The Data Was Always There. It Was Just Wrong.

#aiadoptionintraditionalin

Industrial AI isn't failing because the models aren't good enough — it's failing because companies spent decades building the digital equivalent of a house of cards, and now they're surprised the AI can't play on top of it.

A senior engineer at a midsize European petrochemical refinery once described to a colleague the experience of preparing their operational data for an AI predictive maintenance pilot. Six months of work. Teams pulling sensor readings from equipment installed across three different decades, stored in five incompatible formats, some still transcribed from paper logs by hand. At the end of it, they had a dataset. What they didn't have was a model that could do much with it, because the underlying records were riddled with duplicate entries, mislabeled columns, and timestamps that didn't align across systems. The vendor's demo, naturally, had looked flawless.

This is the story of AI in traditional industry that no one puts on the conference keynote slide. Not a story of technological failure — the models themselves are increasingly capable — but a story of an organizational reckoning that was thirty years in the making. The uncomfortable claim: the real barrier to AI in retail, energy, and manufacturing is not artificial intelligence. It's the data infrastructure that these industries quietly agreed to neglect for decades, and the organizational habits that grew around that neglect.

The Infrastructure That Was Never Built

Many manufacturing enterprises grapple with outdated legacy systems, widespread data silos, and a lack of integrated data governance — limitations that often result in datasets that are noisy, incomplete, or poorly contextualized, requiring laborious and costly pre-processing. This is not new information. It has been the conclusion of every serious industry report for the better part of a decade. What's new is that AI is making the problem visible in a way that ERP consultants and middleware vendors never quite managed to.

Decades of reliance on monolithic systems — mainframes, fragmented ERPs — are now among the greatest barriers to resilience, growth, and compliance. In the energy sector, the situation compounds in a specific way. A single oil, gas, refining, or petrochemical facility can have thousands of sensors measuring everything from temperature and pressure to velocity and viscosity, yet facilities make operating decisions using less than 8% of the data available to them. It's not that the data doesn't exist. Operators already collect most of it — they just can't combine the sensor readings, engineering documentation, and process physics fast enough to actually do anything with it.

The ERP problem specifically deserves more attention than it gets. A recent deployment study on enterprise ERP systems found something that vendors would prefer stayed quiet: initial LLM deployment attempts failed despite sophisticated prompting and retrieval mechanisms, and system performance became acceptable only after implementing automated cleaning — confirming that generative AI cannot compensate for fundamentally corrupted data. This finding directly contradicts vendor claims that LLMs can "clean data on the fly" through contextual understanding.

That last sentence should be printed on a banner and hung in every vendor booth at Hannover Messe.

The J-Curve Nobody Mentions in the Strategy Deck

To be fair to the industries struggling here, the early productivity hit from AI adoption is a documented phenomenon, not a sign of uniquely bad execution. Recent research on AI adoption at U.S. manufacturing firms reveals a more nuanced reality: AI introduction frequently leads to a measurable but temporary decline in performance followed by stronger growth in output, revenue, and employment — a "J-curve" trajectory that helps explain why the economic impact of AI has been underwhelming at times.

The harder finding is who takes the deepest hit at the bottom of that J. The negative impact of AI adoption was most pronounced among established firms — organizations with long-standing routines, layered hierarchies, and legacy systems that resist unwinding. In other words, the companies with the most operational history, the longest track records, and the most accumulated process knowledge are exactly the ones that suffer most when they try to bring AI into the workflow. Not because they're being managed badly, but because their past success literally got in the way.

Organizational memories are often clouded by many failed or painfully stretched technology rollouts — ERP systems, safety tools, telematics systems, and so on. People wonder whether the AI-tools wave is another fad worth waiting out. When you look more closely, the real blocker is change fatigue, not an aversion to technology.

That matters. The floor-level operator at a manufacturing plant who rolls his eyes at the new "AI dashboard" isn't a Luddite. He's someone who watched three previous digital transformation initiatives arrive, produce PowerPoint decks, and disappear. His skepticism is rational. And improved accuracy or productivity boosts mean little to front-line operators, who care more about customer escalations, rework, and operating costs.

Pilots Everywhere, Production Nowhere

Here is the number that should be tattooed on every CDO's forearm before their next board presentation: despite billions in enterprise AI spending, a 2025 study from MIT's NANDA initiative concluded that 95% of generative AI pilot programs fail to produce measurable financial impact — with failures stemming not from model quality but from poor workflow integration and misaligned organizational incentives.

Most companies are stuck in pilot mode. While 88% of organizations use AI in at least one function, only one-third have begun to scale their AI programs at the enterprise level. Bayer, to their credit, gave this syndrome a name that has since spread across the industry: "pilotitis" — the habit of continuous piloting that leaves management increasingly impatient, wondering how to get out of the rut.

Only 26% of organizations have moved beyond proof of concept, and 42% abandoned the majority of their AI initiatives before production. For energy and manufacturing companies, those numbers are likely worse — because they face the additional friction of safety approval chains, functional safety standards, and process engineers who rightly point out that a probabilistic recommendation engine shouldn't be trusted to schedule a pressure vessel inspection without human sign-off.

