the real reason enterprise AI keeps stalling

a product team gets mandate and budget to bring ai into their workflow. they pick a vendor, scope a pilot, and run it in a controlled environment with clean data and a motivated squad. the demo lands. leadership is impressed.

then someone asks who owns the rollout, and the conversation stalls. 6 months later the pilot is still a pilot, a new vendor is under evaluation, and the team is demoralized for a specific reason: the technology worked, and nothing they built reached the people it was supposed to help.

mckinsey¹ calls this pattern "pilot purgatory." in their 2025 state of ai research, covering nearly 2,000 organizations across 105 countries, nearly two-thirds of enterprises remained stuck in experiment or pilot mode. only around 6% were seeing ai move the needle on enterprise-level outcomes. the gap sits in execution infrastructure: the organizational conditions that let a proof-of-concept become a production system. those conditions are almost never technical.

google's DORA team² has studied high-performing software delivery organizations for a decade. their 2025 report, drawing on nearly 5,000 technology professionals worldwide, put the same finding in plain terms: ai amplifies what a team already does. teams with clear ownership, stable priorities, and workflows built to ship use it to move faster. teams with unclear decision rights and fragmented handoffs find that ai surfaces those problems sooner.

before a team can benefit from ai at scale, the work system has to hold: process, roles, rhythm, and the human infrastructure that connects strategy to something in production. most organizations buy the tools first and try to retrofit the structure later. that is why the demos stay better than the deployments.

the workflow redesign gap

mckinsey identified one factor that most distinguishes high-performing ai organizations from everyone else: they redesigned their workflows, rather than layering ai on top of existing processes. only 21% have done this.

inside regulated industries (pharma, fintech, financial services) the existing processes were built for compliance, not for speed. MLR review, change control, and approval chains create friction no ai tool can shortcut. the teams making progress are redesigning the work that sits around that layer (the brief, the handoff, the iteration loop) so what reaches review is already tighter. that is a product design problem. it needs someone who understands both the regulatory environment and how a high-functioning product team actually operates. that intersection is rarer than it should be.

what it looks like to get unstuck

nielsen norman group³ put it bluntly at the start of 2025: "we're seeing a massive regression of the average UX maturity in organizations (and that average wasn't very high to begin with)." it is an assertion from practice, not a tracked dataset. it also matches what we see.

organizations that should be getting better at building user-centered products are, under cost pressure and ai hype, moving backward. design gets compressed. research gets cut. products get built for approval, not adoption.

the teams that break the pattern are not always the ones with the most resources. they keep asking, even when the answers are inconvenient: who is this for, what has to be true for them to actually use it, and what the work needs to look like on a tuesday rather than in a demo. that discipline, more than another pilot plan, is what separates organizations in production from the ones still scheduling the next experiment.