A new collection capability came online in the middle of one of my deployments, and the people who fielded it were certain it would change everything inside a week. The hardware deserved the confidence. It could see patterns the older equipment had only inferred. For the first stretch it changed almost nothing, and the reason had nothing to do with the equipment. The watch floor that was supposed to act on what it saw still ran on the rhythm it carried the month before, the same reporting chain and the same unspoken rule about who could wake a decision-maker at three in the morning. The sensor had jumped a generation overnight while the organization around it stayed where it was.

I have watched a quieter version of that scene play out in several companies this spring, on a very different kind of floor. An executive team licenses an AI platform, picks a date by which it expects to see real impact, and then measures progress against how quickly the tool produces output. The tool is fast. The organization holding it works on a different clock, and almost no one names the distance between the two until the deadline arrives and the impact has not.

A report this month from HCLTech, its "AI Impact Imperatives" study, put a number on the consequence. It projected that 43 percent of enterprise AI initiatives are at risk of failing, and the cause it named had little to do with the models themselves. Success, the analysts argued, depends less on how widely a tool gets adopted and more on whether a company can bring its ambition into line with its execution inside the window leadership has set, and whether anyone owns the gap when the two come apart. Workday's researchers said something close in plainer language, that AI can take friction out of a process but cannot make up for fragmented systems and weak operating design.

Two clocks, one deadline

Software runs on a fast clock. You can stand up a model, point it at a data source, and watch it produce something useful in an afternoon. That speed is real, and it is seductive, because it tells a leader the hard part is nearly finished. The hard part has barely started. The value of the tool is gated by the system it sits inside, and that system moves at the speed of human agreement. Workflows have to be redrawn so the output reaches whoever acts on it. The question of which decisions a person still owns has to be reopened, sometimes painfully. And the people who spent years becoming good at the old way have to be brought along, which takes longer than any rollout plan tends to budget for.

A deadline built on the tool's clock asks the organization to produce an outcome on a schedule only the software could ever keep, and then reads the shortfall as a performance problem. The date was wrong the day it was set, pointed at the wrong variable from the start.

A model can be ready in an afternoon. The organization it depends on needs the better part of a year, and the deadline almost always gets written for the model.

What the deadline ends up rewarding

Give a capable team an impossible date and they will not tell you it is impossible. They will find something they can show you by the deadline, and usually that something is activity. Adoption climbs because people are told to log in. Dashboards fill with usage numbers. A pilot gets declared a success on the strength of how many seats were active rather than what changed downstream of them. The metrics look healthy and the business outcome the whole effort was meant to produce sits exactly where it started. Anyone who has run a transformation knows the texture of it. The reporting goes green while the value stays missing.

This is the part that should worry a senior leader more than the headline failure rate. A failed initiative at least announces itself. One that hits every adoption target while producing nothing real can run for a year before anyone admits the impact never came, because each number along the way said things were fine.

The work the timeline quietly skips

Underneath a real AI outcome is a stretch of unglamorous organizational work that no model can do for you. Someone has to decide which calls the tool is now allowed to make and which still belong to a person. The workflow has to be rebuilt so the model's output lands in front of whoever acts on it rather than in a report nobody opens. And the data underneath all of it needs people who understand what it means and who are accountable for keeping it honest. That work is slow because it is human, and it is the actual project. The model was always the easy acquisition.

When leaders compress the timeline, the human work is the first thing cut, because it is the part that refuses to be rushed. The tool gets installed on schedule. The rewiring that would have made the tool matter slides into a vague later that the deadline pretends does not exist.

The companies pulling real value out of AI this year tend to share a quiet discipline. They set the deadline against the organization's clock rather than the software's, and they give the slow human work the room it needs to happen. That is a less satisfying answer than a launch date, and it asks more of the people at the top, because it means owning the timeline honestly instead of borrowing the tool's. If your AI initiative is running behind, it is worth asking whether the work is failing or whether the deadline was simply written for the wrong clock. What would change if you set the date around the organization instead of the tool?