That last concern, incidentally, is legitimate. Large organizations prefer human-in-the-loop solutions even when full automation is technically feasible — because the perceived risk of autonomous AI outweighs the efficiency gains. In a refinery, this isn't overcaution. This is correct risk management. The AI vendors who don't understand this are the ones pitching to procurement rather than to safety engineers.

When the Shop Floor Meets the Slide Deck

The gap between what AI can demonstrably do and what actually ships to production has a specific texture in industrial settings. Consider what happened when a research team embedded with an energy firm to assess generative AI readiness across organizational functions in early 2025. They identified 41 use cases consolidated into six categories — reporting, RAG-based solutions, predictive maintenance, anomaly detection, budgeting, and forecasting — with three priorities emerging across functions: automation of reporting, predictive maintenance to minimize downtime, and enhanced forecasting for planning.

All sensible. All things that AI can genuinely help with. And yet the same team found that the smart energy sector encounters numerous obstacles in adopting AI, including insufficient or poor-quality data, technical infrastructure issues, a lack of skilled professionals, integration difficulties, and legal compliance concerns.

The missing expertise problem is particularly acute. Missing expertise is the dominant reason that firms do not adopt AI today. A data scientist who can build a transformer model is not the same person who can navigate a plant historian, understand HAZOP documentation, or explain to a process engineer why the model flagged a false positive on a temperature sensor three months after the bearing had already been replaced. Industrial AI needs people who are bilingual in a way that's genuinely rare — who speak both the language of gradient descent and the language of a distillation column.

Retail faces a slightly different version of the same problem. Global retailers lose more than $1.8 trillion per year due to inefficient demand forecasting and inventory disconnect, and 62% continue to rely on manual or spreadsheet-based reorder processes. The technology to fix this exists. The organizational will and the clean data to feed that technology are substantially harder to find. Companies must manage more SKUs and fragmented orders across multiple fulfillment channels, maintain real-time inventory visibility, handle return rates rising to 40% in sectors like fashion, and allocate inventory dynamically across online and offline channels. These are real operational problems that AI can address — but only with the kind of end-to-end data integration that most retailers haven't actually built.

The Counterargument, Taken Seriously

There's a version of this critique that goes too far, and it's worth naming. Some deployments work, and work well. Shell's collaboration with C3.ai produced a platform that continuously monitored over 10,000 critical refinery assets, analyzed approximately 20 billion data points weekly, successfully identified two imminent and critical equipment failures well in advance, and resulted in estimated savings of approximately $2 million. Michelin has reportedly identified more than 200 AI use cases generating meaningful annual ROI. Barnes Group, an aerospace and industrial components manufacturer, identified, de-duplicated, and indexed all of its product specification documents using generative AI — allowing service technicians to access immediate, accurate information and yielding a five-times return on the AI investment in its first year.

The J-curve exists, but early AI adopters showed stronger growth over time. And the research on what separates successful deployments from failed ones is fairly clear. McKinsey reports that top performers are nearly three times more likely to fundamentally redesign workflows as part of their AI efforts — 55% of high performers redesigned workflows around AI versus only 20% of other companies. The companies that treat AI as a tool to bolt onto existing processes lose. The ones that treat it as a reason to question the process itself have a fighting chance.

The counterargument also holds that the data infrastructure problem is solvable. Modernizing legacy systems isn't a one-time project — it's a continuous journey that can begin with relatively simple integrations between siloed systems. True. It's also, sometimes, a decade-long journey that starts with a three-year SAP migration that goes sideways in year two.

The Real Problem Is What Got Celebrated

Here's the sharpest version of the claim: traditional industries aren't failing at AI. They're being penalized for decades of decisions that were, at the time, entirely rational. Running a thirty-year-old process historian because it works is sensible. Accepting a patchwork ERP because the upgrade cost was too high is a reasonable CFO call. Keeping paper logs as a backup because the SCADA system has crashed before is just good field engineering.

None of those decisions were wrong when they were made. All of them collectively created an organization that is genuinely, structurally difficult to apply modern machine learning to. A growing body of reports and academic studies has highlighted that many firms have struggled to generate meaningful returns from their AI initiatives — the "AI productivity paradox." But paradox implies mystery. There's no mystery here. As one industrial company CEO put it: "The technology is not the hard part. It's the changing-the-company part that's hard."

The AI vendors landing in Houston or Dortmund with their demonstration environments and clean synthetic datasets are not wrong about what the technology can do. They're just performing a demo on a surface that doesn't look much like the actual factory floor. The floor has forty years of grime on it, metaphorically speaking, and also sometimes literally. No amount of model sophistication bridges that gap without someone first deciding to clean the floor.

The companies that figure this out stop asking "how do we implement AI?" and start asking "what would need to be true about our data before AI could help us?" Those are very different questions. Most AI strategy documents, for the record, answer neither of them.