The reason AI rollouts aren’t paying off has nothing to do with AI
Most AI initiatives fail to deliver measurable value. This is not because the technology is lacking, but because organisational design and operating models prevent benefits from being realised, writes Fredrik Hagstroem, Chief Technology Officer at technology consultancy Emergn.
It is not hard to find explanations for why AI initiatives are not delivering value. Common and real issues that are easily pointed to are skill gaps, poor governance, data quality, and a lack of tooling and resources. Nine times out of ten, these were existing preconditions; AI just made them more painful and harder for executives to ignore.
The harder pill to swallow is to realise that the organisation is unlikely to be designed or set up to capture the value from the solutions AI enables. GenAI in particular is great at improving work. It can complete tasks faster, at lower cost, and much more consistently – around the clock, in multiple languages and without the same constraints as human teams.
So finding benefits that, on paper, add up to significant business impact is easy. Logically and practically, getting approval for experimenting and building prototypes is also usually quick, because AI is exciting and the up-front investments and risks are limited.
Pilots everywhere, payoff nowhere
In practice, we get pilot proliferation, but scaled solutions are scarce. And, more important than size, the economic benefits fail to materialise on the bottom line. McKinsey & Company has described this as “Pilot Purgatory”. The data supporting it includes MIT’s finding that 95% of GenAI pilots delivered no measurable profit & loss impact.
The answer to why value is bottled up in limbo is not found in the cloud; it sits with systems thinking. A complex organisation does not operate as a series of discrete tasks. Value is not created as tasks are completed, but only when work has flowed across teams, functions and decision points and the outputs are consumed by external customers and users. Improving one part of that flow only improves the system when it eases the overall constraint.
A simple example makes the point. Creating a meeting summary from a transcript that used to take 30 minutes can now be done in minutes. That is an obvious productivity gain. However, if the decisions or actions taken from the summary are the same, there is no difference in the end product or the value that customers pay for.
This pattern shows up repeatedly in AI and Gen AI programmes. Individual teams become more efficient, which leads to departments reporting gains, and adoption looks encouraging. Despite this, customer experience, sales cycles and delivery timelines remain largely unchanged, because the work is still moving through the same functional boundaries.
Most companies are structured for functional efficiency. Tasks define roles, and performance is measured within departmental boundaries. When AI pilots are introduced into this environment, they are typically applied within those same boundaries.
Our own research – based on a survey of more than 750 enterprise leaders – shows that more than half (55%) of leaders say they will not meet their AI goals without the talent, problem framing, and outcome-focused design typically led by product teams.
For business leaders, and particularly those in human resources, this shifts the nature of the challenge. AI is not another technology rollout. It is a challenge of work design. Training people will not compensate for an operating model built around siloed thinking.
Recalibrating the work operating system
A more useful starting point is to ask what the work is actually for, and who needs to consume the output. This reframes the discussion around outcomes rather than activity, and forces organisations to look at how work moves between teams, where decisions are made, and where value is created or lost.
Seen this way, the reason pilots so rarely reach scaled production comes into focus, and two failures stand out. Task automation is the wrong model for how humans and AI work together: it speeds up individual steps without changing how the work connects. And because the work is not aligned end-to-end, ownership of the outcome is left unclear – each team owns its part, but no one owns the result.
The consequence is that human review and decision points end up in the wrong places, bolted onto tasks rather than sitting where they matter. This exposes the real constraint: execution is limited by readiness to trust – not by models, data or technology. Work stalls wherever people are not yet willing to rely on an output and own what follows from it.
Closing that gap takes more than a redesigned task; it takes a different operating model for how work is executed. The shift is from “humans using tools” to humans as circuit-breakers, placed at the points in the flow where trust has to be built – the tools carry the work, while people sit at the junctions that hold real risk and judgement. Processes are then redefined around those trust points, not around the tasks.
Design for trust, not tasks
None of this is entirely new. The importance of end-to-end thinking has been recognised for years. What AI changes is the structure of the problem. As individual tasks become faster and more efficient, weaknesses in handovers and connections become harder to ignore.
Returns on AI investments will be found in the end-to-end processes that generate value, and only once those processes have been optimised as a whole rather than task by task. Until organisations are designed to capture the value they create, more capable tools will not solve the problem. They will simply make the gap painfully more visible.
The real question is not whether AI works, but whether the organisation is set up to benefit from it.
