- Keep an audit trail. Forty-five percent of executives in the Workiva survey already say they’ll need better traceability tooling as AI use grows — build that habit now rather than after an error surfaces.
The Takeaway
The confidence gap in these surveys isn’t a reason to slow down AI adoption in reporting and analysis — it’s a reason to stop treating the AI’s output as the finish line. The organizations getting the most reliable results aren’t the ones with the newest models; they’re the ones that did the unglamorous work of fixing data access, governance and context first, so the model had something trustworthy to summarize. If you’re reviewing an AI-assisted number this week, the right question isn’t “does this look right?” It’s “what data did this come from, and would I trust that data on its own?” One in four organizations already learned the hard way what happens when nobody asks.
