Why Does Industrial AI Fail? The Data Problem 3 Manufacturers Already Solved

TLDR: Industrial AI fails in most enterprises because organizations automate existing workflows instead of redesigning the data infrastructure that AI requires. When manufacturers solve the dark data problem, as ARM, PepsiCo, and Siemens customers have done, the gains are extraordinary: millions of hours saved, significant capital expenditure reductions, and quality improvements impossible with human operators alone.

Best For: COOs, VP Operations, and transformation leads at manufacturing, logistics, and distribution companies who have run at least one AI pilot and are now asking why the results aren't translating to enterprise-wide ROI.

Industrial AI is the application of artificial intelligence to physical, operational environments: factory floors, transportation networks, energy systems, and industrial equipment. Unlike AI deployed in software workflows, industrial AI must meet near-perfect accuracy standards, integrate with decades-old machinery, and act autonomously in environments where errors carry physical and financial consequences. Most industrial AI initiatives fail not because the technology is unproven but because organizations skip the one prerequisite that makes it work: connecting the data those systems actually run on.

Why Does Industrial AI Fail to Deliver ROI?

Industrial AI fails to deliver ROI in most organizations because they treat AI as an optimization layer on top of existing workflows rather than a reason to redesign how data flows through the enterprise. According to MIT Sloan Management Review and BCG, who have studied corporate AI adoption for over four years, 90% of organizations do not realize significant financial benefit from AI. McKinsey's 2025 State of AI research put the number more bluntly: 94% of respondents report not seeing significant value from their AI investments.

Three patterns explain why industrial manufacturers specifically struggle.

The Automation Trap: Optimizing Yesterday's Workflows

When Dan Diasio, Global AI Consulting Leader and Americas Consulting CTO at EY, appeared on the Emerj AI in Business podcast in March 2026, he named the single most common mistake he sees in enterprise AI: organizations using AI to automate the processes they already have rather than redesigning those processes from the ground up. "AI adoption is forcing a shift from automating yesterday's processes to redesigning the enterprise for differentiation and growth," Diasio said. The distinction sounds subtle but drives fundamentally different outcomes.

A manufacturer that deploys AI to speed up its existing inspection process by 20% is capturing a fraction of what's available. A manufacturer that redesigns its inspection process around AI, as Siemens has done with customers in food and beverage, electronics, and metals, replaces the process entirely. That's the gap between incremental efficiency and structural competitive advantage.

The Dark Data Problem: When AI Can't Connect the Dots

Samir Desai, who leads AI product strategy at Siemens Digital Industries Software, named this problem "dark data" in Siemens' Future-Ready Podcast transcript published June 2026. The concept captures what most manufacturers actually face: they know they have enormous amounts of operational data sitting in ERP systems, PLM platforms, CRM tools, and homegrown shop-floor software. They know it's connected somehow. They just don't know what the connections are.

Take a car manufacturer using Siemens' Teamcenter PLM system alongside separate ERP, CRM, and production databases. When an engineer changes a design, the downstream impact on sourcing, manufacturing, and cost is real, but uncovering it traditionally requires pulling subject matter experts from multiple departments into a room and spending days or weeks on analysis. AI cannot reason across data it cannot see. The data fabric concept, which Siemens implements through its Rapidminer Graph Studio (acquired through Altair), uses knowledge graphs and ontologies to map the semantic relationships between these siloed systems. Once those connections exist, AI can traverse them.

The Visibility Trap: Measuring the Wrong Things

Diasio also identified what he called the "visibility trap": organizations that focus on being seen as AI-forward rather than measuring whether AI is actually generating business value. This shows up in how companies report AI metrics. Counting the number of AI tools deployed, the number of employees who attended AI training, or the number of pilots launched tells you nothing about operational impact. According to BCG's 2025 AI value realization research, only 4-5% of companies are generating substantial value from AI. The rest are managing the appearance of progress.

How 3 Manufacturers Already Solved the Data Problem

When manufacturers do solve the data connectivity problem, the results are documented and specific. These are not projections. They are production outcomes from real industrial deployments.

ARM Semiconductor: From 5 Billion Simulations to 2,000

ARM, the chip design company whose architecture powers most of the world's mobile devices, operates under Six Sigma quality standards. One defect per billion chips is the ceiling. To meet that standard, ARM was running approximately five billion simulations to verify chip designs before committing to manufacturing, a process that took months.

