Most enterprises have bought AI. Far fewer have learned to operate it at scale. That gap, not the technology, keeps promising initiatives from becoming enterprise-wide value.

According to Gartner research published in September 2026, only 22% of organizations have successfully scaled AI across multiple business units or adopted an AI-first approach, even as 85% of functional leaders planned to increase AI spending in 2026. The challenge extends beyond model capability. In practice, dashboards show green while the P&L stays flat and agents that pass every test stumble in production.

The difference is the operating model: who owns the outcome, what value is expected and how readiness, risk and reuse steer investment. One European fintech's AI assistant handled two-thirds of its customer service chats in its first month, yet the company's CEO later acknowledged that cost had been “a too predominant evaluation factor.” In my view, this illustrates how an AI assistant can perform effectively at scale, while the operating model around it fails to balance cost savings with service quality.

Where AI ambitions run into reality

Across the industry, I see recurring dream-breakers in three areas: value, control and scale. None of them shows up in a demo.

Value is lost when teams measure activity rather than outcomes. Green dashboards, closed tickets and “live” agents can coexist with no financial benefit, especially when the old approvals and exception queues stay. That is why a business outcome with a named owner, judged by experts, must be part of the criteria for releasing AI into production.

Control is the second area. Guardrails are essential, but when designed only as prohibitions, they block useful work or bury it in approvals. Decision rights matter just as much: when AI, legacy rules and people all decide the same case, nobody owns the outcome. A high test pass rate can hide failures on real user phrasing when hard-coded rules override the model's interpretation. My rule: systems own facts and limits; AI owns judgment within them.

Then come the constraints of scale. Enterprise data can fail silently and AI turns those errors into actions. Gateway quotas, timeouts and costs often decide outcomes before the model leaderboard does: a larger model does not automatically do better if timeouts and limits were tuned for a smaller one. Fix the plumbing before comparing models.

Governance before autonomy

The first priority is to link AI to value through clear evaluation and release criteria. Treat evaluation as a governed product, with a fixed quality bar and a pinned AI grader, rather than targets that keep changing during development. Moving targets are where agents start to run wild.

That foundation makes autonomy safe. I think of it as an autonomy ladder: assist, recommend, act with approval, act within limits. Agents climb one rung at a time, per decision, only when the controls for that rung exist. Autonomy is earned, not installed.

Governance should enable useful work safely. If every routine action waits for a person, the benefits of agent speed disappear. Equally, removing oversight without reliable evaluation creates risks that only surface after deployment.

Designing those boundaries early is far cheaper than changing requirements once agents are at work.

Redesign the task, including the human role

Software development shows how work itself is changing. Previously, developers searched Stack Overflow, found examples and wrote code manually. Today, an AI coding agent can generate thousands of lines in minutes.

The bottleneck has moved from writing to judging. An AI reviewer may help, but it can also overlook flaws. People still need the experience to challenge what both agents produce.

I have seen this firsthand. An AI agent once proposed a far heavier change than the task needed. When I questioned its plan and suggested a simpler route, it revised the approach and the estimated effort dropped dramatically.

That is task-level redesign in practice: agents generate, people set the standard, challenge the plan and own the decision to stop. Oversight capacity must be planned like any other capacity. Otherwise, faster generation merely creates a growing queue of work for people to check.

Build the foundation, then scale

Every enterprise has different architectures, data and constraints, yet the build order is consistent: the value case and its owner, then data and evaluation, then the AI spine of gateway, quotas, observability and cost control and only then wider autonomy. Platform limits should be understood before deployment, not discovered through rework.

To find what limits your program, score it against the dream-breakers; the lowest score is your cap. Weak on value and more spend is premature. Weak on control and more autonomy is premature. Weak on oversight and more agents will slow you down. At HCLTech, we work through this side by side with technical teams, on their real infrastructure.

This November, more than 6,500 CIOs and industry leaders will come together at Gartner IT Symposium/Xpo™ 2026 in Barcelona to explore the priorities shaping the future of enterprise technology. I look forward to contributing to that conversation by examining the eight dream-breakers that recur across enterprise AI programs and sharing how organizations can identify what's holding them back from achieving measurable business value.

One question to take back to your team: if agents produced ten times more work next quarter, who would review it and who could say stop? AI is now part of the foundation. The operating model is what you build on it.

Join my session, HCLTech: Building the AI Enterprise — An Operating Model for AI at Scale, on Monday, November 9 at 3:35 PM CET.

Register for the session or learn more about HCLTech at Gartner IT Symposium/Xpo™ 2026.