Enterprises winning with AI agents are limiting how much the agents can do alone

What happened

New data on enterprise AI agent deployments suggests that unrestricted autonomy is proving costly rather than beneficial. According to a Gartner forecast reported by VentureBeat, more than 40% of current agentic AI projects will be discontinued before 2028 — not because the underlying models underperform, but due to spiralling costs, unclear return on investment, and insufficient risk controls.

The reporting also points to McKinsey's 2026 AI Trust Maturity Survey as further evidence that organisations are reassessing how much independent decision-making they allow their AI agents. Rather than building agents that plan and act across broad, multi-step workflows with minimal oversight, the enterprises reportedly seeing better results are the ones narrowing agent responsibilities and enforcing clear operating boundaries.

This marks a shift from the dominant assumption of the past two years — that giving agents more autonomy would naturally translate into better performance. In practice, many production deployments are reportedly running into governance, cost and trust problems that outweigh the efficiency gains of open-ended autonomy.

Why it matters

For technology and transformation leaders, this reframes the agentic AI conversation from "how much can we automate" to "how tightly should we scope what we automate." An agent with a narrow, well-defined mandate is easier to monitor, audit and correct — all of which matter when agents are making decisions that touch customers, employees, finance or compliance. The Gartner projection implies that a large share of current agentic AI investment is structurally fragile, regardless of model quality, because it was designed around unconstrained autonomy rather than accountable, bounded action.

This has direct implications for anyone deploying AI in service, operations or support functions. Agents that operate within explicit rules are more predictable for both the business and the people who rely on them — a property that matters as much for trust and adoption as it does for cost control.

By the numbers

• More than 40% of agentic AI projects running today are forecast by Gartner not to survive to 2028, citing rising costs, unclear business value and weak risk controls.

The Renascence take

The industry spent two years treating autonomy as the finish line for agentic AI. What's emerging instead is that autonomy without boundaries is a governance problem wearing an innovation costume.

Unbounded agents fail for the same reason unbounded human processes fail: nobody can trust, audit or improve a system whose scope keeps shifting. The lesson isn't to slow AI adoption — it's to design agents the way good service design has always worked, with clear responsibilities, explicit escalation points and visible limits. Operators chasing autonomy for its own sake are optimising for a demo, not a durable operating model; the ones scoping agents tightly are the ones building something a customer, employee or regulator can actually rely on.