Most manufacturers have tried AI. Nearly 90% of projects don't survive the first year

Manufacturing has moved past the question of whether to try AI. The harder question is whether a pilot ever becomes the way a line actually runs.

That is the picture Sakshi Udavant lays out in Manufacturing Dive on October 15, talking to plant leaders at Pella and Hexion and to people who sell into the floor. Most manufacturers have experimented with AI or automation in some form. Wide-scale adoption, and a solid return on investment, are still the exception.

The numbers that stick are not the demo reels. A 2025 MIT report, cited in the piece, found that 95% of pilot programs for enterprise AI solutions showed no measurable impact. Ciarán O'Mara, co-founder and chief product and technology officer at Protex AI, puts the survival rate bluntly: nearly 90% of AI projects do not make it past their first year.

What plant managers say the tools are for

Jared Pfeiffer, a principal client partner at Augury and a former plant manager, frames the opportunity as knowledge, not headcount cuts. As experienced workers retire and plants stay short-staffed, AI can pull sensor data, operating conditions, maintenance history, SOPs, manuals and institutional knowledge into something a less experienced employee can use. For plant managers, he says that can mean shorter downtime, faster troubleshooting, better shift-to-shift consistency and less dependence on a handful of experts.

Dustin Katje, senior plant manager at Pella Corp., which has 20 locations, pushes the same line from inside the plant. "AI is about supporting people, not removing them from the process," he told Manufacturing Dive. His version of the win is less time mining data and more time driving improvements — reducing repetitive work so people can apply experience where it matters.

Pilots are easy. Scaling is the job

Katje's company is trying to run both sides at once: permanent cross-functional teams for new applications, for data and insights, and for scaling what already works across the manufacturing network. The goal he names is solutions that move across sites without being rebuilt each time. The starting point is not a model. It is where people lose time, where complexity creates mistakes, or where better information would help the customer.

Pfeiffer's warning is that some companies jump from "AI is exciting" to "we need AI everywhere" without defining the problem. The best business cases, he says, connect to metrics the plant already owns: downtime, throughput, yield, scrap, maintenance cost, labor utilization, quality or energy. The prioritization questions are concrete — where production is lost, which problems always need the best people, where operators waste time searching for information, which failures cost the most money.

A great model on fragmented, undocumented plant data just gives you wrong answers faster.

The context gap on the floor

Audra Kirkland, VP of manufacturing and IT at Hexion, a chemical manufacturer in Ohio, is blunt about vendors that chase manufacturing's data volume without understanding safety and compliance. In chemicals, she notes, getting AI and automation wrong is a safety risk, not only a budget miss.

O'Mara's diagnosis is less about the model and more about the job description. Plant managers who spent 20 or 30 years running lines with clipboards and stopwatches are now expected to speak to LLM and agent deployment in front of leadership. His advice is "retrofit and not rebuild": use tools that work with the systems already on the floor. Most of the industry, he says, is still in the pilot phase, and many of those pilots never produce a clear return.

The lesson for everyone else

When projects die in year one, O'Mara says it is usually not because the technology failed. It is because nobody figured out how to fit AI into how the plant actually works. The spend pattern he describes matches that diagnosis: money is rising toward the human capital that cleans siloed data, sets ownership and keeps people accountable for outputs — not only toward models and platforms.

That is a different program from buying another pilot. It starts with one recurring workflow — a shift handover, a work-order close, a quality hold — talking to the people who do it, and showing what still happens outside the MES. Only then does "AI everywhere" have a place to land.

What we still don't know

How Manufacturing Dive's sources define a "pilot" versus production use. How much of the MIT 95% figure is manufacturing-specific versus enterprise AI in general. Whether Pella's three-team model is common among multi-site makers or still rare. How often safety and compliance reviews kill projects that would otherwise look good on throughput alone.

The roundup's own frame is that AI can bridge retiring expertise and support people on the floor. The survival numbers say most projects never get that far.

This story draws on reporting by Sakshi Udavant in Manufacturing Dive on October 15, 2026.