The failure is real, and it is measurable
The numbers are consistent across the firms that study this, and they have not improved as the tools have gotten better.
Roughly 70% of digital transformations fail to reach their stated goals. McKinsey and BCG have both landed near that figure for years, and a 2024 review of the evidence put the range at 60 to 85%. The rate has barely moved in two decades of trying.
For AI specifically, the picture is worse. An MIT NANDA study in 2025 found that 95% of enterprise generative-AI pilots delivered no measurable profit-and-loss impact. Gartner projected that 60% of AI projects would be abandoned through 2026 for lack of AI-ready data, escalating costs, and unclear business value. And even where companies have started, S&P Global found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year, 2025.
The pattern underneath is the same one McKinsey describes for transformations generally: companies run about eight digital projects on average and scale fewer than a third of them. Activity is not the same as adoption, and a pilot is not a result.
Why they fail: six root causes
Read across McKinsey, Gartner, MIT, and Prosci and the same causes recur. None of them is a software problem.
No adoption. The single strongest predictor of whether a project meets its objectives is the quality of change management around it. Prosci's longitudinal data is unambiguous: projects with excellent change management meet their objectives 88% of the time, against 13% for those with poor change management. That is not a rounding error. It is the difference between a project that works and one that does not, and it has almost nothing to do with the tool.
No business outcome. Initiatives get framed as "modernize" or "become data-driven" without a number anyone owns. Prosci finds that projects with an effective executive sponsor meet objectives 79% of the time versus 27% without one. When nobody's budget and reputation ride on the result, the result does not survive contact with the next planning cycle.
Bad data. You cannot build trustworthy analytics, let alone automation, on records that have drifted from reality. Gartner's figure is the blunt one: 60% of AI projects abandoned because the data was not ready. There is a useful rule of thumb from data-quality practice, the 1-10-100 rule: it costs $1 to verify a record as it is captured, $10 to correct it later, and $100 or more to do nothing and let the error propagate downstream. Most plants are paying the $100, continuously, without seeing the invoice.
Pilot graveyard. A successful pilot that depends on a hero, a vendor, or a one-off configuration never becomes a program. This is McKinsey's "fewer than a third get scaled," and it is the trap that flatters a team right up until it strands them.
No change process. Even a good deployment decays if the organization has no mechanism to keep procedures, training, and records current, and no one checking whether people actually adopted the new way. MIT's work on pilots flags the absence of a feedback loop after launch as a defining failure mode.
Wrong starting point. Companies stall for years "evaluating AI," or they pick a first project so integration-heavy that it never returns value. McKinsey names the first failure mode directly: analysis paralysis is one of the top reasons manufacturers fail to move. The instinct to wait until the plan is perfect is itself the thing that prevents a result.
The mechanism: what actually breaks
These six causes are not a list to nod at. Each one is a specific, testable property of an organization, and each one breaks a transformation in a predictable place. That is what a readiness assessment should measure, one dimension at a time.
Adoption is whether the people who must use a system will actually use it once the launch fades. It is the strongest single predictor of outcome in the change-management literature, and the most common place a project dies quietly. The tell is simple. If the people who will live with the tool were not involved in shaping it, and if it gives them nothing back, they will route around it. They always have.
Scalability is whether a thing that works once can be repeated without the original heroes. A pilot that cannot be described as a repeatable playbook is not an early success. It is a dead end you have not reached yet. The test is whether the rollout survives the departure of the two people who understand it.
Business-goal mapping is whether the effort connects to a metric a real person owns, with money and reputation attached. "Efficiency" is not a goal. A named owner, a baseline, and a target is a goal. Without that, there is no one to defend the project when budgets tighten, so no one does.
Data integrity is whether the underlying records and telemetry match the operation. This is the one dimension that corrupts everything built above it. Analytics on drifted data does not produce rough answers. It produces confident wrong ones, and every system downstream inherits the error. It is the single weakness you cannot route around.
Change-process robustness is whether the organization can absorb a new way of working and keep it current. A deployment is an event. Adoption is a condition that has to be maintained. Sites that document a change but never train it, or train it but never check that it stuck, watch their gains erode within months.
Starting point is whether you are beginning somewhere small enough to finish and valuable enough to matter. This one is different from the rest. It is not a weakness in the organization to be remediated. It is a choice. Picking an integration-heavy first project, or staying in evaluation indefinitely, is a decision you can change this quarter.
What the assessment should hand back
A useful assessment does not return a score. A score invites false comfort, because an organization can average out to "fine" while being broken in the one place that matters. The output should be a finding: the single weakest point, and what kind of action it demands.
Not every weakness means the same thing, and treating them identically is how readiness work turns into a generic report nobody acts on. The weaknesses fall into four kinds.
A prerequisite must be fixed before you invest. Data integrity is the clearest example. Building on a broken foundation does not merely underperform. It fails, and it takes everything above it with it. You do not start until this is sound.
A parallel fix can be addressed alongside the work. Adoption and change-process gaps are real, but they are worked on during a deployment rather than before it. You can start, provided you design for them deliberately instead of assuming people will come around.
A sequencing fix matters before you scale, not before you start. Scalability is the classic case. Run the pilot, but do not fund a network-wide program until you have a repeatable way to roll it out. Winning the pilot and losing the rollout is the most expensive version of this mistake.