Siemens introduced a generative AI model into ARM's chip design verification workflow through its Solido product, a platform for AI-accelerated circuit simulation. The result, described in Siemens' June 2026 Future-Ready Podcast transcript, was a reduction from five billion simulations to approximately 2,000. That is a 2.5 million times improvement. The time required dropped from months to hours. ARM did not simply automate what it was already doing. It redesigned the verification workflow around an AI model that could predict failure distributions without running most of the simulations.

Before an AI readiness assessment that surfaces the specific use case (chip verification), the ROI infrastructure (simulation cost and cycle time), and the data requirements (circuit behavior datasets), this deployment would not have happened. ARM and Siemens identified the specific workflow where AI could replace not just speed up a manual process.

PepsiCo: 15% Capital Expenditure Reduction Through Digital Twin

PepsiCo partnered with Siemens and Nvidia to build what Siemens calls an "industrial metaverse" for factory planning. Using Nvidia GPUs to run Siemens' digital twin composer, PepsiCo can now simulate factory layouts, production sequences, and equipment configurations before committing to physical construction or equipment purchases.

The outcome, documented in the same Siemens Future-Ready Podcast transcript from June 2026, was a reduction of up to 15% in capital expenditure. For a company operating at PepsiCo's scale, across hundreds of manufacturing and distribution facilities globally, that figure is substantial. The critical enabler was not the AI model itself but the willingness to build a connected data environment (factory design data, equipment specifications, production requirements) that the digital twin could reason across.

Visual Quality Inspection: Replacing the Conveyor Belt Operator

Matthias Loskyll, who leads manufacturing AI strategy at Siemens, described a third class of industrial AI deployment that is already in production at scale: automated visual quality inspection. The baseline case involves a camera above a conveyor belt and a human operator whose job is to watch products pass and pull defective ones. Across most of manufacturing, this is still a manual role performed by people working eight-hour shifts in conditions that make sustained attention difficult.

Siemens' Inspector product and Visual Inspection Cockpit platform automate this task across metals, electronics, plastics, and food and beverage. In the more complex variant, the inspection system detects a defect, identifies which upstream process parameter likely caused it, and sends a control signal back to the PLC (programmable logic controller) governing that step. The loop closes in real time. Loskyll describes this as "physical AI": AI that perceives, reasons, and acts autonomously within the physical production environment.

For a primer on AI use cases in manufacturing and distribution, many of these patterns apply across industries beyond electronics and food.

What Makes Industrial AI Different From Enterprise Software AI

Industrial AI is not simply enterprise AI installed in a factory. The operating requirements are categorically different, and understanding those differences is the precondition for any deployment strategy that works.

Accuracy Requirements in Physical Environments

Peter Koerte, Siemens' Chief Strategy and Technology Officer, explained on MIT Sloan's Me, Myself, and AI podcast in April 2026 why industrial AI must reach near-perfect accuracy. In a consumer recommendation system, a 90% accuracy rate is commercially acceptable because the cost of a wrong suggestion is low. In a factory, the cost of an incorrect maintenance prediction, a missed defect, or a mis-sequenced process can be equipment failure, product recall, or physical injury.

Koerte cited predictive maintenance in Siemens' rail operations as an example. Siemens AI predicts train door failures days in advance, which allows maintenance crews to schedule repairs before a door fails mid-service. An 80% accurate model is not useful here. The model must reach a threshold that justifies operational trust, or operators will override it within weeks and the deployment stalls.

Domain Data vs. Generic AI

Generic AI models trained on public data do not understand your production environment. Siemens' AI for chip verification works because it was trained on circuit simulation data. Its visual inspection product works because it was trained on defect images specific to the materials and defect types each customer produces. The domain specificity of industrial AI is not an implementation detail; it is the primary determinant of whether the model works at the accuracy threshold the environment demands.

This is also why data connectivity matters so much. A Siemens building energy optimization model that Koerte described, which reduces energy consumption in commercial buildings by learning usage patterns, works only when it has access to sensor data from across the building's systems. Siloed HVAC data without lighting data, occupancy data, and weather data produces a model that cannot identify the full optimization opportunity.

Physical AI: When AI Takes Action, Not Just Recommendations

Loskyll introduced a concept that separates the most advanced industrial deployments from the majority: physical AI. Most enterprise AI produces a recommendation that a human then acts on. Physical AI produces an action directly, integrated into the control system of the machine or production line. The advanced process control systems Siemens deploys in chemical production, for example, continuously tune process parameters based on real-time sensor values. No human is in the loop for routine optimization decisions.

