The constant buzz is that AI is going to transform every business, every industry, and every job function. The process manufacturing industry is no different; as an industry that operates on tight margins, companies are always looking for a way to run better, faster, smarter. The opportunity for transformation is huge, but the opportunity for failure is just as large. The truth is most AI initiatives in process manufacturing will fail, and they will fail quietly. Tools will get selected, pilots will launch, and somewhere in the first six months, confidence will erode. Recommendations will look plausible but feel wrong. Teams will stop trusting the outputs. Projects will stall or get shelved. Leadership will scratch their heads wondering what went wrong.
The problem is rarely the AI. AI can only operate and provide value based on the underlying data beneath it. When that data is fragmented, incomplete, or never captured in the first place, AI has nothing reliable to work with. It produces confident answers to the wrong questions, and a confident wrong answer is more dangerous than no answer at all.
Gartner estimates that 60% of AI projects will be abandoned through 2026 due to data readiness gaps; so, what are the successful 40% doing differently? They started by building a reliable data foundation, then applied AI to amplify what was already working.
The examples below illustrate how data breaks down in process manufacturing and gas distribution — and what fixing each one requires.
AI-driven production scheduling, demand forecasting, and replenishment planning all start with inventory counts. If those counts are wrong, every recommendation built on them is wrong by the same margin — and the system will apply that error with full confidence.
Lewis Chemical’s inventory accuracy was 89% before deploying Datacor WMS. After implementation, their accuracy improved to 99.5%. That difference may look small on paper, but the 10.5 points between them is where AI recommendations go wrong. Every production run scheduled, every raw material committed, every delivery promised against that missing inventory is a decision built on data that doesn't reflect reality.
Most process manufacturers have experienced this without framing it as a data quality problem. They call it a scheduling problem, a stockout problem, a planning problem. The symptoms differ; the root cause is the same: the system of record does not match the physical operation.
A substantial portion of capital in gas distribution and chemical manufacturing sits in returnable containers — cylinders, totes, intermediate bulk containers (IBCs), bulk tanks — that were never serialized and have no event history in any system. There is no record of where they are, how long they have been at a customer site, or whether they have come back. They exist as an aggregate count in a spreadsheet, if at all.
When an asset has no event record of its lifecycle, there is no way to track billing effectively, and the revenue consequence of that lack of visibility accumulates silently over time. Argyle Welding Supply had a bulk tank at a customer site for over 2 years with no invoice attached to it because the asset simply did not exist in the system. After implementing TrackAbout with serialized tracking, rental income increased by 6–8%. Indiana Oxygen found the same problem running in the opposite direction: cylinders that had been written off as lost were still recoverable once they became visible. Tracked rental assets grew by 6%. These two outcomes — recovered revenue and recovered assets — point to the same root cause: the assets were never in the system.
Without a recorded chain of custody (fill, delivery, dwell time, return) there is nothing for AI fleet optimization, utilization forecasting, or rental revenue intelligence to work from. Building that event-level infrastructure is the operational prerequisite, and it has to come before any AI investment can pay off.
The third failure mode is the hardest to see because the data exists but is disconnected. A process manufacturer has a vast body of documentation that captures the how and what of their operations (batch records, compliance documentation, pricing history, and production logs); they’re just spread across disconnected systems, spreadsheets, and paper. They are inconsistent across facilities, incomplete across time, and not structured in a way that AI can aggregate or use for inference. A system with data like that often doesn’t have data that works.
AI models require dense, consistent historical records to identify patterns. What they get with manual entries or disconnected systems is patchwork: some records in a software system, some in spreadsheets, some on paper, and some missing entirely. The model trains on what it can find and produces outputs that look coherent. The problem is that coherence is shaped by the gaps as much as the data — and a plausible-looking wrong answer is harder to catch than an obvious one
The problem usually starts with fragmentation. Pariser Industries was running a discontinued legacy system alongside years of standalone spreadsheets and Access database reports — three sources describing the same operation in formats that couldn’t be reconciled. Datacor ERP replaced that tangle with a unified record set from manufacturing through accounting, and for the first time the business could see its own operation clearly.
Hubbard-Hall is the clearest illustration of what reliable records unlock. By integrating WMS and MES into a single operational record, they doubled production output without spending a dollar on new equipment. The capacity had always been there; it became visible only when the records became accurate enough to reveal it.
The results manufacturers expect from AI — better forecasting, recovered assets, revealed capacity — are already available to operations with a reliable data foundation. AI amplifies that foundation; it doesn’t replace it.
Fixing the data foundation is not a single initiative, and the three failure modes covered here are not an exhaustive list. They are illustrations of a pattern that runs across process manufacturing: when operational data is fragmented, incomplete, or never captured in the first place, AI has nothing reliable to work with.
The perception gap makes this harder to address: most organizations are more confident about their data than the evidence supports. In a 2026 study by Precisely and Drexel LeBow, 88% of business leaders reported adequate infrastructure for AI initiatives. But when the same respondents were asked whether data readiness was a barrier to AI adoption, only 43% said it wasn’t.
What changes the outcome is infrastructure that organizes operational data and makes it actionable — not as a precursor to an AI project, but as the way the business runs. Inventory events captured in real time. Assets tracked through their full lifecycle. Batch, compliance, and financial records held in a single integrated system. When that infrastructure is in place, the data an AI system needs already exists, already flows, and already reflects what is truly happening in the operation.
None of this requires AI to implement. All of it makes AI worth investing in.
For process manufacturers weighing the timing of that investment, our companion post — “Wait and See” vs. “Prepare and Lead”: 2026 Is the Year to Decide — makes the case for why 2026 is the year operational infrastructure becomes a competitive differentiator.
Datacor ERP and TrackAbout are built specifically for this foundation work — purpose-built for process manufacturing and gas distribution, where inventory complexity, returnable asset fleets, and regulatory record requirements make generic solutions a poor fit. The data layer they create is what makes AI investments in this sector pay off.