The Agent Gap: Everyone Is Deploying, Few Are Scaling

Ask three research firms how far enterprise AI agents have got and you will get three answers that are all technically true. Nearly every large company says it is using AI. A minority say they are scaling agents. Almost none can say, in any given function, that agents run at scale. The confusion is not a measurement problem. It is the shape of the market.

McKinsey's most recent State of AI survey puts the pieces in order. Eighty-eight percent of respondents report regular AI use in at least one business function, up from 78% a year earlier. Twenty-three percent say their organisation is scaling an agentic AI system somewhere in the enterprise, and a further 39% have begun experimenting. But in any single business function, no more than 10% report scaling agents. Larger companies are moving faster: the share scaling agents in one or more functions rose from 27% to 40%, while smaller organisations stayed flat at 22%.

Gartner's two forecasts

Gartner has made two predictions that look contradictory and are not. The first: up to 40% of enterprise applications will include integrated task-specific agents by 2026, up from less than 5% in 2025. The second, from a June 2025 release: more than 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value and inadequate risk controls. Gartner analyst Anushree Verma described most current agentic projects as early-stage experiments or proofs of concept, often driven by hype and misapplied, and warned of 'agentwashing', the relabelling of assistants as agents.

Both can be right because embedding an agent in an application is a vendor decision, while getting value from one is a customer decision. Software makers will ship agents into everything. Whether the buyer redesigns the workflow, scopes the permissions and measures the outcome is a separate question, and it is the one that decides cancellation.

The inventory problem

The most revealing statistic is not about adoption at all. SAP's LeanIX Agentic AI Survey 2026 found that 98% of companies have deployed AI agents or plan to, while fewer than half have visibility into an inventory of the agents already running. OutSystems' 2026 State of AI Development report, from a survey of 1,900 IT leaders, found 96% using agents in some capacity and 12% with a centralised platform to manage them.

An organisation that cannot list its agents cannot scope their access, audit their actions or retire the ones that fail. Sprawl compounds. Every quarter of unmanaged deployment makes the eventual inventory harder to build, and it is the reason security teams, not model teams, increasingly decide how fast agents scale.

What production looks like

For a picture of agents at full scale, look at the company building them. In early September 2026 OpenAI published internal metrics alongside an essay by its chief scientist, Jakub Pachocki. The company said it had reached its self-set milestone of an 'automated research intern', a system that completes well-scoped research tasks under human direction, including tasks that would take an experienced researcher several days. Agent runtime inside OpenAI's research organisation had grown to 3.1 agent-workdays for every human workday. Median researchers were spending more than $600 a day on tokens; the top decile exceeded $7,000. More than half of successful four-to-eight-hour tasks still required human intervention.

That is what scaled agent adoption costs and looks like: continuous inference spend, humans in a supervisory loop, and tasks specified tightly enough for a machine to complete. It is also the clearest explanation of why most enterprises are at 10%. Few have the specification discipline, the observability or the budget line.

Who benefits, who is at risk

Beneficiaries: large enterprises with the scale to fund inference and the discipline to inventory agents, systems-of-record vendors that ship agents where the data already lives, and identity and observability vendors selling the control plane. At risk: mid-sized companies stuck at experimentation, vendors relabelling assistants as agents, and any deployment without an owner, a scope and a kill switch.