This level of integration requires a data foundation that most manufacturers have not yet built. Before implementing AI without replacing legacy systems, manufacturers need to solve connectivity at the data layer first. That is the prerequisite Siemens' customers who achieve results have consistently addressed.

What Skeptics Get Wrong About Industrial AI

Operations leaders who have watched AI projects fail once tend to apply lessons that don't transfer. Here are the three most common objections and why they lead teams in the wrong direction.

"We tried AI in our factory and it didn't work." In most cases, the pilot failed because the AI was given disconnected or insufficient data, not because the technology is immature. When Loskyll describes brownfield factories with machines of different ages and vendors, he is describing exactly this situation. The answer is not to stop trying AI but to solve data connectivity first. This is what the data fabric approach addresses: building the infrastructure layer before deploying the application layer.

"Our data is too messy to start with AI." This confuses a symptom with a prerequisite. Every manufacturer Siemens has worked with had messy data before the engagement. The question is not whether the data is clean; it is whether the organization is willing to build the semantic layer (the ontology, in Siemens' framing) that reveals how data items across systems relate to each other. That mapping exercise surfaces the mess and creates the structure AI needs.

"We'll wait until AI matures." According to BCG's 2025 widening AI value gap research, organizations that lead in AI adoption now generate twice the revenue increase and 40% greater cost reductions than laggards. The technology is mature enough to deliver the ARM and PepsiCo results today. The gap between leaders and laggards is not technology maturity; it is organizational commitment to building the data infrastructure.

The 5 Conditions That Separate Industrial AI Wins from Stalled Pilots

The evidence from Siemens' deployments at ARM, PepsiCo, and manufacturing customers across 20-plus verticals points to five conditions that predict whether an industrial AI initiative will deliver measurable ROI or stall.

This framework maps directly to what an AI readiness assessment for manufacturing should surface before any deployment begins. If a manufacturer cannot check all five conditions, the pilot will produce results that don't scale.

McKinsey's research on AI in industrial processing plants estimates that AI could generate $275 to $460 billion annually for global manufacturing and supply chain, and that manufacturers applying AI are three times more likely to improve their key performance indicators. The gap between that potential and the 94% of organizations that aren't seeing significant value is not a technology gap. It is a data infrastructure and process design gap.

BCG estimates that only 5% of organizations are in the "future-built" category where AI is generating substantial value. Those organizations have one structural characteristic in common: they redesigned their operating model around AI rather than applying AI to their existing operating model. The EY, Siemens, and MIT Sloan evidence all point to the same conclusion. The manufacturers who solve the data problem first, and accept that solving it requires building something new rather than deploying something on top of what already exists, are the ones generating the documented results.

Industrial AI is not a technology decision. It is a design decision about how data flows through your operations. The three manufacturers examined in this article made that design decision explicitly. Most enterprises haven't.

Frequently Asked Questions

What is industrial AI?

Industrial AI is the application of AI to physical, operational environments including factories, transportation networks, energy systems, and industrial equipment. Unlike enterprise software AI, industrial AI must meet near-perfect accuracy thresholds, integrate with existing industrial control systems, and in advanced deployments, take autonomous action through connected machinery.

Why does industrial AI fail more often than enterprise software AI?

Industrial AI fails more often because the data requirements are more demanding. Factory environments have fragmented data across PLM systems, ERP platforms, and shop-floor controllers, often from different vendors and decades of accumulation. According to MIT Sloan and BCG research, 90% of organizations fail to realize significant financial benefit from AI, and data connectivity is the primary structural cause in manufacturing.

What is the dark data problem in manufacturing?

Dark data is operational data that exists within an enterprise but cannot be connected or reasoned across because it sits in isolated systems. A manufacturer may have design data in its PLM, supplier data in ERP, and production data in shop-floor systems, none of which are semantically linked. AI deployed on any single silo produces limited value because it cannot see how a change in one system affects others.

What is a data fabric for industrial AI?

A data fabric is an infrastructure layer that maps relationships between data held in different systems using ontologies and knowledge graphs. In Siemens' implementation, the Rapidminer Graph Studio platform builds semantic links between PLM, ERP, and factory systems, enabling AI models to traverse those connections and surface multi-system insights. It replaces siloed analytics with enterprise-wide reasoning.

How did ARM reduce simulations from 5 billion to 2,000 using industrial AI?

