A quarter of organisations have moved forty per cent or more of their artificial intelligence experiments into production. Just over half expect to do so in the next three to six months. Both numbers come from Deloitte's 2026 State of AI in the Enterprise survey [@StateAIEnterprise2026], and they are interesting principally because the second has appeared, in some form, in essentially every major AI adoption survey for nine consecutive years. The first has barely moved.
For 9 years confidence in strategy has consistently outstripped the delivery of strategic outcomes.
If one had been told in 2017 that more than half of all enterprises would, within three to six months, scale their AI experiments into production, and then told the same thing in 2019, and again in 2021, and again in 2023, and again last week, one might be forgiven for suspecting that the three-to-six months in question are not three-to-six months of our earth revolving around our sun. They are a managerial interval: imminent, expected, and always at least a quarter or two away. Far enough to avoid accountability, near enough to keep the promise alive.
This is the scaling number, and it is the cleanest evidence available that the AI conversation in contemporary organisations is operating in a register that has very little to do with the underlying AI technology.
The problem is not that scaling is impossible. The Deloitte survey, with admirable candour, lists the things that scaling actually requires. Infrastructure investment. Integration with existing systems—I'll say that again as it is usually the biggest blocker—Integration with existing systems. Security reviews. Compliance checks. Monitoring systems. Ongoing maintenance. None of these are exotic. Engineering and finance functions have been doing all of them, in some form, for forty years. The question is not whether the work can be done. The question is whether it is being done.
The empirical evidence on this is unambiguous. The same Deloitte report notes that use cases estimated to take three months routinely stretch to eighteen when integration complexities emerge. Models that achieved high accuracy in pilots prove inadequate at scale. Failures that were learning opportunities in a controlled environment become business risks in production. None of which is surprising. All of which is precisely what one might have predicted, in 2017, about the move from pilot to production for any complex organisational technology. The surprise is not that it is hard. The surprise is that organisations have spent nine years discovering it is hard and have responded to the challenge by commissioning further pilots.
The Deloitte report attributes its eighteen-month integration figure to "edge cases". In enterprise AI, these "edge cases" turn out to be the norm. What is sold as "highly configurable out-of-the-box" routinely requires custom code and several developer squads to make functional. What is promised as connected proves not to be, even when the entire stack comes from a single supplier—the consequence of a decade in which the largest technology companies have steadily absorbed their competitors and produced ecosystems whose components do not reliably communicate with one another. The connectivity promise outstrips operating reality by an order of magnitude. Integration complexity is not the edge. It is the centre.