From the outside, enterprise AI in mid-2026 still looks like an expansion story. Microsoft shipped Scout. Salesforce launched Agentforce. Google and IBM just announced a multi-billion dollar agent partnership. Inside enterprises, the math is less accommodating.

Gartner analyst Anushree Verma told IT Pro that while interest in scaling copilots and agents into production has accelerated, “the deployments that do not show a tangible value will be difficult to sustain.” Most organizations completed at least one AI proof-of-concept by the end of 2025. The question in 2026 is which of those survive financial review, and the emerging answer is: fewer than the vendor slides suggest.

The Numbers Behind the Correction

Three independent data points converge on the same conclusion. MIT’s NANDA initiative found that 95 percent of corporate generative AI pilots produced zero measurable ROI despite an estimated $30 to $40 billion in enterprise spend. Sinch’s “AI Production Paradox” report showed 74 percent of enterprises have already pulled live AI customer communications agents back out of production. And BetterUp Labs, working with Stanford researchers, calculated that AI-generated “workslop” drains roughly $9 million per year from a 10,000-person organization.

These are not edge cases. They describe the median enterprise AI experience.

The Per-Seat Copilot Problem

The clearest pressure point is the broad, per-seat copilot. MetrixData 360 reports that organizations are actively reducing their Microsoft Copilot seat counts at renewal. Some are redirecting AI budgets toward OpenAI or Anthropic APIs. One organization absorbed over $500,000 in unexpected AI spend during a single development cycle because consumption-based costs scaled without adequate monitoring.

The structural issue, as Verma frames it: “Use cases with copilots are not leading to a tangible ROI. That kind of a use case gives a return on employees, while a workflow-specific AI tool gives an ROI.” Summarizing meetings and drafting documents generates diffuse productivity gains that satisfy individual users but evaporate under CFO scrutiny. A copilot that saves each employee 20 minutes per day produces anecdotes. A workflow-specific agent that cuts claims processing time by 40 percent produces a line item.

Why Agents Raise the Stakes

Agents compound the problem because they promise more than assistance. A copilot drafts and suggests. An agent completes tasks, triggers actions, and moves work across systems with reduced human oversight. The potential value is higher, but so is the cost of poor design.

Gartner projects that by 2028, an average Fortune 500 enterprise will have over 150,000 agents in use, up from fewer than 15 in 2025, according to IT Pro. That trajectory creates agent sprawl: hundreds or thousands of small AI systems operating across departments, each with its own permissions, data access, ownership, and failure modes. At that scale, rollback becomes IT hygiene.

Verma warns that “all the siloed, ad hoc task-specific agents are leading to an agent sprawl” that demands careful management. The enterprises pulling back are not abandoning AI. They are pruning, consolidating around high-value use cases, and killing deployments that cannot answer the CFO’s question.

What Survives the Correction

The pattern emerging from every data source points in the same direction: workflow-specific agents with clear before-and-after metrics outlast general-purpose copilots with diffuse benefits. Service management, compliance review, software testing, claims processing: these have measurable baselines. “Helps employees be more productive” does not.

Verma’s framework separates surviving AI projects into three categories: defensive use cases improving efficiency, extension use cases supporting growth, and disruptive bets aimed at longer-term innovation. The first two justify continued spend. The third requires explicit executive sponsorship and a tolerance for ambiguity that most finance teams no longer extend to AI projects.

For agent builders, the implication is direct. The market is not contracting. It is selecting. Vendors who can demonstrate per-task or per-workflow ROI in the first 90 days of deployment will absorb the budgets being freed up by copilot consolidation. The question enterprises are asking has shifted from “should we deploy agents?” to “which agents earn the right to stay?”