Why Your Enterprise AI Programme Is Failing: It Isn't the Technology

01.Capability Is Rising. Returns Are Not.

If your AI programme has stalled, the reason you have been given is probably the wrong one. It is usually the model, the data quality, the integration work, or the vendor. On the best available evidence, every one of those is getting better year on year, and the returns still are not arriving.

In April 2026, BARC put the question to 639 senior AI leaders at enterprises turning over $100 million or more, on behalf of Domino Data Lab. Ninety-three per cent said they had got better at getting AI into production, up from 88 per cent the year before. Fifty-seven per cent said their return on investment still fails to outpace what they are spending, a number that has not moved since 2025 ^[1].

Those two findings are the whole problem in miniature. Enterprises are getting materially better at the engineering and no better at converting it into money. Capability is close to solved. Value capture is not. That is the gap this article is about, and none of it closes by buying a better model.

The corroborating evidence points the same way. BCG's survey of 1,250 senior executives and AI decision-makers found 60 per cent achieving no material value from AI, reporting minimal revenue and cost gains despite substantial investment, though the same study found the cohort seeing some returns had grown 13 percentage points year on year ^[4]. McKinsey's November 2025 global survey of 1,993 respondents found 39 per cent attributing any enterprise-level EBIT impact to AI, and most of those put it below 5 per cent of EBIT ^[5]. And Gartner's survey of 782 infrastructure and operations leaders, fielded in November and December 2025 and published the following April, found only 28 per cent of AI use cases in infrastructure and operations fully succeeding against ROI expectations, with 20 per cent failing outright ^[3].

One clarification worth making, because it is routinely fudged. Gartner forecasts worldwide AI spending will reach $2.59 trillion in 2026, a 47 per cent year-on-year rise ^[2]. That number is regularly set beside enterprise failure rates to imply trillions being burned on doomed corporate projects. It does not mean that. Gartner's own release is subtitled to the effect that the spend is dominated by vendors and hyperscalers, with enterprises yet to flex their spending potential; AI infrastructure alone is over 45 per cent of the total. It is chip, server and data-centre capex, not enterprise project budgets. Forrester, meanwhile, predicted in October 2025 that enterprises would defer a quarter of planned AI spend into 2027 as financial scrutiny tightens ^[6]. That is a forecast, not an observation, and it describes a different pot of money entirely.

02.The Awkward Part: It Demonstrably Works

Any honest account of enterprise AI has to deal with an inconvenient body of evidence: at the level of individual tasks, the technology works, and it has been measured properly.

The canonical study is Brynjolfsson, Li and Raymond's field experiment tracking the staggered rollout of a generative AI assistant to 5,172 customer support agents at a Fortune 500 company. Issues resolved per hour rose 15 per cent on average, and around 36 per cent for workers in the lowest skill quintile ^[7]. A February 2026 review of the empirical literature found controlled field experiments and randomised trials consistently documenting task-completion-time reductions of 15 per cent to more than 50 per cent across writing, customer support, software development, accounting, law and translation, replicated across firms, occupations and experimental designs, with consistently larger gains for less experienced workers ^[8].

On the enterprise side, Wharton's third annual study with GBK Collective, surveying more than 800 senior leaders at companies with over 1,000 employees and at least $50 million in revenue, found 74 per cent already reporting positive return on generative AI, and 72 per cent formally measuring it ^[9]. Stanford's Digital Economy Lab went further and documented 51 enterprise AI deployments across 41 organisations, seven countries and more than a million combined employees that are live in production, used consistently, and delivering measurable value ^[10]. And adoption is still spreading, though less evenly than the headlines suggest. The US Census Bureau's Business Trends and Outlook Survey put firm-level AI use at 19.8 per cent in early May 2026, having hovered between 17 and 20 per cent since December 2025, with firms expecting 20 to 23 per cent within six months ^[30]. A Census working paper puts adoption at 32 per cent on an employment-weighted basis over the November 2025 to January 2026 period, which is the number to use when you care about how many workers are affected rather than how many firms ^[31]. The growth is concentrated in firms of 20 employees or more; use among smaller firms has not moved significantly. It is also heavily skewed by sector, from 39.7 per cent in Information down to around 14 per cent in Retail Trade ^[30]. Be careful comparing any two of these figures: a Federal Reserve review of the competing instruments found the same underlying question producing about 18 per cent of firms in one survey and an employment-weighted 78 per cent in another ^[11].