ARM deployed Siemens' Solido AI platform for chip design verification. Rather than running all simulations, the AI model predicts which circuits are most likely to fail, reducing the required simulations from 5 billion to 2,000, a 2.5 million times improvement. Verification time dropped from months to hours. This required redesigning the verification workflow around AI, not simply accelerating the existing process.

What was PepsiCo's result from Siemens' digital twin technology?

PepsiCo partnered with Siemens and Nvidia to build an industrial metaverse for factory planning. Using Siemens' digital twin composer running on Nvidia GPUs, PepsiCo can simulate factory layouts and production configurations before committing to physical investment. The result documented in Siemens' June 2026 podcast was up to 15% reduction in capital expenditure across its manufacturing operations.

What is physical AI in manufacturing?

Physical AI takes autonomous action by integrating directly with industrial control systems, rather than producing a recommendation for a human to act on. Siemens' advanced process control for chemical manufacturing modifies process parameters in real time through the PLC governing that step. Their Visual Inspection Cockpit sends defect signals back to upstream production steps automatically, with no human in the loop.

What accuracy does industrial AI need to reach?

Industrial AI must reach the accuracy threshold at which human operators trust it enough to stop overriding it. Siemens' CSTO Peter Koerte notes this differs by application: predictive maintenance for train doors requires accuracy high enough to schedule repairs days in advance with confidence. Visual quality inspection must detect defect rates that human operators cannot sustain across a full shift. The threshold is defined by the cost of a wrong answer in the specific environment.

What is the automation trap in enterprise AI?

The automation trap is the tendency to apply AI to existing workflows rather than redesigning those workflows around AI's capabilities. According to EY's Dan Diasio on the Emerj AI in Business podcast, leading organizations are redesigning the enterprise for differentiation and growth, not simply making current processes faster. The trap produces marginal efficiency gains rather than structural competitive advantage.

How do successful manufacturers approach data connectivity before AI deployment?

Successful manufacturers standardize their data infrastructure before deploying AI use cases. Siemens recommends building a common connectivity layer reused across deployments, establishing a semantic ontology that maps relationships between PLM, ERP, and factory data, and piloting AI use cases only after that foundation exists. This prevents the pattern of building a new integration for every AI use case and accumulating technical debt that blocks scaling.

What separates industrial AI leaders from laggards?

BCG's 2025 research found that AI leaders generate twice the revenue increase and 40% greater cost reductions than laggards. The structural difference is that leaders redesigned their operating model around AI rather than applying AI to their existing model. In manufacturing, this means solving data connectivity as a precondition, not treating it as a follow-on integration project.

What is predictive maintenance and how does it work in industrial AI?

Predictive maintenance uses AI to forecast equipment failures before they occur based on sensor data, enabling maintenance crews to schedule repairs during planned downtime rather than responding to breakdowns. Peter Koerte described Siemens' rail application: the system predicts train door failures days in advance. Compared to time-based maintenance schedules, predictive maintenance reduces unplanned downtime and extends equipment life.

How does AI for visual quality inspection work?

AI visual inspection uses cameras and AI models to detect defects, missing components, or contamination on production lines. In basic deployments, the system flags defects for human review. In advanced deployments like Siemens' Visual Inspection Cockpit, the system sends real-time control signals to upstream production steps, closing the quality loop automatically. This replaces human operators who cannot sustain consistent attention across full production shifts.

Should a manufacturer build its own AI or use an embedded vendor solution?

Most manufacturers should use embedded vendor solutions for initial industrial AI deployments. According to Siemens' deployment model, the most successful path is running proof-of-concept deployments in your top three use cases with vendor support before any licensing commitment. This surfaces whether the solution fits your specific data environment. Fully custom AI development requires domain data, model development capability, and integration depth that most manufacturers do not have internally.

What is the first step for a manufacturer serious about industrial AI ROI?

The first step is mapping your data environment before choosing an AI use case. This means identifying what data exists across PLM, ERP, shop-floor, and maintenance systems, how those systems are (or are not) connected, and where your highest-cost manual processes actually sit. An AI readiness assessment focused on manufacturing typically surfaces data gaps that would have killed any AI deployment attempted without that diagnostic first.

What does EY recommend for enterprises that want to move from AI experiments to enterprise value?

EY's Dan Diasio recommends three aligned shifts: mindset (from AI as a productivity tool to AI as a business redesign lever), skill set (building change management and AI integration capability alongside technical skills), and tool set (selecting AI investments that connect to your core operating processes, not peripheral workflows). Enterprises that align all three see returns; those that invest only in tools without changing the operating model do not.