What happened
Deloitte published its *State of AI in the Enterprise* report in October 2026. The report frames the enterprise challenge as shifting from access and experimentation toward making AI economically durable and integrating it into how work is organized. It describes agentic AI as beginning to move systems from answering questions and supporting decisions to taking bounded, multistep actions within workflows. It also discusses AI economics, work redesign, governance, physical AI, and sovereign AI.
The headline adoption finding is that 81% of surveyed companies say AI has improved the speed of decision-making and execution. Deloitte contrasts that with 25% using AI for what it calls deep transformation: creating new products or services, reengineering core processes, or fundamentally reinventing a business model. The report’s distinction matters: faster execution is a widespread benefit, but it does not by itself establish that the business has changed how it creates value.
The economics findings point to a visibility gap. Seventy-one percent of companies track total AI inference and API spend, while 51% track token volume and 44% track per-unit inference cost. Deloitte argues that organizations may therefore know the total bill without knowing what is producing it. The material supplied here does not provide breakdowns by company size, industry, or specific AI application, so these figures should be read as survey-level findings rather than a forecast for any one team.
The report’s recommendations connect measurement to operating choices. Its listed priorities include redesigning work rather than only workflows, tracing cost to the token rather than only the total bill, connecting AI initiatives to commercial decisions that produce revenue, and matching governance to the autonomy AI is actually given. The report presents autonomy in categories ranging from informing and recommending to executing within limits and acting autonomously, with approval, oversight, escalation, and accountable ownership varying by level.
The supplied report material identifies the work as a survey of 3,500 global business leaders. It gives no survey field dates, sampling details, or question wording in the excerpt available here. The reported percentages therefore describe respondents’ reported experience; they do not independently prove that AI caused faster decisions or that tracking unit costs leads to higher returns.
What the archive adds
A prior Telegram post on Deloitte’s *Scaling Agentic AI to realize business value* emphasized the operating-model angle: shared platforms, reusable components, and lifecycle governance. The new report adds survey figures on decision speed, transformation, and cost visibility; it is a new fact in a related discussion, not a restatement. We have no genuinely related prior Field Note link to add here.
Trend signal
This report supports AI-native organizations move toward review and orchestration: the report argues that value depends on redesigning work, assigning accountability, and governing the autonomy given to AI—not simply increasing usage. Its survey figures are evidence of the gap it describes, rather than proof that any single operating model will close it.
Why now
Deloitte says enterprises are moving beyond questions of access and adoption toward cost, token consumption, and return on investment. In its framing, agentic systems are beginning to take bounded actions inside workflows, making the relationship between autonomy, oversight, and business value a more immediate operating question.
Connections to other events
The earlier Deloitte agentic-AI operating-model post connects this report’s new survey findings to a practical management concern: scaling requires more than individual tools. This report adds a measurement angle—how much of the bill is visible at the unit level—and a distinction between accelerating existing work and transforming products, processes, or business models.
Why it matters
For teams building with AI, the findings caution against treating usage growth or faster task completion as sufficient evidence of business impact. A rising total invoice can conceal which workflows consume tokens and whether their cost is justified by a useful result. Conversely, visibility into per-unit cost is not itself ROI; it is an input to decisions about which work to automate, redesign, or keep under human review.
The report’s emphasis on work redesign is relevant to agent builders because systems that take multistep actions change where decisions and accountability sit. Teams need to define boundaries, oversight, escalation, and an accountable owner in proportion to what an agent is allowed to do. That makes evaluation a business-design problem as well as a model-quality problem: measure the outcome the workflow is meant to improve, the cost of producing it, and the review burden it creates.
What happens next
- More enterprise AI programs will track inference cost per unit alongside total spend. (6 months, 65%). Watch for: organizations reporting unit-cost metrics in AI program reviews.
Limits and uncertainty
The supplied material does not establish that AI caused the reported improvements, nor does it provide the underlying survey questionnaire, field dates, sampling method, or respondent-level data. The percentages are self-reported and should not be treated as universal rates for every sector or company. The figures also do not show whether the organizations tracking token volume or per-unit inference cost achieve better returns. Deloitte’s recommendations are guidance, not evidence that adopting them will reliably produce transformation or revenue gains.
Skill to build now
Pick one AI-enabled workflow this week. Record its intended business outcome, the human review and escalation points, and the available inference or API cost at the workflow level. If you can see only the aggregate bill, note that as a measurement gap rather than inferring unit economics. This small baseline helps distinguish a faster existing process from a change that affects the business outcome.
Reader takeaway
The report’s central contrast is simple: AI-enabled speed is common in respondents’ accounts, but deep transformation is not. For builders, the next useful question is not only whether an agent can complete a task faster, but what outcome it changes, what that outcome costs, and who remains accountable when the system acts.