The numbers are starting to come in, and they are not flattering. McKinsey, Gartner, Deloitte, BCG. Pick a consultancy and pick a survey. Every one of them is reporting the same finding from a different angle: somewhere between seventy and eighty percent of enterprise GenAI initiatives that started in 2023 are still pilots in 2024. They never shipped. They never scaled. They never moved past the demo deck.
This is not because the technology does not work. The technology works. ChatGPT crossed a hundred million users in two months. Claude, Gemini, Llama, the open-source models, the closed ones. Anyone who has spent ten minutes with a current frontier model knows that what these systems can do is genuinely new. The blocker is not capability. The blocker is everything that surrounds capability inside a real enterprise.
The mistake most organisations made in 2023 was treating GenAI as a technology problem. Procure the model. Fine-tune on internal data. Deploy. Done. The mistake is understandable. That is roughly how the previous wave of enterprise software adoption worked. With GenAI, it does not.
A working LLM in production is not a piece of software. It is a system that interacts with messy enterprise data, makes probabilistic decisions, fails in ways that are difficult to predict, and changes its behaviour every time the underlying model gets updated. It requires governance frameworks that did not exist a year ago. It requires data infrastructure most organisations have not yet built. It requires human oversight processes that compliance teams are still figuring out how to write. It requires explainability work that the model vendors themselves have not solved. It requires evaluation harnesses that most engineering teams have never built before because deterministic software did not need them.
Each of these problems is solvable. None of them gets solved by a pilot project running in a corner of the innovation team. They get solved by treating GenAI deployment as an organisational programme that touches data engineering, legal, risk, operations, and frontline business units simultaneously. The organisations that have shipped real GenAI into production this year, and there are some, all look similar in one respect. They stopped treating the model as the product. They treated the integration as the product.
There is a harder truth underneath this, which is that a lot of the 2023 pilots were never going to ship regardless. They were exploratory exercises designed to be seen doing something with AI, not to deliver business value. The board wanted a tick in a box. The pilot delivered the tick. The pilot is now over, the box is ticked, and there was never a plan for what came next. This is not a GenAI problem. This is an organisational behaviour problem that has happened with every hyped technology since cloud, and it will happen with the next one too.
For the organisations that are serious, 2025 is the year the pilots either become real or get quietly retired. The cost of doing GenAI properly has become clear. The cost of doing it badly is becoming clearer. The middle ground, where you have a pilot and a press release and nothing in production, is no longer a comfortable place to sit.