Why 95% of Industrial AI Projects Fail Before They Reach the Factory Floor

The pitch is always compelling: deploy artificial intelligence on the factory floor, and watch productivity climb while costs fall. A KPMG analysis found that manufacturers adopting digitally enabled aftermarket and field services have targeted 30 to 40 percent reductions in truck rolls and 15 to 30 percent reductions in service part inventory. A 2025 Fluke Corporation study put the cost of unplanned downtime at up to $207 million weekly across U.S. manufacturers alone. The business case, on paper, writes itself.

And yet, according to widely cited MIT research, up to 95 percent of enterprise AI projects fail to produce meaningful return on investment. In manufacturing, the reasons are not algorithmic. They are architectural: fragmented data locked in legacy systems, unstructured service records spanning decades and multiple languages, and organizational silos that keep the people who understand the problems separated from the people building the models.

Lanfu Fu has spent the past several years working at the intersection of those failures. As AI Solutions Delivery Manager at Videojet Technologies, a Veralto operating company, he leads initiatives to embed AI into field service diagnostics, production scheduling, and institutional knowledge management.

His background is unusual for the role: a Chinese CPA (CICPA) credential earned while working at PwC in Shanghai, then a STEM MBA from the University of Michigan’s Ross School of Business with concentrations in data analytics and finance. Before moving into manufacturing technology, he spent seven years in audit and business development, including due diligence work at PwC that surfaced a multi-million-dollar asset overstatement and a role leading financial restructuring across dozens of subsidiaries at HNA Group, where his initiatives cut tax burden by roughly $2 million and reduced labor costs by 40 percent. The through-line from audit to AI deployment is the same discipline: start with what you can measure, and don’t deploy what you can’t defend.

That financial and technical fluency shapes his central thesis: most industrial AI projects don’t fail because the models aren’t good enough. They fail because the data isn’t clean, the governance isn’t in place, and the people closest to the problem weren’t involved in designing the solution.

The tribal knowledge time bomb

In field service organizations of any meaningful scale, critical diagnostic knowledge concentrates in a small number of senior technicians. When those people retire, transfer, or are unavailable, the organization faces two compounding problems: blind dispatch, where technicians arrive at customer sites without adequate diagnostic context, and needle-in-a-haystack retrieval, where field engineers cannot locate relevant solutions among thousands of historical cases.

Fu built the business case for what became the Service Intelligence Engine at Videojet, an AI-powered diagnostic platform that combines historical service data analytics with structured troubleshooting workflows. The system analyzes tens of thousands of historical service work orders to flag probable root causes before a technician arrives on site, reducing reliance on any single expert’s availability.

The data layer nobody wants to fix

The gap between AI’s theoretical promise and its practical delivery almost always traces back to data. The problem is not volume. It is structure, or rather, the absence of it.

“Manufacturing environments generate enormous volumes of service records, but much of it is unstructured, multilingual, and inconsistently categorized,” Fu notes. Research in reliability engineering has demonstrated that building enterprise data taxonomies based on standards like ISO 14224, which distinguishes between failure modes and failure causes, can dramatically improve classifier accuracy. But the prerequisite is painstaking manual review of label inconsistencies, work that few organizations are willing to prioritize.

“If the labels are wrong, the AI will confidently give you wrong answers,” he says.

That conviction drove a hackathon project at Videojet, applying natural language processing to historical service work orders. The initiative used transformer-based text classification to parse raw, multilingual work order narratives into five structured output categories: issue classification at two levels of granularity, issue summary, troubleshooting steps, and resolution actions.

Across the dataset, the system achieved 87 percent classification consistency, demonstrating that unstructured field records spanning decades and multiple languages could be converted into a structured, searchable knowledge asset. The approach drew on Technical Language Processing, a framework developed at NIST for applying NLP methods to specialized industrial vocabularies where general-purpose language models often misinterpret domain-specific terminology.

When decades of knowledge become queryable

The hackathon addressed service data. But a parallel challenge existed on the production floor: scheduling knowledge accumulated over decades by long-tenured operators, locked in PDFs, spreadsheets, and individual memory.

Fu’s response was AskVin, a retrieval-augmented generation system that ingests over 940 production scheduling documents and serves as an interactive knowledge base at a North American manufacturing facility. The system covers production rules, tank configurations, changeover sequences, material constraints, operator qualifications, and demand patterns, essentially the full decision-making context a production planner needs.