These findings deserve their caveats. Wharton's ROI is self-reported by leaders with an incentive to claim success. Stanford selected on success, so its sample says nothing about base rates. But taken together they make one thing untenable: you cannot claim the technology does not work.

The same literature also contains the opposite result, and it is more interesting than either camp's summary. Harvard's "jagged frontier" study found that on a task deliberately built to sit beyond AI's capability boundary, consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it ^[12]. The honest synthesis is not "AI works" or "AI fails". It is that AI reliably helps less experienced people on tasks inside its frontier, and can actively harm experts on tasks outside it. Knowing which side of that line a given task falls on requires knowing the task in detail, which is where this article is going.

03.Five Numbers You Should Stop Quoting

Before going further, a detour that will save you an embarrassing board meeting. The enterprise AI failure discourse runs on a handful of statistics that do not mean what they are used to mean. We checked them against primary sources. Five of the most repeated do not hold up.

This matters beyond pedantry. Boards are making capital allocation decisions on these numbers, and a strategy justified by a statistic that collapses under a single click is a strategy with no floor under it. It also cuts both ways: the same laundering that produces "80 per cent fail" produces the vendor deck claiming 300 per cent ROI. If you are going to be sceptical, be sceptical symmetrically.

04.The Real Bottleneck Is Tacit Knowledge

So the models work, the engineering is improving, adoption is climbing, and the returns are flat. What sits in the middle?

Every organisation that starts an automation project brings documentation: process flows, procedure manuals, compliance checklists. On paper the process exists. In practice it is rarely what people actually do. Staff develop their own methods. Rules nobody wrote down. Exceptions handled differently depending on who picks up the case. The gap between the documented process and the real one is almost always wider than anyone expects, and it is precisely where automation projects die, because an AI system given the documented process automates a process that does not exist.

This is not a new observation about work. What is new is that it has become the binding constraint. When automation meant rules engines, you could ship something useful from the documented process and handle exceptions manually. When automation means a model making judgement calls, the exceptions are the product. The value is in the 20 per cent of cases that need judgement, and judgement is exactly the part nobody wrote down.

Which puts a specific, unglamorous question at the centre of enterprise AI strategy: how do you get what is in your experienced people's heads into a form a system can use? Every serious framework arrives at the same answer by a different route. BCG's well-known 10-20-70 allocation (roughly 10 per cent of effort on algorithms, 20 per cent on technology and data, 70 per cent on people and processes ^[27]) is a statement that the hard part is organisational. And in Stanford's catalogue of deployments that worked, 77 per cent of the hardest implementation challenges were non-technical: change management, data quality and process redesign ^[10].

That framing has become almost universal in 2026 commentary, and on its own it is not useful. "It's a people problem" is where most analyses stop. It is where the interesting question starts: why, specifically, does the knowledge not move? Not as a matter of culture or communication, but as a matter of incentives.

05.Nobody Documents Themselves Out of a Job

Here is the trap, stated plainly. To automate a process you must document it to a fidelity that includes every exception, every judgement call, every workaround. An employee who does that has written the specification for their own replacement. The single act that makes AI implementation possible is the act that makes the person doing it redundant.

The organisation is asking its most experienced people to volunteer the one asset that makes them hard to replace, and offering them nothing for it.

For a long time this was an anecdote that consultants swapped over coffee. In March 2026 it became peer-reviewed evidence. A three-phase longitudinal study published in Humanities and Social Sciences Communications, a Nature Portfolio journal, tracked 324 R&D employees across nine manufacturing firms and found that organisational AI adoption does not produce one uniform response: it bifurcates the workforce ^[19]. Employees with an external locus of control appraised AI adoption as a hindrance, and hindrance appraisal drove measurable knowledge hiding. Employees with an internal locus appraised the same change as a challenge, and increased knowledge sharing. Same technology, same announcement, opposite behaviour.