A redirect is not a fix at all. It is a change of plan. A bad starting point means choose a lighter, high-impact first move, not "improve the organization." If you are stuck evaluating, the remedy is to commit to something small. If your first project needs heavy integration before it returns anything, the remedy is a different first project.
The result should also say, in plain language, the first concrete thing to do. Not a maturity level. A next action.
How this differs from the frameworks you may already know
Manufacturing is not short of readiness and maturity models. The most rigorous of them, SIRI (the Smart Industry Readiness Index), AIMRI, and the ARC Advisory Group readiness model, are genuinely good at what they do. They are also built for a different question than the one a plant manager is usually asking.
SIRI scores a factory across process, technology, and organization, and is explicitly a prioritization framework. It is thorough, and it takes weeks to months plus a certified assessor. AIMRI, out of the University of Auckland, is notable because it is one of the few models that treats culture and change management as the critical element, which the evidence supports, but it is a full benchmarking study, not something you can act on this month. ARC is honest about its own limits: it deliberately excludes culture and change management, the very factors the research says dominate outcomes, and it focuses on technology readiness.
Three things separate a useful pre-investment assessment from all of them.
It includes culture, and weights it heavily. Adoption and change absorption are not soft topics to note in an appendix. They are the strongest predictors of whether the money comes back, and a model that leaves them out is measuring the wrong thing with great precision.
It returns a weakest link, not an average. A composite maturity score hides the one broken dimension that will sink the project. The point is to find the thing that fails first, and to say so plainly.
It is short enough to actually take. A framework that requires a certified assessor and a quarter of elapsed time will not be used at the moment it would change a decision, which is before the purchase order. The overview snapshot here takes about five to eight minutes and returns a direction. A deeper, evidence-backed pass comes after, once you know where to look.
| This assessment | SIRI | AIMRI | ARC | |
|---|---|---|---|---|
| Measures culture and adoption | Yes, heavily weighted | Partly | Yes, central | No, by design |
| Returns a weakest link, not an average | Yes | No, a profile | No, a benchmark | No, a score |
| Time to a first answer | 5 to 8 minutes | Weeks to months | Full study | Facilitated |
| Hands back a next action | Yes | A roadmap | A benchmark | A roadmap |
Start where integration is light
There is a practical implication that falls straight out of the starting-point dimension, and it is the part most assessments are too polite to say.
The first project should be one that returns value without a heavy integration burden. Every month spent on integration before any result is a month in which the organization has seen nothing, which is precisely how adoption dies and how projects get cancelled. The evidence favors a wedge. High impact, low dependency, fast to first value, and only then the harder integrations, funded by a result you have already shown.
This is not a technology preference. It is a sequencing preference, and it is the difference between a program that earns the right to expand and one that is still promising value when the budget review arrives.
If you want the direction for a specific operation rather than the argument, the overview snapshot takes a few minutes and returns the single thing to address first. If you would rather walk through a particular plant, start a conversation.
Questions worth asking first
Is this only about AI, or any digital project?
Any of them. The failure modes, adoption, data, ownership, and starting point, long predate AI. AI only punishes them faster and more publicly, because it is confident even when it is wrong.
We already did a maturity assessment. Why would this say something different?
Most maturity models return an average across many dimensions, which lets one critical weakness disappear into a respectable overall number. This returns the single weakest link and tells you whether to fix it first, fix it alongside, or change your plan. The two are not in conflict. They answer different questions.
What if we score badly? Does that mean stop?
It means sequence, not stop. Only one kind of finding, a broken data foundation, says do not start. The others say start, but design for the weakness, or start somewhere lighter. A low result is useful precisely because it is cheaper to learn it now than after the implementation.
How is this different from what a systems integrator would tell us?
An integrator is paid to implement, so the answer tends toward a project. This is taken before a vendor is chosen, and its only output is a diagnosis and a recommended first move. It is meant to make the later conversation with an integrator sharper, not to replace it.
Where does the evidence come from?
Every figure on this page is sourced, and the sources are listed below. Where a number is a projection or a survey rather than a measured result, it is described as one. The snapshot returns the weakest link, not an average, so the result says which area fails first.
Sources
- McKinsey. Losing from day one: why even successful transformations fall short (2021); The state of AI 2025; three-quarters of digital transformations fail, and fewer than a third of digital projects are scaled.
- Boston Consulting Group. Flipping the odds of digital transformation success (2020). About 30% succeed, and the figure has been flat for years.
- Gartner. Gartner predicts 60% of AI projects abandoned through 2026 for lack of AI-ready data (2025).
- MIT NANDA / Project NANDA. State of AI in Business 2025. 95% of enterprise generative-AI pilots showed no measurable P&L impact.
- S&P Global Market Intelligence. Share of companies abandoning most AI initiatives rose from 17% to 42% in 2025.
- Prosci. Best practices in change management. Excellent change management correlates with meeting objectives 88% versus 13%. Effective sponsorship, 79% versus 27%.
- 1-10-100 rule. Data-quality cost heuristic, widely attributed to data-quality practice. $1 to verify, $10 to correct, $100 or more to do nothing.