“The goal was straightforward,” Fu explains. “A planner used to have to track down the one person who knew why a particular tank sequence mattered, and hope that person was on shift. Now they type the question and get an answer grounded in verified documentation.”

The system uses hybrid search, combining keyword matching with vector-based semantic retrieval, so that queries phrased in everyday language still surface technically precise answers. Fu designed explicit confidence thresholds into the architecture: when the system cannot find a grounded answer, it says so rather than generating a plausible-sounding guess. “I’d rather have a system that’s right 90 percent of the time and honest about the other 10 percent than one that’s confident 100 percent of the time,” he says.

The broader design principle is knowledge preservation. When institutional expertise is embedded in a queryable system rather than held in individual memory, the organization becomes less fragile. New planners gain access to decades of accumulated scheduling logic from their first day, while experienced operators are freed from routine interruptions to focus on exception handling.

Making the business case

The financial discipline gap is where many promising AI initiatives stall. Organizations invest in pilots, produce encouraging demos, and then struggle to justify full deployment because they never established a rigorous cost baseline.

“I insist on building a clear cost-to-serve model first,” Fu explains. “What does it cost today when a technician can’t resolve an issue on the first visit? What does it cost when a planner makes a scheduling decision based on incomplete information?” His audit background shows in the method: every initiative begins with a measurable baseline and a specific target, whether the domain is service resolution, production scheduling, or data processing. The question is never “can AI do this?” but “what does the current state cost, and what is the minimum improvement that justifies the investment?”

From scheduling chaos to measurable results

The most compelling evidence that industrial AI can deliver comes not from vendor benchmarks but from production environments where the stakes are real. Against the backdrop of the Fluke Corporation’s finding that more than half of U.S. manufacturers experience unplanned downtime, the pressure on scheduling accuracy is acute.

At Videojet, Fu leads the C3 AI production scheduling initiative, applying constraint-based optimization to determine which orders run on which equipment, in what sequence, with which operators, while balancing material availability and customer delivery commitments at a North American facility. The system is designed to eliminate hundreds of hours per year of manual scheduling work and reduce error rates in production sequencing. “Sometimes, the highest-impact improvement is making sure the right information reaches the right person before they make a costly decision,” he notes.

The same pragmatism applies to field service technology adoption. Research from Field Service East found that 58 percent of field service organizations using augmented reality encountered difficulties integrating the technology into legacy systems. Rather than forcing adoption, Fu developed a structured evaluation framework that assesses each training technology along two dimensions: operational criticality and delivery practicality.

“The most impressive demo adds no value if it addresses the wrong skill gap or can’t be delivered at scale within the constraints of a manufacturing operation,” he says. The framework led his team to design AR as an augmentation layer, available when technicians need visual guidance on complex procedures but never mandated, so that adoption happens organically rather than through mandates.

Bridging the adoption gap

If the MIT research on AI failure rates reveals anything, it is that the bottleneck is organizational, not algorithmic. Poor integration, data fragmentation, and insufficient collaboration between technical and operational teams account for far more failed projects than model performance.

Rather than mandating AI tool usage, Fu designs rollouts that let practitioners experience value firsthand. “Framing it as ‘giving you superpowers for the tedious parts so you can focus on the hard, interesting problems’ drives adoption,” he explains.

The approach is deliberate: start with the use case where the pain is most obvious, let the first users become advocates, and expand from there. When AskVin launched at the Elk Grove facility, the initial pilot group was the scheduling team that had the most to gain. Their early feedback shaped the system’s confidence thresholds and search behavior before broader rollout.

What the next five years look like

The manufacturing industry is transitioning from static information retrieval toward dynamic, AI-driven execution workflows. Agentic AI platforms purpose-built for industrial environments are emerging, capable of serving as autonomous decision-support systems that any employee can query in natural language, regardless of technical background.

“Within five years, AI agents will handle the routine pattern-matching across service records, maintenance logs, and scheduling rules,” Fu says. “The question is whether organizations will have built the data architecture and the human literacy to actually use them.”

That may be the most important insight for manufacturing leaders weighing their next technology investment. The 95 percent failure rate will not decline because models improve. It will decline when organizations treat data architecture, cross-functional governance, and human-centered design as prerequisites rather than afterthoughts, and when they invest as deliberately in the people who work alongside AI as they do in the AI itself. The organizations that built systems like AskVin today, not because the technology was perfect but because the institutional knowledge could not wait, are the ones most likely to be ready.