The narrative is seductive: a company adopts AI across every function, productivity soars, costs plummet, and the organisation enters a new era of data-driven decision-making. McKinsey publishes a case study. A conference keynote uses it as an example. The narrative spreads.
The problem is that the narrative is mostly wrong, not because AI does not work, but because the way it transforms organisations looks nothing like the story.
What the hype promises vs. what research finds
The hype promises transformation: AI will change how companies operate at every level. The research finds something more modest and more interesting. AI produces significant gains in narrow, well-defined tasks. It produces mixed results in broad, complex processes. And it produces organisational disruption that often exceeds the direct productivity gains.
A Stanford study on AI-assisted customer service found a 14% productivity increase. Impressive, but not transformational. And the gains were concentrated among less experienced workers, who used the AI as a learning tool. Experienced workers saw smaller gains because they already had the tacit knowledge the AI provided.
A Harvard Business School study on AI-assisted consulting found that consultants using AI produced higher-quality work on tasks they were already good at, but lower-quality work on tasks outside their expertise. The AI gave them false confidence on unfamiliar problems. They did not know enough to evaluate the AI’s output, so they accepted flawed work.
These findings are consistent across the literature. AI amplifies existing capability. It does not create capability where none exists. The organisations that benefit most from AI are the ones that were already good at the underlying work. The organisations that struggle are the ones that hoped AI would compensate for gaps in their fundamentals.
The survivorship bias in case studies
Most published case studies about AI success share a common flaw: they only report on the successes. The company that deployed AI and saw a 30% reduction in processing time gets a case study. The company that deployed the same AI and saw no improvement, or a decline, does not.
This survivorship bias creates a systematically distorted picture of AI’s impact. We see the winners. We do not see the base rate. For every company that successfully deployed AI at scale, there are multiple companies that attempted it and either failed quietly or achieved marginal results that were not worth the investment.
The research that accounts for this bias paints a different picture. A McKinsey Global Survey found that while 65% of respondents reported their organisations were regularly using generative AI, less than 10% reported seeing meaningful impact on EBITDA from those deployments. The adoption is widespread. The impact is concentrated.
This does not mean AI does not work. It means AI works under specific conditions that most organisations have not met.
The conditions that matter
The research consistently identifies three conditions that separate successful AI deployments from unsuccessful ones.
First, the task must be well-defined. AI produces reliable gains when the input, output, and success criteria are clear. Code generation, document summarisation, and data extraction are well-defined. Strategic planning, organisational design, and creative problem-solving are not. The tasks where AI has the most impact are the ones where human judgment is least needed, which means the tasks where AI replaces the easiest human work, not the most valuable.
Second, the user must be able to evaluate the output. AI is most dangerous when the user cannot tell whether the output is correct. A senior engineer can evaluate AI-generated code. A junior analyst may not be able to evaluate AI-generated analysis. The value of AI depends on the user’s expertise, which means AI creates a paradox: it is most useful to the people who need it least.
Third, the organisation must have the infrastructure to support AI. Not just the technical infrastructure (the data pipelines, the monitoring, the deployment systems) but the organisational infrastructure. Clear data governance, defined quality standards, and a culture that treats AI outputs as hypotheses rather than answers. Without this infrastructure, AI produces outputs that nobody trusts and nobody acts on.
The honest assessment
AI is a real capability with real applications. It is also the most overhyped technology since the blockchain boom, and the overhype creates real damage. Organisations invest in AI when they should be investing in data quality. They hire prompt engineers when they need data engineers. They build AI strategies when they need business strategies.
The honest assessment is that AI is a powerful tool that amplifies existing organisational capability. If your organisation has clean data, skilled people, clear processes, and a culture of critical thinking, AI will make you significantly more productive. If your organisation lacks these things, AI will give you the ability to produce wrong answers faster.
The research is clear on this point. AI does not fix broken organisations. It makes broken organisations produce broken outputs at a higher velocity. The companies that succeed with AI are the ones that invested in the fundamentals first.
Do not ask “how do we adopt AI?” Ask “are we ready to adopt AI?” The answer to the second question is harder and more important.
