Most AI post-mortems blame the model: not capable enough, too expensive, hallucinated. The data points somewhere else. MIT's 2025 State of AI in Business found that roughly 95% of enterprise generative AI efforts had produced no measurable return on the P&L yet, and the efforts that stalled were rarely undone by weak technology. McKinsey's 2025 State of AI ties realized EBIT impact to one thing above all: redesigning the workflow around the tool. Only about 21% of organizations had done that.
In other words, the technology mostly works. The organization around it does not change to absorb it. Four patterns account for most of the stalls.
The most common failure is the quietest: teams point AI at the process exactly as it exists today, workarounds and all. Automating a broken process gives you a faster broken process. The pilot looks impressive in the demo and produces nothing the business can use, because the work was never sound to begin with.
The fix is sequence. Map how the work actually moves, not the documented version, the lived one, then redesign it, then automate. This is why we start every engagement with the operating model rather than the tool. The model is the last thing you add, not the first.
The second failure hides until late. A workflow assumes data that turns out to be scattered across systems, inconsistent, or locked inside PDFs and email. The prototype demos beautifully on a clean sample and then dies the moment it meets production data.
The fix is to verify, not assume. Before committing to a build, confirm the data exists, is reachable, and is good enough, and treat the data layer as part of the work rather than a precondition someone else will handle. Most stalls we are asked to rescue trace back to data that was assumed into the plan.
The third failure is the one consultancies skip. The prototype works, and nobody changes how they work. There is no training, no new standard operating procedure, no decision about which exceptions go to a person. The tool sits beside the old process instead of replacing it, and within a quarter usage drifts to zero.
The fix is to treat adoption as part of the build, not an afterthought. Communications, training, and governance belong in the engagement from day one, with people clearly assigned to the judgment and the exceptions the AI hands them. EY's Q4 2025 AI Pulse found that 62% of private-equity leaders struggle to link productivity gains to AI; a large part of that gap is adoption that was never engineered.
The fourth failure is about attention. A senior sponsor shows up for the kickoff and is gone by week six. Without an owner who cares about a specific number, the initiative loses oxygen the first time it competes with something urgent, which is immediately.
The fix is to tie the work to a lever the sponsor already owns, and to keep the cadence tight enough that drift is visible: weekly working sessions, a clear metric, and an early go/no-go that forces a decision while the sponsor is still in the room.
None of these failures are exotic. They are predictable, which means you can design against them up front:
- Pick one workflow with a named operating lever and a metric, not a portfolio of ambitions.
- Redesign before you automate, so you are not codifying a broken process.
- Verify the data exists and is reachable before the build, not during it.
- Engineer adoption from day one, with humans on the exceptions.
- Keep a sponsor accountable to the number, with a checkpoint to confirm, rescope, or stop.
- Match the architecture to the data: private, hybrid, or external, so sensitive data is never the reason the work stalls.
That is the shape of the First 5 Weeks. It is deliberately built around the places initiatives fail: one workflow, a real lever, verified data, adoption designed in, and a week-1 checkpoint that confirms feasibility before any build work begins. The goal is not a demo that impresses. It is a workflow that survives contact with the organization.
Sources: MIT, The State of AI in Business (2025); McKinsey, The State of AI (2025); EY, Private Equity AI Pulse (Q4 2025). Figures are self-reported by the cited studies and describe the market, not WCG client results.