AI agent adoption is widespread at the experimentation level but limited at production scale. McKinsey found 23% of respondents said their organizations were scaling at least one agentic AI system and another 39% were experimenting. Yet no individual business function had more than 10% of respondents reporting scaled agents, and official firm data show even general AI integration remains narrow.
AI
AI agent adoption is widespread at the experimentation level but limited at production scale. McKinsey found 23% of respondents said their organizations were scaling at least one agentic AI system and another 39% were experimenting. Yet no individual business function had more than 10% of respondents reporting scaled agents, and official firm data show even general AI integration remains narrow.
Key findings
• In McKinsey’s mid-2025 survey, 23% reported scaling an agentic AI system somewhere in the organization and 39% reported experimentation.
• “Somewhere” is the critical denominator. No individual function exceeded 10% scaled agent adoption.
• Stanford’s 2026 AI Index characterized agent deployment as single-digit across nearly all functions while generative AI use in at least one function reached 70% of surveyed organizations.
• US Census research found only 18% of firms used any AI in a business function in the November 2025 to January 2026 reference period. Among adopters, 57% used it in three or fewer functions.
• Production adoption should be measured by workflow coverage, permissions, successful completion, human review, and outcome reliability, not by access to an agent product.
Methodology and definitions
This report compares agent-specific survey evidence with official evidence about the broader AI environment in which agents are deployed. The sources are not combined into a single market share.
McKinsey’s 2025 State of AI survey defines AI agents as systems based on foundation models that can act in the real world and plan and execute multiple steps in a workflow. The underlying survey included 1,993 participants from June 25 to July 29, 2025.
The US Census Bureau’s 2026 AI diffusion working paper is nationally representative of employer firms, but it measures any AI across functions and worker tasks, not agents specifically. We use it as a ceiling and context for how deeply AI is integrated.
The UK Office for National Statistics measures use of defined AI technologies and the extent of adoption. Anthropic analyzes privacy-preserved use of one vendor’s consumer and API products. These sources illuminate depth and interaction mode, but they are not counts of agent deployments.
For this report:
• Assistant responds to a user and generally waits for the next instruction.
• Agentic workflow can select or execute multiple steps toward a defined goal, often using tools or systems.
• Experimenting means testing without dependable production coverage.
• Piloting means a bounded real-world use case with active evaluation.
• Scaling means deployment and adoption are expanding somewhere in the organization.
• Fully operational means a named workflow runs under production permissions, controls, monitoring, and service expectations.
These stages should not be collapsed into “adopted.”
The 23% scaling and 39% experimentation figures
McKinsey reported that 23% of respondents said their organizations were scaling an agentic AI system somewhere in the enterprise. Another 39% said they had begun experimenting with agents. Combined, that means 62% reported at least agent experimentation.
That is strong evidence that agent evaluation became mainstream among the organizations represented in the survey. It is not evidence that 62% of companies have autonomous production agents.
The more precise finding appears in McKinsey’s full report: no more than 10% of respondents reported scaling agents in any individual business function. Agent use was most commonly reported in IT and knowledge management. Many organizations that had started scaling did so in only one or two functions.
This difference between an organization-level yes and function-level deployment is the central adoption gap. A company with one expanding service-desk agent belongs in the 23%, even if every other workflow remains manual.
Stanford’s conclusion: agents remain early
Stanford’s 2026 AI Index economy chapter reports organizational AI adoption at 88% and generative AI use in at least one business function at 70%, while describing agent deployment as single-digit across nearly all business functions.
The contrast matters more than the absolute percentages. Generative tools can be adopted by individuals with a login. Agents require dependable tool access, structured context, permission boundaries, exception handling, logs, and process ownership. That makes the transition from assistant to agent an operational change, not just a model upgrade.
Stanford’s agent adoption summary relies in part on organizational survey data also represented in McKinsey’s report. It should be treated as authoritative synthesis, not as a second independent survey confirming the same number.
