97% of executives override AI decisions — culture, not code, blocks adoption

Ninety-seven percent of executives admit they have overridden AI decisions because they could not see the reasoning behind them, according to an industry study by SAS. That number, more than any compute forecast, explains why enterprise AI stalls after procurement: the technology works, and nobody trusts it.

KPMG's research points the same direction. Only 19% of leaders believe their workforce is AI-ready, the firm reported in June. Meanwhile Gartner projects worldwide AI spending will grow 47% in 2026. Companies are buying the race car. Employees are refusing to drive it.

For HR leaders, the implications sit squarely in people-process territory: training design, role definition, oversight structures and the mechanisms by which domain experts feed knowledge into AI systems.

What are the three layers of AI adoption?

The source framework — drawn from a veteran IT security perspective — argues that AI adoption rests on three layers: foundations, trust and adoption. Foundations are the technical and regulatory guardrails. Trust is employee confidence in AI outputs. Adoption is the infusion of organizational expertise that turns a generic model into a working capability.

Foundations go beyond hardware. Any company will soon be able to buy or rent enough computing power to run AI at scale, the argument runs. The harder task is governing what AI agents (autonomous systems that can act, not merely monitor) actually do: tracing their decisions, understanding why they make them, and meeting compliance obligations that vary by industry and region. Financial services firms must satisfy FINRA in the U.S.; healthcare institutions face HIPAA and patient-privacy rules.

Governance also demands flexibility. Systems built around a single model, vendor or use case become slow to update, and slowness creates gaps — reputational, financial or operational. The human-in-the-loop question follows the risk profile: low-risk processes may run without intervention, while a function like insurance underwriting still requires human final judgment.

Why does trust fail?

The trust problem is not new. Machine learning faced the same wall in its early days: security professionals distrusted black-box models even when those models were likely more accurate than humans at spotting vulnerabilities. Accuracy without transparency did not sell then. It does not sell now.

Trust also has a security dimension, and it cuts both ways. AI is described here as more powerful than any existing defense — capable of finding and exploiting any vulnerability, breaking out of sandbox environments and reaching the open web. Long-dormant flaws are effectively zero-day vulnerabilities now. The recommended posture is offensive: companies should run AI attacks against their own networks to surface gaps before adversaries do, and benchmark agents against realistic scenarios rather than chasing the elimination of risk.

How should organizations handle adoption?

The analogy is blunt: AI currently operates like a junior associate. No CISO would put a first-year employee alone on the front lines; no one would hire dozens of new engineers and grant them full autonomy. Yet that is roughly what unsupervised agent deployment amounts to.

Agents lack organizational context — the unwritten rules, the decision-ownership that an org chart never shows. The fix mirrors junior-employee training: domain experts supply design specifications, training material and, above all, continuous feedback. An AI produces output, experienced professionals evaluate it, and the feedback becomes system context for the next task. That loop, not model size, converts a generic technology into an organizational capability.

Employees, in turn, need training, visibility into what AI is doing, and confidence that actions are governed with appropriate human oversight. The winners, the argument concludes, will treat employees as part of the system — because the largest computing environments matter little if the workforce and the technology are not designed to work together.