Why Enterprise AI Stalls — and What the 1.7% Are Doing Right

Ninety-seven per cent of CEOs say they plan to integrate AI into their business. One point seven per cent feel ready to do it. That gap is the most honest data point in enterprise technology right now, and it's not a technology gap — it's a leadership, data architecture, and process redesign gap.

The interesting story for technology leaders is not the size of the gap. It's that we now have unusually good research on what closes it. Prosci's 2025 study of 1,107 participants — 525 frontline employees, 393 people leaders, 193 executives — singled out specific executive behaviours that differentiated successful adoption from stalled pilots. IBM's May 6, 2025, study of 2,000 CEOs put numbers on the data and architecture problems sitting underneath. Cutter's CEO Insights 2025 is direct about why most enterprises are stuck in pilot purgatory while a smaller group is already in production. Read together, the picture is consistent.

The four behaviours that move the needle

Prosci's framework is unusually concrete for change-management research. Executive impact on AI adoption sorts into four areas — what they call the four V's: Visibility, Vision, Voice, and Value. Leadership clarity on these four differentiated successful from unsuccessful adoption by +1.65 points on a standard maturity scale. The same study showed adoption initiatives suffering trust gaps of +1.09 and ease-of-use gaps of +1.19 when these behaviours were absent.

In plain terms, the executive actions that compound:

Visibility. Executives are visibly using the tools themselves — model interactions, agent outputs, dashboards. Not in a quarterly demo. In their actual working day. The behavioural signal to the rest of the organisation is "this is real, I trust it enough to put my time in it."

Vision. AI is positioned as an enterprise change blueprint, owned by the executive, with a roadmap that ties to core drivers — customer experience, operating margin, competitive resilience — not as a technical backlog owned by IT.

Voice. Executives talk about AI internally with the same regularity they talk about quarterly results. Pilots that succeed have CEOs mentioning specific outcomes by name; pilots that stall have CEOs delegating the conversation entirely to a chief AI officer who is then disempowered.

Value. Initiatives are scoped against measurable outcomes (cycle time, defect rate, conversion lift) rather than activity (pilots completed, models trained). When the executive is the one defining the value metric, the initiative inherits accountability.

The teams shipping autonomous AI systems into production — sales agents, support agents, research agents, internal-ops agents — are almost without exception working inside an organisation where these four behaviours are visible. The teams running stalled pilots usually have one or two of them at most.

Data architecture is the bottleneck

The technology bottleneck most CEOs name is wrong. It isn't the model. It's the data layer.

IBM's CEO study put four numbers in front of the boardroom that everyone in this conversation should know:

• 68% of CEOs say integrated enterprise-wide data architecture is critical to unlock generative-AI value.
• 72% say proprietary data is the differentiator.
• 61% are actively adopting AI agents today and preparing for scale.
• 50% report that their technology investments are disconnected from each other and from rapid AI investment.

That last one is the killer. Half of enterprises are buying frontier models, agent platforms, and copilots while their underlying systems remain fragmented — which means even when the model is right, the agent is operating on partial context. Eighty per cent of the investment goes into the visible model layer; eighty per cent of the failure modes come from the invisible data layer.

This is the part of enterprise AI that maps most directly to what we ship at Axccelerate. Every operational AI workflow automation build runs into the same rate-limiting step — clean data plumbing has to come first. That's why our AI infrastructure work (model routing, eval harnesses, drift detectors, the boring layer) pays for itself many times over against the more visible model-and-prompt layer above it. The CEOs in IBM's study who already have integrated data architectures are the ones running production agents. The ones who don't are running pilots that look impressive in a deck and then quietly fail when they meet real workflow.

Proprietary data is the second leg of the same point. Public-model capability is increasingly commoditised — within a quarter or two, every frontier lab catches up on every other lab's headline benchmark. The durable advantage for an enterprise is the combination of a frontier model with proprietary, well-instrumented, well-labelled data the competition cannot replicate. CEOs who delegate data strategy to the IT function are usually surprised, twelve months later, that their AI initiatives don't compound.

Pilots are not the strategy

Cutter's CEO Insights 2025 frames the stall directly. There are now over 900 AI use cases in active development across enterprise engineering organisations per the ADL benchmarking report — but the gap between cases-in-development and cases-in-production is enormous, and most large companies are in what Cutter calls a "proof-of-concept trap." A pilot succeeds in isolation; the executive sponsor takes it as evidence the approach works; the investment splits across more pilots; nothing crosses the production threshold.

The shift Cutter advocates is from pilot orchestration to what they call Wave 3 disruption — replacing entire workflows, not augmenting individual tasks. Wave 1 was AI-assisted productivity (copilots in the IDE, chat in the docs app). Wave 2 was specific automated tasks (summarisation, classification, retrieval). Wave 3 is process redesign — taking a multi-step workflow that previously needed seven hand-offs and rebuilding it as one agentic loop with humans only at decision gates.

The startups Cutter flags as outpacing incumbents are not winning on raw model quality. They're winning because they were built around Wave 3 from day one — no legacy process to retrofit, no twelve-year-old data warehouse to integrate against, no internal politics around which department owns the redesigned workflow. Incumbents that match them are doing it by treating AI not as a feature roll-out but as a workflow rewrite, with the CEO sponsoring the redesign rather than the model selection.

MIT Sloan's emerging-agentic-enterprise framing is consistent. Agentic AI is spreading faster than process redesign can absorb. Without intentional redesign, agents are deployed inside legacy workflows — and the result is a faster version of the same flawed process, not a better outcome.

The CEO as systems architect, not pilot sponsor

Here's the role the research consistently points to. The CEOs who get AI working at enterprise scale are not the ones with the largest AI budget, nor the ones with the most credentialled chief AI officer, nor the ones running the most pilots. They are the ones who treat AI as work transformation — investing in people, data, and process redesign alongside the technology, with executive visibility over all four.

Three concrete moves separate the early movers from the laggards over the next four quarters.

Take ownership of the data architecture decision. This is not an IT call. The CEOs in the 1.7% of "ready" companies in the Gartner / Cisco data are universally the ones who personally backed an integrated data programme one to two years before they started buying agent platforms. If you're considering an agent rollout against a fragmented data estate, the agent will fail — not because the model is wrong, but because the agent is reasoning over an incomplete picture of your business.

Pick one Wave 3 workflow and rebuild it end-to-end. Not "add AI to" the workflow. Replace the workflow with one that assumes agents are part of it. Lead generation, support triage, financial close, employee onboarding, customer research — pick one process where you have clean signal and rebuild it from the agent outward. Treat the rest of the AI portfolio as Wave 1 / Wave 2 productivity gains around the edges, but pour your strategic attention into the Wave 3 case.

Make AI use part of how you operate, not how you communicate. The visibility behaviour Prosci flags is the cheapest of the four V's to get right and the one most often skipped. CEOs who model active tool use — including in front of their executive team — accelerate cultural adoption far more than CEOs who only talk about AI in earnings calls and town halls.

Ninety-seven per cent of executives believe generative AI will transform their industry. Accenture's research found that 93% of AI investments outperform when they're strategic — and quietly underperform when they're not. The difference between the two is almost never the model. It's whether the CEO treats AI as a technology line item or as a fundamental redesign of how the business does work.

The 1.7% know which of those it is. The remaining 95.3% have, on the available evidence, between two and four quarters to figure it out.