The demo wasn’t the hard part: Getting manufacturing AI into production

You know the meeting: The AI pilot project lands, the demo goes well, someone says, “Let’s get this into production” — and then nothing happens. Weeks become quarters; the model still works but it never ships. It’s tempting to blame the stall on one team dropping the ball. But far more often, it’s built into the way the pilot was set up in the first place.

Walk any manufacturing plant this year and you’ll encounter the same scenario on the shop floor: a vision model that flagged defects beautifully in the pilot phase but never made it onto the production line, or a demand analysis pilot that impressed planners but never touched the enterprise resource planning (ERP) system.

The numbers are blunt about it. An MIT study of more than 300 deployments found that roughly 95% of GenAI pilots produced no measurable profit and loss impact, against an estimated $30 billion to $40 billion in spending. S&P Global reports that the average organization scraps 46% of their proof-of-concept projects prior to production. The RAND Corporation, looking at projects rather than pilots, has put the share that fail to deliver their intended value at more than 80%. Different lenses, same picture.

• ALSO READ → Beyond the pilot trap: How to build a manufacturing AI strategy that scales

What keeps pilot projects from scaling?

When we trace a stalled pilot back to its cause, the trail almost never ends at the model itself but rather with the operating gaps around it: OT data that was clean in the sandbox but doesn’t align with historical and ERP data on the plant floor, manufacturing execution system and ERP handoffs that break the moment the pilot leaves the lab, process changes that affect different shifts inconsistently, and a lack of governance and relevant metrics. Meanwhile, the sponsor’s enthusiasm cooled the week after the demo.

None of these are purely technology problems — and that’s exactly why a pilot that performs beautifully in the lab can die at scale while a simpler solution can thrive if it fits the workflow, has a clear owner and changes a business outcome.

The five seams where pilot projects tend to spring a leak

In reality, some pilots shouldn’t scale at all. A technically inferior tool that supports how people already work will beat a better one that asks them to change everything. We’ve watched stronger systems fail to gain traction because a weaker option was already available to users all day. Fit beats sophistication more often than anyone wants to admit.

What production-ready teams build before the model ships

Production-ready teams treat data preparation and integration as the main event. They design the workflow where the AI will operate — the hand-offs, exception paths and human checkpoints — before they fine-tune the model that operates in it. And they build governance into the architecture on day one, because everyone who bolts it on after an incident pays for it twice.

But production readiness can’t depend on every team relearning the same lessons from scratch, and too many organizations have nowhere for those lessons to live. That’s where an enterprise center of excellence (CoE) can help. In a multiplant environment, this matters even more: a corporate AI CoE has to coordinate with individual site leaders rather than dictate to them, or the lessons never make it past the plant gate.

NTT DATA helps clients create an AI CoE operating model led by a named senior leader and scoped as advisory-led capability building rather than a tool drop. The point of this structure is that it outlives any single model.

The agent question, asked honestly

Successfully scaled AI initiatives are led by teams that understand that not every problem is an agent problem, even when agents are the exciting answer. Deterministic, stable, rules-based work is still best left to ordinary automation. Agentic systems earn their extra complexity only where the work is genuinely multistep and full of judgment.

The market is pulling in both directions at once. According to Gartner®, “33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.” However, it also says: “Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.”*

One large metal recycling client got this right by being narrow on purpose, putting agents only into the specific logistics execution stages where multistep coordination genuinely paid off. Using agentic AI in the right place to get quantifiable results is the difference between the projects that scale and the 40% that get canceled.

Remember this for your next pilot review

One last thing the survivors share: They don’t inherit software’s old lifecycle unchanged. Models drift and degrade, so continuous evaluation, monitoring and human checkpoints must be part of the process from the start. Underneath all of it sits a cost structure that grows with the portfolio, which is what the next installment of this blog will explore.

Before that, here’s a question worth carrying into your next pilot review: Are we running this as a technology experiment or as a test of how we actually work?

Take your single most promising pilot project and pre-mortem it before you spend anything scaling it:

1. Score the five seams from 1 (low readiness) to 5 (high readiness): data foundation, change readiness, governance, outcome metrics, total cost of operation and executive sponsorship. Be uncomfortable about it.
2. Ignore the average; find the lowest. That seam is where the pilot is likely to fail in production, no matter good the demo looked.
3. Fix that one first — or don’t scale yet. The seam was always the biggest risk here.

Perhaps the most useful way to think about a pilot project is not as a test of technology but as a test of reality. The model has to work, of course, but so do the data, the process, the governance, the economics and the people expected to use it. If any one of those breaks under pressure, the pilot has already given you its answer.

This is part two in a three-part series for manufacturing CIOs, CDOs and operations leaders on moving from AI ambition to AI advantage — deliberately, defensibly and at scale. Also read part one. Part three will be published soon.

• ALSO READ → For manufacturers, AI value depends on the cloud decisions being made now

* Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. June 25, 2025.
GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.