The promise versus the floor
For three years now, the technology industry has been telling manufacturing leaders that generative AI will transform their operations. Conference keynotes are full of it. Vendor decks are full of it. LinkedIn is full of it. The actual factory floor, however, is largely empty of it.
The gap between promise and deployment is not a marketing problem. It is a structural problem with how enterprise AI has been funded, scoped, and shipped. And in 2025, that gap is starting to show up in the only place that matters: the financial statements of the companies that bought in.
What the numbers actually say
The funding numbers tell one story. The deployment numbers tell another. The ROI numbers, when anyone bothers to measure them, tell a third.
Recent industry analyses suggest that a substantial majority of enterprise AI projects fail to deliver measurable business value. Not "fail to meet aggressive targets." Fail to deliver any value that can be attributed to the AI itself, separate from what the existing system was already doing.
This is not a technology failure. The models work. The math works. The failure is in the application: companies are buying AI as if it were a product, when in reality it is a capability that has to be wrapped in workflow, data, and human review before it produces anything useful.
The high-profile warning signs
When a company announces 700 layoffs and attributes them partly to an AI implementation that didn't deliver, that is not a one-off. It is a signal.
When a category-defining consumer tech company pulls back from AI customer service after public complaints about service quality, that is not a one-off. It is a signal.
When the funding cycle tightens and the valuations of AI infrastructure companies start to compress, that is not a one-off. It is a signal.
The signal is this: the AI hype cycle is over. The next phase is the AI value cycle, and the rules are different. In the value cycle, you do not get credit for deploying a model. You get credit for changing an outcome.
Where the value actually sits
For manufacturing and supply chain companies, the value of AI does not sit in chatbots. It does not sit in document summarisation. It does not sit in copilots that write emails faster.
The value sits in three places:
-
Forecast accuracy. A 2% improvement in forecast accuracy at a mid-market manufacturer with $200M in inventory is worth real money. Not vanity money. Inventory money. Working capital money.
-
Exception triage. A planner who reviews 50 exceptions a day instead of 500 is a planner who can actually think. The AI does not replace the planner. It removes the noise so the planner can do the job they were hired to do.
-
Audit and version control. The boring part. Every override, every approval, every change captured with actor and timestamp. This is not exciting. It is what makes planning defensible to finance, to auditors, and to the board.
What to do about it
If you are a manufacturing leader being pitched AI right now, here is the test:
Ask the vendor what specific operational outcome their product changes. Not what it "enables." Not what it "unlocks." What number moves, and by how much, and over what timeframe.
If the answer is a circle back to "transform your business" or "unlock the power of," you are being sold hype. Walk away.
If the answer is "we reduce S&OP cycle time from three days to four hours by replacing the spreadsheet reconciliation with a sequential gate workflow," you are being sold a product. Stay in the room.
The emperor's code has no clothes. But underneath the clothes, there is a real body of operational intelligence that does work. The companies that find it will win the next decade. The companies that keep buying the clothes will keep losing money.
This article is part of prompt/ed, the Tensor Analytics education hub. Written for manufacturing practitioners, not AI enthusiasts.