Enterprise AI faces challenges related to accountability and cost

Deloitte’s New AI Report Reveals Shift Towards Autonomous Decision-Making in Enterprises

In its latest study on enterprise AI, Deloitte reveals a significant transition where artificial intelligence is moving from advisory roles to executing decisions independently, often without human validation. While advancements in responsibility and governance frameworks lag behind this rapid development, the acceleration witnessed by companies has not been translated into substantial transformations as it was earlier this year.

On October 29, 2026, during an event in New York, the Deloitte AI Institute unveiled its latest edition of “The State of AI in the Enterprise.” The survey, conducted in May and June 2026, involved 3,500 executives from 25 countries across six industries, with roughly half of the respondents from IT departments. All participating companies have at least one AI deployed in production, and the report is further supported by in-depth interviews with 15 leaders and AI specialists. In comparison, January’s edition, which surveyed 3,235 executives, highlighted a pressing urge to deploy AI agents without mature governance structures in place.

Fast forward nine months, and 81% of companies report that AI is speeding up decision-making and execution processes. Notably, 40% allow AI to execute decisions or actions with minimal human validation. Deloitte now identifies the key bottleneck for value creation as the organization’s capacity to redefine work, accountability, and governance commensurate with the autonomy granted to AI systems.

Execution Without Human Validation and Ambiguous Accountability

Deloitte differentiates between AI that recommends actions and systems that execute them autonomously; the latter necessitates tighter oversight. Yet, it’s the latter category that is gaining momentum. According to the report, when an AI system produces a negative outcome, 36% of companies hold the responsible team accountable, while 24% point to a centralized governance function and 10% to the employee who engaged the system. Alarmingly, nearly a quarter (23%) admit that accountability remains unclear. Consequently, Deloitte advocates for clear human approval tiers on a case-by-case basis, determining what requires prior human endorsement, what can be reviewed post-factum, and what mandates immediate escalation.

Acknowledging the variety of accountability models, Deloitte rejects ambiguity. “Different accountability frameworks can be justified based on usage; however, uncertainty is not permissible. When accountability is unclear, issues can shift from team to team, leaving no one with the authority or obligation to resolve them,” the authors state. Furthermore, the limited transfer of decision-making authority exacerbates the issue; although 60% of companies have redefined roles around AI capabilities to some extent, only 15% report a significant shift in decision-making power. Employees may have better access to information and quicker decision-making capabilities, but they lack the authority to act differently due to static roles, outdated approval processes, and limited decision rights.

Governance for Agent Chains in Less Than One in Ten Cases

The governance gap widens when multiple agents collaborate on a single process. Among the 2,372 companies at this stage, which make up more than two-thirds of the sample, half lack any formal governance or only implement basic oversight, with just 9% deeming these chains as fully governed. This finding echoes previous studies in France that noted the deployment of AI agents before organizations fully understood how to monitor them.

Deloitte suggests calibrating governance to match the effective authority of the system. It distinguishes four levels of AI intervention: systems that inform decisions, those that recommend actions, those that execute within defined parameters, and those that act autonomously. Each level should have tailored rules for approval, oversight, escalation, and accountability. For agent systems, Deloitte advises implementing controls at the execution stage through clear identities, permissions, rule enforcement, monitoring, and traceability, instead of relying solely on policies and periodic reviews. Tools employed by IT departments for human accounts and applications are seen as enablers for adoption, signaling to teams where they can operate independently and when approvals are necessary.

A Clear Invoice Yet Opaque Unit Costs

The second significant blind spot pertains to financial investments. According to Deloitte, 86% of companies plan to increase their AI budgets in the upcoming fiscal year, with 34% anticipating a rise of 20% or more. While overall spending is closely monitored—with 71% tracking their inference and API costs—visibility diminishes at the granular level, with 51% tracking token consumption and 44% monitoring unit costs per token or API call. Technology companies, in particular, track unit costs 1.6 times more often than consumer goods firms.

This expense, charged on a per-token basis, varies depending on model complexity, reasoning chain length, and infrastructure choices. Deloitte describes this as resembling a variable economic system rather than a fixed cost of licenses or virtual machines. The nature of agent workloads exacerbates this impact, as repeated model calls, new attempts, tool usage, and agent interactions multiply the computational requirements for a single business task. Some organizations incentivize AI adoption without linking it to measurable outcomes, risking costs that may subtly exceed the value generated by tasks it replaces. “While many organizations better monitor their total AI spending, far fewer grasp the underlying unit costs, where expenses arise, and whether they deliver corresponding value. It’s in this disparity that the return on investment can evaporate,” cautions Costi Perricos, Deloitte’s global head of generative AI.

To address this gap, Deloitte recommends tracing overall expenses back to processes, models, calls, and tokens, necessitating consistent measurement and assignment of consumption to the activities that create it. The firm also advocates for designating an accountable owner for the economics of each significant workload. With this foundation, companies can compare models, optimize queries, and ensure that reductions in unit prices lead to overall cost savings.

Accelerated Speed, Yet Less Transformation Than in January

Regarding transformation, only 25% of companies leverage AI to fundamentally alter their operations through new products or services, core process reinventions, or modified business models, a drop from 34% earlier this year. The remaining three-quarters are accelerating processes while leaving their fundamentals intact. Nonetheless, financial gains are on the rise, with 28% of companies reporting revenue increases attributed to AI, up from 20% in January. Efficiency and productivity gains have also risen to 74%, up from 66%, and cost reductions now stand at 45%, compared to 40% earlier. Revenue remains the most anticipated benefit, with 67% of firms citing this as a primary goal; Deloitte notes that those deriving revenue from AI also report a wider array of additional benefits.

“It is clear that AI is already helping companies make and execute decisions more swiftly,” observes Jim Rowan, Deloitte’s AI leader in the United States. “The next challenge is to transform that speed, safely, into meaningful change. Leaders need to look beyond isolated use cases and question whether the underlying work still makes sense,” he continues. The firm advises eliminating unnecessary steps, handoffs, and approvals, while streamlining decision-making closer to the work itself. It also urges clarity on what AI can recommend, what teams can do directly, and what remains subject to approval. For every project expected to generate revenue, companies should articulate the mechanism—be it pricing, demand, conversion, or retention—and assign a business leader with the authority to act on insights provided by AI.

Tiered Governance Within Reach of IT Departments

The gaps identified by Deloitte necessitate responses that IT departments are equipped to build, such as establishing identities and permissions for each agent, logging their actions, and measuring consumption relative to business processes. However, the decision on what an agent can autonomously execute is shared with business units, as it redistributes decision-making authority that Deloitte finds is rarely changing.

These insights are primarily declarative—reflecting perceptions and intentions of executives rather than quantitative measures—and stem from a firm that also markets governance and economic consulting services in AI. The U.S. accounts for 1,200 respondents out of 3,500, and no significant results arise from the 130 French executives surveyed. The trend between editions is clear: AI is gaining autonomy faster than organizations are transferring accountability and recognizing associated costs.

Scroll to Top