What official firm data say about the deployment environment
The Census diffusion paper found that 18% of firms used AI in a business function during the November 2025 to January 2026 reference period. Employment-weighting raised that share to 32%, showing that workers are more likely than firms to be located in an adopting business because larger firms adopt more.
Among AI-using firms, 57% integrated AI in three or fewer business functions. The leading functions were sales and marketing at 52%, strategy and business development at 45%, and IT at 41%. At the worker-task layer, 65% of firms limited AI to three or fewer tasks.
These are not agent statistics. They show that the base layer for agent adoption is still narrow in many firms. It is not plausible to infer universal agent scale while most general AI adopters report limited functional and task breadth.
The paper also found that 66% of AI-using firms relied on AI solely to augment tasks. AI-related employment decreases occurred in 2% of firms. That suggests the observed operating model was still predominantly human-led rather than autonomous.
Product telemetry shows modes can shift
Anthropic’s January 2026 Economic Index analyzed a sample of Claude activity from November 2025. On Claude.ai, augmentation represented 52% of conversations and automation 45%. In August 2025, automation had briefly led augmentation, showing that interaction modes move with product features, models, and user behavior.
Vendor telemetry is useful because it observes behavior rather than survey memory. It is also narrow: Claude users and first-party API customers are not the full economy, and an automated conversation is not necessarily a production agent connected to a business system.
A production adoption scorecard
SMBs should count an agent as operational only when all six conditions are true:
1. Named workflow: The trigger, goal, inputs, outputs, and owner are documented.
2. Bounded permissions: The agent can access and change only what the workflow requires.
3. Measured completion: Success and failure can be observed at the workflow level.
4. Human review rule: High-impact, ambiguous, or low-confidence actions stop for approval.
5. Auditability: Tool calls, source context, changes, and exceptions are logged.
6. Business outcome: Cycle time, quality, cost, revenue, or service performance is compared with a baseline.
This avoids conflating an agent demo, a chat interface, and a governed production workflow. The distinction between an AI operator and an AI agent is relevant here: the agent performs work, while the operator owns the workflow, controls, evaluation, and improvement around it.
For a first deployment, use read-only research, internal summarization, or draft creation before permitting external messages, record changes, refunds, approvals, or financial actions. Expansion should happen action by action, not through a blanket autonomy setting.
Limitations
There is no official worldwide agent census. Definitions vary across vendors and surveys. Some products labeled agents are assistants with tool calls; some workflow automations include agentic decisions without that label.
McKinsey’s percentages are self-reported by survey participants and may overrepresent organizations engaged with AI. Census offers a stronger firm denominator but does not isolate agents. Anthropic observes one product ecosystem. Stanford synthesizes multiple sources, including some already cited here.
What would change our view
We would conclude that agents had moved from experimentation to broad scale if representative data showed sustained function-level deployment, recurring multi-step task completion, wider employee coverage, stable exception rates, and measured business outcomes. We would become more cautious if scaling claims grew without corresponding workflow breadth, permissions, logs, or successful production runs.
FAQs
What percentage of companies use AI agents in 2026?
No representative global percentage exists. McKinsey found 23% of respondents reported scaling at least one agentic system somewhere in their organization, but no individual function exceeded 10% scaled adoption.
Is experimenting with an AI agent the same as adoption?
No. Experimentation proves that a team can test the technology. Production adoption requires a repeatable workflow, permissions, evaluation, human review, monitoring, and an accountable owner.
Which departments use AI agents most?
McKinsey found agent use most commonly reported in IT and knowledge management. This does not mean those functions have majority adoption; function-level scaled use remained at or below 10%.
Why are agents harder to scale than chatbots?
Agents act across systems and multiple steps. That introduces permission, data, reliability, exception, audit, and process-change requirements that a standalone chatbot may not have.
How should an SMB start with an AI agent?
Choose one bounded workflow, begin with read-only or draft-only permissions, evaluate real cases, require human approval for high-impact actions, and expand only after quality and value are stable.
Get a 20-Minute AI Workflow Audit
AI Operator can assess whether your agent is a demo, pilot, or production workflow and define the controls and evidence needed for the next stage.