That is a sharper claim than "people resist change". It says AI adoption actively sorts your staff into those who feed the system and those who starve it, and it identifies what determines which way someone falls. The study's limits should be stated: single country, single sector, R&D employees, self-reported measures. It is not proof that everyone hoards. It is solid evidence that the mechanism is real and predictable.

Survey work points the same way, with the same caveat about who paid for it. The Adaptavist Group, an IT services vendor, surveyed 4,000 knowledge workers across the UK, US, Germany and Canada and found 35 per cent hoarding knowledge for fear of being replaced, and 38 per cent reluctant to train colleagues in areas they consider personal strengths ^[22]. Harvard Business School researchers Das Narayandas and Shunyuan Zhang describe the underlying dynamic as an identity threat with three components (role compression, control shift and span erosion) and give the resulting behaviour a name: symbolic adoption, in which employees give every appearance of using the new system while quietly withholding the contextual knowledge that would make it work ^[20]. In June 2026 Harvard Business Review documented the adjacent behaviour from the other direction: employees concealing their AI usage, for fear of being judged, handed more work, or looking replaceable ^[21].

The uncomfortable part is that this fear is not irrational, and the best evidence for that comes from the success stories. In Stanford's catalogue of 51 deployments that genuinely worked, headcount reduction was the largest single outcome category, appearing in 45 per cent of them ^[10]. Workers watching successful AI deployments and concluding that success means fewer colleagues are reading the evidence correctly. Change programmes that treat this as a communication failure to be managed with reassurance are trying to talk people out of an accurate belief.

There is a measurable cost to getting this wrong. A separate study in the same journal, tracking 381 South Korean employees across three time-lagged waves, found AI adoption associated with lower psychological safety, which in turn was associated with higher employee depression, with ethical leadership moderating the effect ^[23]. Meanwhile PwC's global workforce research found the opposite pole is achievable: daily AI users report higher pay, job security and productivity, with PwC's Pete Brown noting that to scale those benefits "businesses must go beyond training. Work itself needs to be redesigned" ^[25].

One note of caution on this territory. A large share of the widely circulated statistics about employees actively sabotaging AI rollouts (the "29 per cent of workers, 44 per cent of Gen Z" numbers that drove a news cycle in April 2026) come from a single survey commissioned by a vendor that sells enterprise AI software, with the headline findings written up and promoted by that vendor's own chief marketing officer ^[26]. The fieldwork itself was run by an independent research firm across 2,400 respondents, so the figures may well be sound. But the research was paid for and framed by an interested party, and a case built on it is a case built on one interested party.

06.The Consultant Has the Same Incentive

Now apply the same lens one level up, because this is where the two halves of the problem turn out to be one problem.

Faced with internal resistance and a shortage of in-house expertise, executives reach for consultants. The engagement produces a strategy document, a reference architecture and a proof of concept that works beautifully in a controlled environment. Then the team rotates off. What remains is a system the client owns but nobody inside the client organisation fully understands. That is the same failure as the undocumented process, one layer of abstraction higher.

And the incentive structure is identical. A consultancy whose client becomes genuinely self-sufficient has ended its own revenue stream. A consultancy whose client remains dependent has an annuity. Nobody needs to act in bad faith for this to shape outcomes; it is simply what the contract rewards. The employee has a rational incentive not to document themselves out of a job, and the supplier has a rational incentive not to train themselves out of a contract. The knowledge the AI needs sits behind two separate parties who both lose by handing it over. That, rather than any shortage of models or GPUs, is the enterprise AI bottleneck.

A caveat that matters, because the obvious conclusion is wrong. "Build it internally instead" is not what the evidence supports. MIT's Project NANDA data, whatever its methodological limits, found that purchasing AI tools from specialist vendors succeeded roughly twice as often as purely internal builds ^[14]. External expertise is not the problem. The problem is engagements structured so that knowledge leaves when the people do.

The practical test is simple, and it is a contracting question rather than a technology one. Does the engagement have knowledge transfer as a deliverable with acceptance criteria, or as a line in the closing slide deck? Who operates the system in month nine? If the honest answer to the second question is "we would have to call them back", the project has not been delivered. It has been rented.

