SAS study finds 97% of workers override unexplained AI decisions

A global survey of 2,700 decision-makers reveals that two-thirds trust agentic systems, with unexplainable conclusions driving routine employee overrides.

Corporate adoption of autonomous artificial intelligence systems is outpacing employee confidence in the technology.¹² In a global study commissioned by software firm SAS of 2,699 decision-makers across 28 countries, nearly 90 percent of respondents reported that AI agents now play a role in their organizations.¹² Yet trust in autonomous agents has fallen behind broader sentiment toward generative models.

According to the survey, 66 percent of enterprise decision-makers said they trust agentic AI systems, compared to roughly three-quarters who express confidence in generative AI tools more broadly.¹² That gap comes as 97.2 percent of respondents acknowledged that workers override automated recommendations in at least some operational scenarios.²³

The cost of unexplained conclusions

The primary driver of manual intervention was not incorrect answers, but opacity. Survey participants cited an inability to explain how a model reached its conclusions as the top reason for overturning an AI decision.¹² That lack of transparency was cited more than twice as often as an outright factual error.¹

"When AI works, it’s incredibly impactful," Bryan Harris, chief technology officer at SAS, said in a statement accompanying the research. "However, it is well documented that state-of-the-art agents can have error rates that exceed 25% on complex tasks – which is unacceptable in high-stakes decision-making. In order to achieve accuracy and repeatability, organizations must embed domain expertise into agentic workflows, while keeping people at the center of governance and oversight."²³

The study found that only 17.5 percent of enterprises operate a data infrastructure mature enough to handle the technical demands of agentic workflows.²³ That structural deficit limits an organization's capacity to audit automated outputs, forcing staff to second-guess tools deployed across core business lines.

Chris Marshall, vice president at market intelligence firm IDC, which assisted with the research, pointed out that enterprise scaling depends directly on technical visibility. "As AI becomes more autonomous, organizations face a new challenge: maintaining confidence in systems people don't fully understand," Marshall said. "Our findings show that stronger oversight, explainability, accountability and data foundations are becoming prerequisites for scaling AI successfully."²³

Sector gaps across banking and government

The tension between rollout and governance varied sharply across regulated industries. In banking, 23 percent of surveyed institutions scored in the top tier of IDC's Trustworthy AI Index, outpacing insurance, government, and life sciences.⁴ Yet only 11 percent of banks achieved an ideal balance of high internal confidence and demonstrably verified systems.⁴

The study found that 47 percent of financial institutions fell into what researchers termed a trust dilemma: either underutilizing dependable AI tools due to skepticism or relying on unverified models without adequate controls.⁴ Data fragmentation compounded the issue, with 19 percent of banks operating on siloed architectures and 45 percent lacking formal data governance policies.⁴

The public sector displayed an even wider disconnect between technical ambition and institutional guardrails. Government agencies reported the highest rate of agentic AI adoption among analyzed sectors at 52 percent, running ahead of banking, retail, and healthcare.⁵⁶ However, only 6 percent of public agencies met the criteria for high confidence paired with verifiable safeguards.⁵⁶

Thirty-eight percent of government organizations reported overrelying on unproven tools while failing to implement standard trustworthy AI protections.⁵⁶ In addition, public sector respondents expressed greater trust in generative AI models than in established machine-learning systems, despite long-running public use of machine learning for tax enforcement and fraud detection.⁵⁶

Returns tied to governance practices

Despite foundational challenges, the survey data showed clear financial divergence between organizations with disciplined governance frameworks and those lagging behind. Companies that scored 80 or higher across five trustworthy AI dimensions were 15 times more likely to report strong returns on investment compared to laggards, recording a 62 percent rate versus 4 percent.²³

High-scoring organizations realized 1.85 times greater gains across 13 distinct performance metrics, including revenue growth and operational resilience.²³ Among industry leaders with established governance practices, 85 percent reported plans to expand their investments in trustworthy systems by more than 10 percent in the coming year.²³

References

This article is based on 7 sources, listed in the order they are cited.

1. 1 JC J.D. Capelouto announcement · 2 Sep 2026 AI deployment in businesses outpaces trust, study finds See the source
2. 2 S sas.com Organizations with trustworthy AI practices are 15 times more likely to see strong ROI, per study findings See the source
3. 3 S sas.com Organizations with trustworthy AI practices are 15 times more likely to see strong ROI, per study findings See the source
4. 4 S sas.com Study: Only 11% of banks have cracked the code on trustworthy AI See the source
5. 5 S sas.com Global study finds many government organizations overrelying on unproven AI See the source
6. 6 S sas.com Global study finds many government organizations overrelying on unproven AI See the source
7. 7 S sas.com ‘An imperative, not a magic pill’: Study finds insurers looking to build trust, value in AI See the source