The Real Prize in Factory AI Isn t the AI
When Infor and AWS announced AI agents for factory floors last week, the most important number was not in the press release. Eighty-six percent of manufacturing AI pilots never make it to full production, and the reason tells you who really wins.
When Infor and Amazon Web Services unveiled a suite of AI agents for factory floors at Hannover Messe on April 20, the announcement arrived with the usual choreography: live customer testimonials, percentage gains, a stage full of slides. Buried in the fine print was the constraint that makes the whole thing interesting. The agents, currently available only through a product called Infor Velocity Suite, do not try to reason their way into manufacturing from scratch. They start with data the manufacturer already has. Without it, the system does not work.
That prerequisite is the part Infor wants to sell. The company, a Koch Industries-owned enterprise software firm with 60,000 manufacturing customers worldwide, is not primarily pitching a smarter AI model. It is pitching access to the data infrastructure those customers built over decades — and arguing that the AI is the easy part. Rick Rider, Infor's senior vice president for industry product, put it plainly in the announcement: "Generic AI doesn't work in manufacturing — you need agents that understand manufacturing-specific operational processes, bill of materials, supply chains, and shop floor realities."
Xpress Boats, a US manufacturer that makes boats to order, ran an early version of the system against its own operations and reported a 50 percent reduction in expedited shipping costs, a 98 percent improvement in how quickly it could diagnose a process problem, and a 95 percent reduction in the time it took to handle returned goods. Those are real numbers from a real customer, but they are a single data point, not a trend. The solution is in limited availability. Independent confirmation does not yet exist.
The gap the announcement is trying to address is real and well-documented. Only 14 percent of companies that started a generative AI pilot in manufacturing made it to full production by the middle of last year, according to ERP Today; the other 86 percent stalled somewhere between proof-of-concept and the actual shop floor. What makes Infor think it can move that number is the same thing that makes it an interesting test case: the company already has the ERP relationship. It has the bill of materials libraries, the inventory records, the shop floor integration. The AI agent is the layer on top. The data is the foundation.
That distinction reframes where the competitive boundary lies. Vendors pitching generic large language models into manufacturing environments are discovering the same thing: without structured data underneath, the AI recommendations are unreliable. A factory running messy, incomplete ERP records gets messy, incomplete AI recommendations. The companies that spent years cleaning up their enterprise software have built something the AI vendors need more than the vendors want to admit publicly.
The implication cuts both ways. Manufacturers with well-organized enterprise data now have something AI vendors want: usable context. And for the AI vendors themselves, the race is not just to build better agents — it is to sign up the manufacturers who already have the data infrastructure the agents need. Infor's expanded partnership with Deloitte, targeting the layer between enterprise planning systems and actual shop floor operations, suggests where the competitive boundary is being drawn: not at the model, but at the integration layer that connects AI reasoning to factory reality.
What to watch next is straightforward. Infor has not disclosed how many customers are currently running the system in production. Whether this approach actually moves the needle on that 14 percent production success rate depends entirely on whether the limited availability rollout produces independently verifiable results from additional manufacturers. The structural argument is coherent. The data point that would make it a trend — another manufacturer besides Xpress Boats reporting comparable numbers, with the methodology visible — has not arrived yet.
The manufacturing ERP market, at $23 billion in 2025 and growing at roughly 8 percent annually, is large enough and slow-moving enough that a single announcement does not restructure it. But Infor's positioning — as the company that already has the data, now building the AI on top of it — is a pattern worth watching. Enterprise software vendors are discovering they have more leverage in the AI era than the AI vendors would prefer to admit.
