97% of Executives Override AI Decisions Over Trust Gaps

Ninety-seven percent of executives admit they have overridden AI decisions because they could not see the reasoning behind them, according to a SAS Data and AI Impact Report. That number, more than any infrastructure benchmark, explains why enterprise AI initiatives stall: the barrier is no longer technical capability but trust.

A second figure sharpens the point. Only 19% of leaders believe their workforce is AI-ready, per KPMG research. Companies are racing to provision compute, unify data and modernize security — but employees still have to know how to wield the tools and understand their role in agentic workflows, both as a security check and as a mentor to the system.

A three-layer framework — foundations, trust, adoption — maps how to manage the technical and organizational change together.

What do the foundations actually require?

Hardware is the easy part. Any company will soon be able to buy or rent enough compute to run AI at scale. The hard part is governing agent activity: tracing what agents do, understanding why they make specific decisions, and governing those actions across security, reporting, data access and compliance.

Regulation varies by industry and region. Financial services firms must comply with FINRA in the U.S.; medical institutions face HIPAA and patient privacy rules. Governance must be designed in from the start, not bolted on.

Flexibility matters as much as coverage. When governance is built around a single model, vendor or use case, every shift becomes a manual, time-consuming update — and slow adaptation opens gaps that cause reputational, financial or operational harm. Risk-based autonomy is the balancing act: low-risk processes can run without human intervention, while functions like insurance underwriting still need human final judgment.

The takeaway: foundations mean traceability and compliance, not just compute.

Why does trust fail even when AI works?

Technology projects fail when organizational change is ignored. You can build the fastest race car in the world; if nobody wants to drive it, the speedometer never moves.

Machine learning already ran this experiment. Early ML security tools operated as black boxes, and practitioners — even facing models likely more accurate than humans at finding vulnerabilities — did not trust outputs they could not verify. For users, trust means proof, and agents must outline their thought process and actions.

Security adds urgency. AI is more powerful than existing defenses: agents will find and exploit any vulnerability, break out of sandbox environments and reach the open web. Flaws that sat undiscovered for decades are becoming zero-day vulnerabilities overnight. The countermeasure is offensive: companies should attack their own networks with AI to expose gaps before bad actors do, and benchmark new agents against realistic scenarios. The goal is not eliminating risk but managing how agents change the risk equation.

The takeaway: transparency and adversarial testing, not reassurance, build trust.

How do organizations turn AI into an organizational capability?

A general-purpose model does not know why an engineer configures things a certain way, what precedent applies, or who owns decisions in each domain. Reading an org chart, an agent still cannot tell who holds decision-making authority.

Today's AI operates like a junior associate. No CISO would put a first-year hire alone on the front lines, and no one would grant dozens of new engineers full autonomy. Agents need the same treatment: domain experts supplying training material — pages of design specifications, power requirements, site-selection criteria — and continuous feedback on outputs.

The loop is the mechanism: AI produces output, experienced professionals evaluate it, feedback is captured, and that knowledge becomes system context for the next task. That is what converts a generic technology into an organizational capability.

Employees need training, visibility into what AI is doing, and confidence that actions are governed with appropriate human oversight. The winners will treat employees as part of the system, because the largest compute environments matter less than how workforce and technology work together.

The takeaway: mentorship loops, not model size, compound AI value over time.

As agents take on independent action across operations, product and security, the competitive edge will sit with organizations that solve the human layer first — governance, trust and expertise — rather than the ones that simply buy the most compute.