What enterprises are actually testing with AI and agents now
Enterprise AI has moved beyond isolated chat experiments. But adoption statistics conceal several different realities: embedded assistance, bounded workflow automation, agent pilots, production deployments and genuine operating-model redesign are not the same state.
Adoption is broad. Depth is uneven.
Recent enterprise surveys point in the same direction despite different samples and definitions. AI use is widespread; agent experimentation is substantial; production deployment is materially narrower; enterprise-wide redesign is narrower again.
McKinsey’s 2025 global AI survey reports that 23% of respondents say their organizations are scaling an agentic AI system somewhere in the enterprise and another 39% are experimenting. In any individual business function, no more than 10% report scaling agents.
Deloitte’s 2025 emerging-technology study gives a similarly layered picture: 30% of surveyed organizations were exploring agentic options and 38% piloting, while 14% had solutions ready to deploy and 11% were actively using them in production. These measures are not directly interchangeable with McKinsey’s, but both show the same structural gap between interest, experimentation and operational depth.
Deloitte’s 2026 State of AI in the Enterprise sharpens that picture. Across 3,235 surveyed leaders, 25% said their organizations had moved at least 40% of AI experiments into production, while only 34% reported genuinely reimagining the business. Agentic use is accelerating faster than governance: only one in five companies reported a mature governance model for autonomous AI agents.
Read left to right as different depths of enterprise adoption, not a maturity score. Source: Deloitte, State of AI in the Enterprise 2026. DENFORTH synthesis.
What work is attracting agents first?
Current deployment patterns concentrate where work is digitally accessible, high-volume, information-intensive and connected to systems that already expose usable data or APIs. PwC’s 2025 US executive survey found current or near-term agent use concentrated in customer service, sales and marketing, and IT/cybersecurity, with lower reported use in supply chain, manufacturing, procurement, legal and tax.
Customer service
Current or planned agent use within six months among PwC respondents using or planning agents.
Sales & marketing
Research, content, account support, CRM interaction and workflow assistance create accessible entry points.
IT & cybersecurity
Ticket handling, coding, diagnostics, operations and security workflows combine digital evidence with executable tools.
The pattern matters more than the ranking. Enterprises are starting where AI can observe digital work, produce or retrieve information, and increasingly take bounded actions. Physical operations and heavily regulated or consequence-sensitive work generally demand more integration, validation and control before comparable autonomy is practical.
Embedded assistance is not operating-model transformation.
PwC’s May 2025 survey of 308 US business executives reported 79% saying AI agents were already being adopted in their companies. Among adopters, 66% reported measurable productivity value. Yet only 45% said they were fundamentally rethinking operating models and 42% redesigning processes around agents.
This distinction explains why apparently high adoption figures can coexist with modest enterprise transformation. An agent embedded in an existing application may save minutes, summarize information or update a record without changing decision rights, role boundaries, process architecture or cross-functional coordination.
The production gap is an architecture problem.
Moving from a successful agent demonstration to dependable operational use requires more than model capability. Enterprise systems have permissions, inconsistent data, legacy interfaces, exception paths, regulatory constraints and consequences. Deloitte identifies legacy-system integration as a major obstacle to production agentic systems; agents still depend heavily on APIs and conventional data pipelines to interact with enterprise applications.
Scaling remains rarer than experimentation.
IBM’s 2025 C-suite study reported that 25% of AI initiatives had delivered expected ROI and 16% had scaled enterprise-wide. PwC’s January 2026 CEO analysis likewise reported only one in eight CEOs seeing both increased revenue and reduced costs from AI. These figures measure different populations and outcomes, but they reinforce a consistent point: access to capable AI is no longer the main differentiator.
The harder work sits around selecting material problems, redesigning workflows, connecting systems, establishing reliable context, allocating authority, changing roles, evaluating outcomes and deciding what deserves to scale.
Agent autonomy is being allocated, not simply switched on.
Current enterprise use spans a spectrum: retrieval and recommendation; drafting and analysis; proposed actions requiring approval; bounded execution with human exception handling; and more autonomous multi-step workflows. High-consequence environments push the design toward tighter permissions, stronger evidence and explicit escalation.
This makes the useful unit of analysis the work system, not the agent. The same model can be safe and valuable in one workflow and unacceptable in another because the surrounding data, authority, consequence and recovery conditions differ.
The emerging pattern
Enterprise AI is moving from individual assistance toward workflow participation. Agents are increasingly being tested as actors that can retrieve context, reason across steps, call tools and change system state. But the leading edge of capability should not be confused with the median enterprise deployment.
The evidence suggests a staged reality: broad experimentation, selective production, narrower scaling and still-narrower organizational redesign. The enterprises that move beyond the first two stages will have to redesign the system around the AI—not merely add AI to the system they already have.
The enterprise question is moving from “Can AI do this?” to “What operating system lets AI do this reliably?”
That changes the transformation agenda. Model selection remains relevant, but workflow, data, integration, permissions, evaluation, Human–AI allocation and operating-model design increasingly determine whether technical capability becomes enterprise capability.