Enterprise AI rollouts stall as worker confusion outruns strategy

Malcolm, an AI engineer at a data-analysis firm, delivered advice his employer did not want to hear. A traditional machine-learning model would handle the task more accurately than generative AI, produce repeatable results, and cost far less. “‘Don’t use AI,’ was his advice,” the BBC reported this month. The company ignored him, deploying a large language model anyway — less accurate, more expensive, but carrying a label the organisation wanted: embracing AI.

The episode captures the shape of enterprise AI adoption as it enters 2026’s messy middle. Companies are pushing staff to use generative AI tools faster than they can define why, where, or to what end. Dan Boyles, CEO of consultancy Hello AI Collective, describes sitting with an oil-and-gas C-suite and asking what ought to be the first question. “What’s the reason for using AI?” None of them could agree. “I just went, ‘what’s the reason for using AI?’ And none of them could agree,” he told the BBC.

The problem has shifted decisively from procurement to management discipline. KPMG has told its US workforce to use AI for 75 per cent of their tasks. Accenture has issued similar mandates. But across the businesses now racing to deploy, the scaffolding is missing. Culture Amp research found that 90 per cent of HR professionals expect to increase generative AI use — yet one in three say no one at their company currently owns the AI strategy. A separate 2026 industry survey, published by Fast Company, found 93 per cent of AI and data leaders identify human issues — culture, change management, workforce readiness — as the primary barrier to adoption, not the technology itself.

What employees are experiencing is a rollout shaped by executive urgency rather than operational design. The UK’s FDA union, which represents civil servants, found that fewer than a third of workers had been consulted on how AI tools could be deployed into their workflows. “Change is being done to workers, not with them,” FDA general secretary Dave Penman told the BBC.

Caroline Rawlinson, CEO of employee-experience platform Culture Amp, warns that the cultural foundation matters as much as the software. “If you’re putting AI technology on top of a fragmented culture or a fear-based culture, it is not going to succeed,” she said. “At best it becomes a very slow roll out as people don’t understand what they’re being asked to achieve. At worst, it ends up as quite a big, wasted effort.”

The stakes are turning tangible. Teradata chief executive Steve McMillan told 5,100 employees in early June that the company would freeze annual salary increases and reallocate the budget toward AI investment. “We will fund this AI investment by reallocating the budget from 2026 annual salary adjustments,” McMillan wrote to staff.

If executive mandates are racing ahead of strategy, the data suggests workers are not waiting for permission. A Harvard Business Review study of the top 100 generative AI use cases found that 63 were work-related — yet none originated as a top-down corporate initiative. Instead, people are adopting AI individually, using it to draft emails, summarise documents, debug code, and prepare presentations. Marc Zao-Sanders, the study’s lead author and co-founder of filtered.com, calls this the “AI in the Wild” pattern: usage that happens at desks every day, invisible to the IT and strategy functions that are supposed to govern it.

The shadow-IT dynamic that defined cloud adoption a decade ago is replaying with AI, and the security exposure is mounting. Okta’s latest survey, reported by The Register, found that more than half of organisations experienced an AI-related security incident or near-miss in the past 12 months. Bosses remain “blinded by confidence” about how much shadow AI their staff are using, the report concluded.

Companies are now scrambling to track what they cannot govern. Business Insider reported this week that JPMorgan, Meta, Amazon and KPMG have all begun monitoring how employees use AI tools, with some feeding the data into performance decisions. Uber, meanwhile, has imposed a monthly spending cap of $US1,500 per employee on AI tokens — a recognition that unmanaged usage has a real cost, and that individual workers are spending far more than managers assumed.

The emerging picture is what Fast Company has labelled trophy-style AI adoption: celebrating usage for its own sake, where activity becomes a substitute for impact. Dashboards show adoption numbers rising and executives cite them on earnings calls, but the underlying work rarely changes.

A compounding inefficiency sits underneath. Most enterprise AI workloads still run on the most expensive frontier models, even for simple tasks: CNBC reported that 95 per cent of enterprise usage flows through top-tier models from OpenAI and Anthropic, when cheaper task-specific alternatives would suffice for many jobs. The parallel with Malcolm’s account is exact — organisations are buying the most conspicuous AI, not the most appropriate one. Zao-Sanders and his HBR co-authors found that the gains from individual AI use were real but typically marginal — faster email drafting, quicker document summaries — rather than the transformational productivity leap that investment cases promise. Moving from marginal efficiency to genuine business change, they argue, requires integrating AI into workflows by design, not by default.

The Australian context sharpens the argument. Commonwealth Bank chief economist Luke Yeaman, writing in the Australian Financial Review, noted that deep AI integration could add 0.8 to 1.0 percentage points to annual labour productivity growth — but the Productivity Commission rates Australia’s likely gain at just 0.4 points. The infrastructure pipeline is enormous: $150 billion in data-centre projects are underway and capacity could triple by 2030. What is missing is the managerial capability to turn that compute into changed work. Planning approvals, competition policy and SME technology adoption are the binding constraints, Yeaman argues — not GPU supply. For Australian CIOs and software teams, the message is that the hardware will arrive long before the organisation is ready to use it.

A counter-narrative is forming at the organisations that got the sequence right. Snowflake’s 2026 summit spotlighted what product teams call an “enterprise context layer”: a governance and data-access framework that sits between AI agents and proprietary information, so workers can use the tools without exposing sensitive data or operating without guardrails. ZDNet Australia reported on firms that paused AI rollouts after discovering long-forgotten data assets suddenly made searchable — a governance problem that would have been catastrophic had the deployment proceeded without the pause. Vanta, the trust-management platform, recently launched an agent designed to surface shadow-AI usage across organisations before it becomes a security incident, a tacit acknowledgment that the control problem now runs ahead of the deployment one.

The common thread in these examples is not more technology. It is governance decisions made before tools are handed to staff, co-design with the people expected to use them, and measurement frameworks that distinguish genuine adoption from dashboard theatre.

Where the next chapter lands — whether organisations build that muscle or keep buying licences and hoping — will determine which half of the Productivity Commission’s 0.4-point estimate actually materialises.

“I think the wreckage is organisations not getting the ROI from it that they were expecting and not getting their people engaging with it.”
— Senior consultant, large consulting firm, to the BBC

The companies that succeed, the evidence from 2026 suggests, will be those that treat AI adoption as a change-management exercise with a technology component, not the other way around. For everyone else, the gap between what the dashboard says and what the workers are doing will keep getting wider.