Companies are letting AI act—but who is accountable when it goes wrong?
In Deloitte's survey of 3,500 AI-involved leaders, 40% said their organisations let AI execute decisions or actions without human approval to at least a moderate degree, while 23% said accountability for a negative outcome remained unclear. The figures are self-reported, not audited incident data.
The short version
- What happened
- Deloitte surveyed 3,500 director-to-C-suite business and IT leaders involved in AI initiatives across 25 countries and six industries in May and June 2026.
- Who it affects
- Workers and managers may be asked to supervise systems that can act before approval. Clear decision rights, escalation routes and the power to reverse an action matter more than a nominal human-in-the-loop label.
- What remains unproven
- Consultancy survey
Keep in mind
Consultancy survey
3,500 leaders
25 countries
Six industries
Fielded May–June 2026
AI-involved respondents
Self-reported outcomes
No response rate disclosed in the public materials
No audited ROI
No incident denominator
Deloitte commercial interest
Source analysis · 2 linked sources. Primary analysed source date: 29 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
Autonomy is spreading faster than clear responsibility
Deloitte's latest enterprise survey points to a practical governance gap: AI is beginning to act inside organisations while responsibility for those actions is not always settled. Among 3,500 leaders directly involved in their companies' AI initiatives, 40% said their organisation allowed AI to execute decisions or actions without human approval to at least a moderate degree. Separately, 23% said accountability remained unclear when an AI system produced a negative outcome.
Those percentages do not mean 40% of all companies have fully autonomous agents. The survey question uses a broad threshold—at least a moderate degree—and respondents were selected because they had direct involvement in AI work. Deloitte does not publish a denominator of automated decisions, the types of decisions covered or the severity and frequency of failures. The result is best read as a warning about governance reported by AI-active leaders, not a measured rate of uncontrolled systems.[1][2]
Speed was common; deep business change was not
The same survey found that 81% of respondents said AI had improved the speed of decision-making and execution, but only 25% described their organisations as deeply transforming the business. Sixty per cent reported redesigning jobs around AI to at least a moderate degree, while 15% called those changes extensive or complete. For workers, that distinction matters: adding a faster tool to an old approval chain is different from changing who decides, who checks and who carries the consequences.
Deloitte also reports that 28% of leaders attributed revenue gains to AI, up from 20% in its January study. That comparison is not an audited return-on-investment measure. The public materials do not show company accounts, a matched control group or a design that isolates AI from demand, staffing, pricing, acquisitions or other technology investments. A respondent's attribution can reveal what leaders believe is creating value without proving how much AI caused.[1][2]
The distinction that matters
Three different claims in the survey
Leaders reported faster work, broader autonomy and attributed business gains, but each statement needs a different evidence test.
Speed
81% reported faster decisions or execution
A perceived process improvement, not a measured accuracy or safety result
Autonomy
40% reported some action without human approval
Scope and decision denominator were not published
Value
28% attributed revenue gains to AI
No audited causal return was reported
Cost visibility is another weak point
Seventy-one per cent said their organisations tracked total AI inference and application-programming-interface spending, but 44% tracked the cost per unit of inference. A total bill can show that spending is rising; unit costs are needed to compare models, workflows and volumes. Even then, cost per call does not establish value unless organisations connect it to a defined outcome, error rate and human review burden.
Deloitte is both the author of the research and a seller of AI consulting services. Its report is useful because it supplies a large, multinational sample and a fieldwork window, but it is not independent evidence of client outcomes. The public release identifies 25 countries and six industries—consumer; energy, resources and industrials; financial services; life sciences and healthcare; technology, media and telecom; and government and public services—but does not disclose a response rate or enough weighting detail to establish population representativeness.[1]
What would strengthen the assessment
The next evidence should connect system permissions to consequences: how many actions bypassed human approval, which decisions affected customers or workers, what controls could stop or reverse an action, and how often incidents occurred. Independent audits could test whether named accountability owners have authority, information and time to intervene rather than existing only on an organisation chart.
For business value, comparable financial and operational data would matter more than another confidence survey: verified revenue or cost changes, a pre-AI baseline, a credible comparator, implementation and review costs, and error-adjusted outcomes over time. Until then, the survey's strongest signal is organisational rather than financial: companies are giving AI more scope to act before all of them can clearly say who answers for the result.[1][2]