Why Do So Many AI Programmes Fail? 5 Things Organisations Need to Check
Sass Allard is a strategic coach and change consultant helping leaders and high-performing women navigate complex change with clarity and resilience. She brings over 20 years of experience in global organisations, offering practical insights to support meaningful transformation and growth.
Global IT spending on AI is projected to reach $409 billion this year. MIT's Project NANDA found that around 95 percent of generative AI pilots inside large organisations produce no measurable return, and RAND puts the broader AI project failure rate above 80 percent, roughly twice that of conventional IT projects.
For the most part, the technology works. McKinsey's analysis found that 61 percent of organisations that fail at AI treat it as an IT project rather than a business transformation, and 73 percent of failed projects began without executive agreement on what success would look like. Twenty years of running change programmes tells me neither of those numbers is unusual.
Change management has been quietly demoted
Somewhere over the past two decades, change management was reduced to a delivery function. Communications plans, training schedules, adoption dashboards, a RAG status reported into a steering committee by someone with no authority to alter the thing being reported on. All of it is the administration of a change rather than the management of one.
What determines change sticking is whether people's sense of their own value survives it and whether they trust the organisation enough to say what is really happening on the ground. That is culture change, and it sits closer to organisational psychology than to project delivery.
There is also the matter of people already using it, whatever the company policy says. Many organisations aren’t introducing this change. They’re arriving late to one already underway, which means their principal lever is trust rather than rollout.
Research by Writer with Workplace Intelligence found that nearly a third of workers admitted to actively undermining their employer's AI efforts by refusing to use the tools, feeding them poor input, or slowing implementation down. That number describes a relationship problem.
5 things worth checking before spending
Ask three people at different levels what problem AI is solving here. Three different answers mean the organisation has done procurement but not strategy. This is the single most reliable predictor of what happens eighteen months later.
Find out whether people are using it in the open. Covert use is widely read as a policy failure, when what it actually reports on is trust. Where employees hide productivity gains, they have concluded the gain will be taken from them rather than shared with them, and they are frequently correct.
Check whether anyone has said out loud what happens to jobs. People do not read silence as neutrality. Silence gets filled, and rarely with the generous interpretation. An uncomfortable answer given honestly is more workable than an absence because people can plan around information, but they can’t plan around a vacuum.
Ask a middle manager a difficult question about it. Adoption lives or dies in the layer that fields questions daily. If that manager can only escalate, the change has no transmission mechanism.
Look at the last significant change the organisation made and whether it stuck. Track record predicts far more than current enthusiasm does. An organisation that has struggled to successfully embed change is unlikely to start with the hardest thing it has attempted.
If the answers are uncomfortable
Stop procuring and go back to the question of what the organisation is actually trying to change. Deal with the employment question openly, including where the answer is unwelcome. Give managers access to the reasoning rather than the messaging, because people asked to defend a position they do not understand will defend it badly. Measure behaviour rather than activity, since licences deployed and training completed tell you almost nothing about whether anyone's work has altered.
A colleague and I have spent recent months building an assessment for this situation, largely because the diagnosis is nearly always available but rarely requested. What surfaces tends to be unglamorous yet fixable, provided it surfaces before the money has been committed.
The part that gets skipped
Change management earned its reputation for box-ticking, having been handed to programme teams as a communications workstream with a modest budget line. The discipline required now is of a different order, concerned with how organisations absorb disruption, what conditions allow people to adapt without being diminished, and where the informal power actually sits when the announcement has been made, and everyone has returned to their desks.
Organisations extracting real value from artificial intelligence (AI) have generally been competent at change for years, and the technology has done little beyond rewarding them for it. The remainder are discovering that a technology which learns quickly is unforgiving of an organisation that does not.
Read more from Sass Allard
Sass Allard, Strategic Coach & Change Consultant
Sass Allard works at the intersection of leadership, behaviour, and wellbeing, supporting individuals and organisations as they navigate demanding periods of change. Her background spans two decades in global companies, where she has helped senior leaders strengthen culture, clarity, and capability. She brings a grounded understanding of how hormonal shifts shape women’s experience at work without limiting the broader conversation. As a UN Women delegate to the Commission on the Status of Women, she brings a global lens to agency and progress. Sass writes about adaptation, resilience, and the practical shifts that create real movement in work and life.
