Why AI programmes fail, and what the numbers actually say

Every organisation adopting AI eventually meets the same statistic, usually in a board paper, usually without a source. The numbers are worth stating carefully, because the specific shape of the failure tells you where to spend.

The failure rate is real, and it is high

BCG research puts the proportion of AI projects that fail to deliver expected benefits at 70–85%, roughly twice the rate of traditional IT projects. RAND has reported figures above 80%. Even conservative estimates put major enterprise AI initiative failure at around 43%.

The spread between those numbers is mostly definitional. What does not vary is the direction: most AI programmes underdeliver, and they underdeliver more often than comparable technology programmes.

The cause is not technical

This is the part that gets lost. Analyses of AI transformation consistently attribute roughly 30% of success to technology and 70% to people, process and culture.

The clearest evidence sits in the gap between two adoption numbers. Around 88% of organisations now use AI in at least one business function. Only 39% report measurable impact on the bottom line.

Deployment is nearly universal. Benefit is not. Whatever separates those two figures, it is not model quality — the same models are available to everyone in both groups.

Organisations underfund the transition by a factor of three

Programmes that succeed put 30–40% of resources into change management: communication, training, workflow redesign, and support after go-live. Typical organisations allocate about 10%.

The effect of getting this right is large and well documented. Projects with excellent organisational change management report a 73% success rate, against 39% for those with fair programmes.

A programme that spends 90% of its budget on a platform and 10% on the transition has, statistically, bought the failure case.

The workforce has not been prepared

Two figures explain most resistance you will encounter.

• Only 13% of workers report receiving any AI training from their employer.
• Around 40% of employees fear losing their job to AI, up from roughly 28% two years earlier.

Worse, the share of organisations offering formal AI upskilling fell to about 26% in 2026, from roughly 35% the year before. Investment in tools went up. Investment in the people expected to use them went down.

Untrained and anxious is a poor starting position for adoption. It produces quiet non-use, which is much harder to detect than open refusal, because the dashboard still shows the licence as active.

What follows from this

If the failure mode is organisational, the remedy has to be organisational.

1. Budget the transition properly. If change management is under 20% of programme cost, the plan is optimistic.
2. Train by role, not by awareness. An all-hands session is not training. Executives, managers and frontline users need different curricula.
3. Name the job-security question. It is the actual conversation happening in your organisation. Silence from leadership gets filled with worse assumptions.
4. Measure adoption, not deployment. Licences issued is a procurement metric. Behaviour change is the one that predicts benefit.
5. Sequence by value and feasibility. The loudest request is rarely the highest return.

None of this is novel. Change management has been a mature discipline for decades. What is new is how quickly AI raises the cost of skipping it — because unlike most systems, an AI tool that people quietly stop trusting will still return answers, and nobody will tell you.

---

Figures cited are drawn from published industry research current at the time of writing. We confirm the applicable data and regulatory position for each client’s jurisdiction and sector at the start of an engagement.