> ## Content Index > Fetch the complete content index at: https://aiadoption.org/llms.txt > Use this file to discover other available public pages before exploring further. # Agent deployment is still single digits - URL: https://aiadoption.org/ai-analysis/agent-deployment-is-still-single-digits/ - Published: 2026-09-25T03:12:57.000Z - Updated: 2026-09-25T03:12:57.000Z - Description: Agent benchmarks soared in 2025 (OSWorld 12%→66.3%, GAIA 20%→74.5%), but McKinsey shows scaled agent deployment stuck in single digits across most business functions. Capability and deployment have diverged, and the integration layer is the bottleneck. - Author: Jassie - Tags: AI Analysis, AI Economics, AI Automation, AI Strategy If your strategic plan reads the rapid agent capability advances of 2024–2025 (OSWorld 12%→66.3%, GAIA 20%→74.5%) as evidence that agent deployment is correspondingly widespread, the McKinsey deployment data resets the picture. Scaled agent use is in single-digit percentages across most business functions in 2025\. The capability and the deployment have diverged. The data shape. McKinsey asks organisations about three deployment stages: piloting, deploying at scale, and not using. The data reports the share of organisations at the "deploying at scale" stage by business function. The pattern: experimentation is widespread, scaled deployment is rare. Most functions show double-digit percentages of organisations piloting agentic AI. The same functions show single-digit percentages of organisations deploying at scale. The gap between pilot and scale is large and consistent across functions. Specific function breakdowns: marketing and sales, customer service, IT, and HR all show notable pilot activity. The scaled-deployment rates for the same functions are low. The software engineering function shows somewhat higher scaled deployment, consistent with the GitHub Copilot deployment patterns (#74 Productivity gains are real but uneven). Most other functions show scaled deployment in the 2–7% range. The methodological observation that follows: the agent capability narrative and the agent deployment narrative are measuring different things, and they are increasingly disconnected. The capability narrative (Chapter 2; see Article #45 The agent capability jump) shows agent benchmarks rising rapidly. Models can now navigate operating systems, complete complex multi-step tasks, and work with diverse tools. The benchmarks are real and the capability is real. The deployment narrative (Chapter 4, this article) shows scaled deployment lagging the capability by a substantial margin. The capability is available; the deployment infrastructure, governance, security model, and organisational readiness to absorb agentic AI at scale are not. ## Several measurement issues complicate the picture. **First, "scaled deployment" is defined differently by different organisations.** An organisation with one agent handling 20% of a function's volume might describe this as "scaled" or as "still in pilot." The survey methodology relies on self-reporting, which produces variance in interpretation. **Second, agent deployment is often hidden inside other deployment categories.** An enterprise software vendor that uses agents to handle customer queries internally would not necessarily report this as "agent deployment": the function might be reported as customer service, with the agent layer subsumed into the workflow. The "agent deployment is single digits" finding may underestimate the actual agent usage if these embedded deployments are systematically miscounted. **Third, the survey measures organisations, not agents.** An organisation that has deployed one critical agent at scale shows as one data point, equivalent to an organisation with 50 agents. The "share of organisations" metric does not track the depth or breadth of deployment within those organisations. These methodological issues affect the absolute numbers but probably not the direction. Even with generous interpretations, agent deployment at scale across business functions remains a minority of organisations. The capability is widespread, the deployment is not. ## For strategic planning, three implications follow. **The first: planning frameworks that assume "capability is available, so deployment will follow rapidly" need adjustment.** The 2024–2025 trajectory shows capability and deployment diverging, not converging. The reasons for the divergence, security and risk concerns (#60 Why agentic AI hits a security wall), RAI tooling gaps, and regulatory uncertainty, are structural and slow to resolve. Plans that assumed deployment would track capability with a 12-month lag are operating against data showing the lag is 24+ months and may extend further. **The second: organisations that successfully bridge the capability-to-deployment gap have structural advantages.** The data shows most organisations are not bridging it yet. The few that do (the \~5–10% deploying at scale in the leading functions) will accumulate experience, infrastructure, and competitive advantage that the lagging majority will struggle to replicate. The strategic question for executive teams is whether to be in the small group bridging the gap now or to defer until the operational patterns are more established. **The third: the deployment bottleneck is not capability.** It is the integration layer underneath. The agents themselves work. What is missing in most organisations: authentication and authorisation frameworks designed for agent-as-actor (rather than human-as-actor); audit logging that captures agent actions at the granularity needed for compliance review; reversibility infrastructure for the actions agents take; integration with existing identity and access management systems; and incident response procedures that include agent-caused incidents. Building this integration layer takes 12–24 months and substantial investment. Organisations that are bridging the gap are investing in this infrastructure; organisations that are not are blocked by its absence. > The distinction underneath all of this: capability benchmarks are useful measurements but they are not deployment indicators. This distinction is clearer now than it was in prior years. Strategic planning that conflates the two will produce overestimates of where deployment actually is and underestimates of what is required to bridge the gap. Strategic planning that distinguishes them, and budgets for the integration layer separately from capability evaluation, will produce more accurate plans and better deployment outcomes. For executives setting AI direction, the planning anchor is that agent capability is now substantial but agent deployment infrastructure is not yet built in most organisations. The strategic choice is whether to invest in building it (high cost, multi-year horizon, structural competitive advantage if successful) or to defer (low cost short-term, accept a position in the trailing majority). The data does not say which choice is correct for any specific organisation. It does say that the question is now real and pressing in a way it was not in 2024. --- ### Sources - **Primary**: Stanford AI Index 2026, Chapter 4 (Economy) 4.3 — [hai.stanford.edu/ai-index/2026](http://hai.stanford.edu/ai-index/2026?ref=aiadoption.org) - **Survey data**: McKinsey & Company "State of AI" Survey, 2025 — agentic AI deployment by function and stage - **Productivity-evidence cross-reference**: Brynjolfsson, Cui, Ju & Aral, Becker et al. — software-engineering and customer-service productivity studies, 2024–2025