Enterprise AI ROI just crossed from productivity story to revenue system

The table below consolidates the most significant statistics from five major research programs published in late 2025 and the first quarter of 2026. Each row represents a verified finding, not a projected estimate. Where two sources use different methodologies or ask different questions, that is noted in context.

The vertical picture is not random. AI is scaling fastest in industries with dense, repeatable workflows, measurable customer interactions, and high-cost decisions made many times per day. Those conditions make it possible to test agent performance, measure the result, and improve the system quickly.

Manufacturing leads because factory and supply chain operations involve continuous decisions about scheduling, quality, inventory, and maintenance, all of which are trackable. Financial services is close behind because document-heavy processes like loan underwriting, compliance review, and fraud detection involve expensive knowledge work that agents can accelerate. Telecom and retail share similar profiles: high transaction volume, large customer contact surfaces, and digital infrastructure already in place.

Healthcare sits lower on the production scale partly because of regulatory caution around decisions that affect patient safety. But its workflow density, meaning the volume of repeatable clinical and administrative tasks, is massive. As governance standards mature, it may have the largest addressable surface of any sector.

More than half of large enterprises now have AI agents running in production, but governance, meaning the systems for monitoring, approving, and auditing what those agents do, has not kept pace. EY's March 2026 Technology Pulse Poll puts numbers to that gap directly.

97% of technology leaders say autonomous AI is a high priority. Autonomous AI, or agentic AI, refers to software systems that plan and execute multi-step tasks on their own, rather than responding to a human prompt each time. 95% expect AI spending to rise in the coming year. And yet 78% say adoption is outpacing their risk management practices, while 52% of department-level AI initiatives lack formal approval from leadership.

The pattern is consistent with how enterprise software has always spread: individual teams move faster than central governance can track. The difference with AI agents is that these systems can take consequential actions, send communications, make purchasing recommendations, or route customer requests, without a human review at each step.

Deloitte adds one more piece of context: 60% of workers now have access to sanctioned generative AI tools, up from roughly 40% the year before. As the number of users grows and agents begin operating across more workflows, the surface area that needs monitoring gets wider every quarter.

The right frame for governance is not caution as a brake. It is oversight as scale infrastructure. The organizations that can approve, instrument, and when necessary retire agent workflows quickly are the ones that will be able to run a real portfolio of agents rather than a handful of isolated experiments.

The practical move for strategy teams is to change how they evaluate AI. Evaluating at the feature level, meaning asking whether a particular AI tool is useful, captures only a fraction of the available return. Evaluating at the workflow level, meaning asking whether an agent can compress the handoffs, reduce the error rate, and accelerate the decision cycle of an entire business process, captures the compounding value.

The Microsoft and IDC finding about 3.7x average and 10.3x leader returns is the financial frame for this. The organizations near 10x are not using fundamentally different AI. They are using AI differently, placed inside workflows where the decisions repeat, the outcomes are measurable, and the learning loop is short.

2025 was largely about proving that AI copilots, meaning tools that work alongside employees and suggest actions rather than acting independently, could help with individual tasks. 2026 is about proving that networks of agents can produce measurable business outcomes at scale.