Throughout my career, and through my work at Aperio, I’ve had a unique vantage point on industrial AI adoption. Working with some of the world’s largest energy, chemicals, and manufacturing companies, I’ve seen digital transformation efforts up close – the excitement of early pilots, the challenge of scaling, and what happens when AI meets the complexity of real operations. That experience has given me a clear view of what works, what doesn’t, and why.

One pattern has stood out: Most industrial AI projects don’t fail dramatically – they slowly fade away. AI models (such as those used for predictive maintenance, yield optimization, or demand forecasting) may look promising in pilot stage. However, soon the reliability of those models starts to degrade. In response, models are tuned or replaced but it’s like playing whack-a-mole – fix one issue and another surfaces elsewhere, because the underlying problem was never the model to begin with. At some point the initiative loses the confidence of end users, and then management. The initiative may still be technically running, but nobody’s relying on it anymore.

Tracking down the culprit

Apart from our own experience with this recurring pattern, the numbers back this up: Boston Consulting Group’s research, based on a survey of 1,000 senior executives across 59 countries, found that 74% of companies have yet to demonstrate tangible value from their AI investments. RAND Corporation’s analysis puts the AI project failure rate above 80% – roughly double that of non-AI technology initiatives. S&P Global’s 2025 survey found that 42% of companies abandoned most of their AI work that year, up from 17% the year before.

When you dig into the causes, faulty algorithms are rarely the culprit. BCG found that around 70% of AI implementation failures trace back to people and processes, with only 10% attributable to the models themselves.

Informatica’s 2025 CDO Insights survey of 600 data leaders found that data quality, completeness, and readiness ranked as the single biggest obstacle to getting AI into production – cited by 43% of respondents, ahead of technical maturity, skills gaps, and regulatory constraints.

What makes operational data so hard to trust

Plant operations generate data differently from the rest of the enterprise. Firstly, there is the issue of sheer volume – thousands of sensors (sometimes millions) continuously measure temperature, pressure, flow, vibration, and hundreds of other process variables. Each one of those signals passes through historians, control systems, and increasingly into cloud platforms. At each handoff, quality can degrade in ways that nobody notices in real time.

Research published in the Journal of Big Data found that the most common forms of sensor data error in industrial environments are missing values, outliers, bias, and signal drift. These are not rare failure modes. Studies in complex process manufacturing environments have found that up to 40% of machine and sensor data requires significant cleaning before it can support analytics work.

In contrast to transactional data, time series has no built-in correction mechanism. When a sensor begins to drift or a telemetry stream starts dropping packets, the equipment keeps running, the data keeps flowing, and the error compounds quietly until it surfaces somewhere downstream – usually in a model that’s started behaving strangely, or a forecast that engineers have stopped trusting.

As Brett Roscoe, SVP and General Manager at Informatica put it: attaining the data readiness required for AI to deliver transformative impact remains the central unresolved challenge for most organisations.

Where accountability breaks down

The structural problem is that operational data sits at the intersection of teams that don’t naturally talk to each other. OT teams own the sensors and the historians. IT teams own the cloud infrastructure and the data lakes. Data scientists build and maintain the models. Each group is accountable for their own layer, but nobody is formally accountable for the integrity of data as it moves across all three.

Informatica’s 2025 survey found that 92% of data leaders are worried that AI pilots are advancing without the underlying data challenges being addressed first. That figure points to something systemic: it’s not that organisations are unaware of the problem, it’s that they haven’t structured ownership around it.

The shift to cloud-based OT data lakes has added further complexity. Data that looks clean in local historian systems frequently arrives in the cloud with timing misalignments, duplicate records, and inconsistencies introduced during migration or replication.

What the companies getting it right actually do

Organisations that are getting significant financial returns from AI are much likelier to have redesigned their data workflows end-to-end before choosing their modelling approach. That sequencing matters. It means companies must start treating data infrastructure as the precondition for AI investment, not something to be sorted out after a pilot has already been launched.

Devon Energy offers a concrete example of what this looks like at scale. As one of the largest independent oil and gas producers in the US, they’ve built an AI operating model that runs across daily workflows (drilling, production, and management) with historian data feeding directly into cloud platforms for real-time optimization. At that level of operational dependency, manual data quality checks are not a viable option. The pipeline from sensor to model has to be continuously validated, and problems have to be caught before they reach the systems making live decisions.

Getting to that standard requires three things that most industrial organizations haven’t yet formalized. Data quality needs an owner – someone accountable for it in the same way that site managers are accountable for uptime and safety. It needs to be measured, with data quality KPIs defined at both site and enterprise level. And it needs to be validated continuously across the full chain, not checked once at project launch and assumed to hold.

The gap is widening

Industrial AI is past the experimental phase. The data shows that companies who have successfully scaled AI have achieved 1.5 times higher revenue growth, 1.6 times greater shareholder returns, and 1.4 times higher returns on invested capital over the past three years compared to those still struggling to get there. That gap isn’t driven by model sophistication. It’s driven by operational readiness. And a big part of that is whether the data feeding those models can actually be trusted.

Continuous, automated validation of time-series data is achievable now across millions of signals. The manual bottleneck that made rigorous data quality impractical at scale is no longer a constraint. What’s still missing in most organizations is the decision to treat data trust as a core operational discipline – not a data science problem, not a one-time migration task, but something that gets measured and managed every day, like any other variable that determines whether a plant runs well.

The AI investments have been made. The question is what’s standing between those investments and the returns they were supposed to deliver. In most industrial settings, the answer is still the data.

Jane Arnold is Chief Operating Officer at Aperio, a data quality platform for industrial operations. She has over 30 years of experience in manufacturing, process control, and digital transformation, and works with leading energy, chemicals, and industrial companies to improve the reliability of operational data for analytics and AI.

Sources

• Boston Consulting Group (BCG), Where’s the Value in AI?, October 2024
• RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, August 2024
• S&P Global Market Intelligence, Enterprise AI Survey, 2025
• Informatica, CDO Insights 2025: Racing Ahead on GenAI and Data Investments, January 2025
• McKinsey & Company, State of AI, 2025
• Teh et al., Sensor data quality: a systematic review, Journal of Big Data, 2020
• World Economic Forum, Unlocking Value from Industrial Data, 2023