For years, the AI conversation has focused on what systems can automate, predict, and now, generate. As organizations increase their investments in AI, many are finding that the business value increasingly depends on whether people understand the reasoning behind automation, predictions and generation well enough to trust it.
In early September, SAS released the 2026 Data and AI Impact Report: The New Economics of Trust, the second edition of an annual flagship study examining AI adoption, trustworthiness and business impact. Developed by SAS with research insights from IDC, the report draws on responses from 2,700 decision-makers across 28 countries and is part of a three-year global initiative to understand what separates organizations creating measurable value from AI from those still struggling to move beyond experimentation.
While the inaugural report measured what people believed about AI, the new one measures what they actually do with AI. The findings suggest those turn out to be very different things, something I’ll get to in a moment. But one thing that remains consistent across both reports is that trust and trustworthiness, anchored by explainability, remain key to successful adoption.
For executives, the implication is straightforward: explainability is not a compliance sidebar or a technical preference. It is a condition for scaling AI into decisions where speed, accountability and financial return all matter.
Explainability is not a compliance sidebar or a technical preference. It is a condition for scaling AI into decisions where speed, accountability and financial return all matter.
Nearly 89% of respondents say AI agents already play some role in decision-making, from informing human-approved decisions to acting autonomously. Trust has not kept pace with that transition. Generative AI is trusted by 76% of respondents, while trust in agentic AI falls to 66%. As responsibility shifts from a person using AI to a system acting on a person's behalf, the standard for confidence rises with it.
The cost of unexplained AI
That helps explain one of the report's most revealing findings. Nearly every respondent said they override AI-generated recommendations at least occasionally, and 79% do so in more than one out of every 10 cases. Across thousands of organizations, people are routinely stopping, reviewing and reworking recommendations produced by systems their employers have already paid to build, deploy and maintain.
At the Fortune 500 scale, that hidden cost compounds quickly. A single extra review in one workflow is a necessary good judgment. However, when repeated across customer interactions, fraud decisions, supply chain exceptions, HR processes and financial approvals, that review absorbs the productivity gains AI was expected to deliver.
What makes this more consequential is that factual inaccuracy is not the leading reason for intervention. The top reason, cited by 34.5% of respondents, is insufficient explanation. Missing context follows at 25.6%, while factual error ranks lower at 14.8%. People often step in because they can’t see the rationale AI used to make a decision, so they exercise their own judgment instead.
In consequential decisions, we want to understand the reasoning, see the relevant context and know which evidence mattered. We expect that from one another, and there is little reason to believe we will expect less from machines acting with growing autonomy.
Explainability requires the model, the data lineage and the interface to surface reasoning that a human can follow and, ideally, address. A recommendation that is not easy to understand is harder to implement. A decision that cannot be traced is harder to defend. A system that cannot explain itself shifts the burden of interpretation back to the person on the other side of the screen, leaving the organization dependent on individual review rather than institutional confidence.
If trust is how much confidence people have in AI, then trustworthiness is the evidence that justifies that confidence. The report shows that organizations have made real progress on trustworthiness this year. The Trustworthiness Index, which measures data quality and governance, model governance and oversight, explainability and fairness, responsible AI policy, and audit and accountability, rose from 49.7 to 61.3. Perceived trust, however, slipped from 73 to 70.7. For leaders, the gap is a warning that better controls do not automatically translate into greater confidence unless employees can see how those controls improve the decisions they are being asked to make.

Among the five dimensions in the index, Explainability and Fairness improved the least, and insufficient explanation remained the leading cause of overrides. Policies can be written, committees established and accountability assigned, all of which matter, but explainability requires connecting source data, model behavior and the user experience so that a person can follow the reasoning and respond to it.
In short, explainability begins with data.
Explainability starts with data
A system can explain only what it can trace, yet just 17.5% of organizations report operating at the highest level of data infrastructure maturity. Without high-quality, well-governed data and clear lineage, an explanation may appear transparent without the evidence needed to support it.
For business leaders, the return on getting this right is difficult to dismiss. Organizations with high trustworthiness scores are 15 times more likely to report strong or high returns on AI investment than those with low scores (62% vs. 4%). They also realize 1.85 times the gain across 13 business outcomes. These figures do not prove that a single explainability investment will produce a prescribed return, but they show that organizations building trust through sound data, governance, validation, accountability and transparency are converting AI spending into value at a very different rate.
Leaders, as defined by this research, are not investing in trust because they are afraid of AI, but because trust allows AI to move from demonstration to deployment and from isolated productivity gains to enterprise-scale value. They optimize their data infrastructure, validate continuously, train employees, govern by default and keep investing in the practices that make responsible autonomy possible.
The leaders in this research are not investing in trust because they are afraid of AI. ... trust allows AI to move from demonstration to deployment and from isolated productivity gains to enterprise-scale value.
That is why I view explainability and governance as growth capabilities and economic enablers. The organizations that are sustainably successful may not be the companies with the most advanced models, but the ones that reduce the distance between system output, human understanding and business action. When leaders make that connection visible, trust becomes more than a safeguard; it becomes the mechanism that turns experimentation into return at enterprise scale.