07.What Actually Moves the Number

If the constraint is knowledge transfer that nobody is incentivised to perform, the fix is to change the incentive rather than to buy better software. Four things follow from the evidence.

Pay for the knowledge, explicitly. If documenting a process to automation fidelity is genuinely valuable, and it is the scarcest input in the whole programme, then treat it as work, compensate it, and credit it. Organisations routinely budget seven figures for a platform and nothing at all for the elicitation that determines whether the platform does anything. Reversing that ratio is the single highest-leverage change available.

Make the post-AI role visible before asking for the knowledge. Narayandas and Zhang prescribe five complementary mechanisms: recharter roles so employees can see a more valued version of their job on the other side, build decision guardrails that preserve discretionary override authority, add analytical overlays that enhance rather than replace judgement, open credible redeployment pathways, and put executive sponsorship behind the redesigned roles ^[20]. An employee who can see their role after the project has a reason to help build it. One who cannot will optimise for staying indispensable, and will be right to.

Expect the first attempt to fail, and budget for it. The most useful single number in Stanford's study of successful deployments is that 61 per cent of them had a failed attempt first ^[10]. Early failure is a stage, not a verdict. Governance that treats the first unsuccessful pilot as grounds for cancellation systematically eliminates the projects that were about to work. And given that 42 per cent of companies scrapped most of their AI initiatives in 2025, up from 17 per cent the year before, while the average organisation abandoned 46 per cent of its proofs of concept before they reached production ^[24], this is not a hypothetical failure mode.

Pick tasks where you can already tell right from wrong. The jagged-frontier finding means the deciding factor is whether a task sits inside the model's competence, and you cannot know that from a vendor demo. Start where the output is checkable, the comparison case is measurable, and the AI's contribution can be isolated. That is also, conveniently, where the knowledge you need to elicit is smallest.

One caution on the agentic wave specifically. Gartner predicted in June 2025 that more than 40 per cent of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls ^[28]. Agents raise the knowledge requirement rather than lowering it: a system that takes actions needs to know your exceptions far more precisely than one that drafts text for a human to check.

08.What This Means If You Are Buying AI

For the mid-sized companies we work with across Ireland and the UK, the practical implication is mostly a relief: you are not behind because you lack a research team. The World Economic Forum expects 39 per cent of workers' existing skill sets to be transformed or become outdated by 2030 ^[29], and none of the evidence above suggests the winners will be whoever bought the largest platform. The differentiator is whether the people who know how the work actually happens are willing to say so.

Smaller organisations have a real structural advantage here. The distance between the person doing the work and the person deciding what to automate is measured in desks rather than org-chart layers. The tacit knowledge is reachable. What is usually missing is a method for getting it out and a reason for anyone to cooperate. Both of those are cheaper to fix than a failed platform migration.

That is the shape of how we approach automation and custom development work: map what actually happens before choosing anything, and build so the team can operate it without us. If the goal is internal capability rather than a dependency, training is usually the better first purchase. Our AI training programmes exist for exactly that. We also run regular community events in Dublin where this comes up constantly, usually in more candid terms than anyone uses in a boardroom. If you want to talk through a specific process, get in touch.

09.Conclusion

"Enterprise AI is failing" is the wrong sentence, and the volume of commentary repeating it in 2026 has made it harder rather than easier to see the actual problem. The models work. Adoption is rising. Production capability is improving on every measure available. What has not moved in a year is the conversion of that capability into profit.

The reason is not mysterious, but it is unflattering, which may be why it gets less airtime than the technology. Making AI valuable requires transferring knowledge that currently lives in people's heads and in suppliers' methodologies. Both parties are worse off for handing it over. Changing that arithmetic means paying for the knowledge, showing people the role that exists on the other side, and contracting for capability instead of deliverables. Until an organisation does that, it can buy any model it likes and the number will stay flat.

The organisations pulling ahead are not the ones that found better technology. They are the ones that made it safe, and worthwhile, to explain how the work is really done.

10.Frequently Asked Questions

11.References

Every statistic in this article was traced back to its original source before publication, and the survey dates, sample sizes and scope conditions are stated wherever they change what the number means. Where a widely repeated figure did not survive that check, it appears in section 03 rather than in the argument.