An upstream oil company in Asia built 140 AI applications across 11 offshore assets. Every application was critical to operations. None of them scaled beyond the site where it was built. Jason Schern, Field CTO at Cognite, identifies fragmented data architecture as the root cause blocking industrial AI from delivering measurable business value.
Why Does Industrial AI Fail to Scale
Industrial AI fails to scale when companies build applications that work at one site but cannot be replicated at others. According to Jason Schern, Field CTO at Cognite, a large upstream oil company in Asia built exactly 140 such applications across 11 offshore assets, and zero had been deployed to a second site.
During a recent audit, a large upstream oil company in Asia discovered something shocking: they had developed 140 applications and agents across 11 offshore assets. Every single one was critical to operations; they couldn’t turn off a single application without disrupting the business.
However, here’s the kicker: zero of those 140 solutions had been deployed beyond the original asset for which they were built.
“Value scaling is key,” Jason explains. “If I build an agent for one site and I can’t lift and shift it to the next site, things that provide incremental value in one place never become measurable value to the business.”
Why Do Industrial AI Projects Stay in Bubbles?
Industrial AI projects stay in isolated bubbles because they are built on top of fragmented, context-poor data foundations. The problem is not technical capability. As Jason Schern of Cognite puts it: “If you want to be good at AI, you’ve got to be great at data.” Companies that achieve measurable results are those that can scale and rapidly validate value across multiple assets.
The problem isn’t technical capability—it’s data architecture. Companies are building AI solutions on top of fragmented, context-poor data foundations that make scaling impossible.
Jason puts it bluntly: “If you want to be good at AI, you’ve got to be great at data.”
The difference between companies that achieve measurable results and those stuck in limbo: the ability to scale and rapidly validate value across multiple assets. “You don’t want some magical use case that drives outsized advantage in one place, in a bubble,” Jason says. “You want things you can scale rapidly across them all.”
Why Does Industrial Data Lack Context?
Industrial time-series data contains almost no built-in context. Unlike language, which carries meaning through grammar and structure, raw streams of sensor readings carry no information about equipment relationships, work orders, or plant schematics. That missing context is what prevents AI models from reasoning accurately across industrial assets.
Large language models work because context is built into the content: grammar, word choice, paragraph structure. But with industrial data it’s a different story. “If I’ve got streams and streams of time series data, how much context is there in that? Almost none,” Jason notes.
A midstream operator in the U.S. proved this point by achieving a 15% reduction in energy consumption, not through better sensors, but by combining operational data with financial forecasting, energy cost predictions, and equipment relationships. They developed virtual flow meters that eliminated laboratory testing by using analytics from pressure, temperature, and flow rates combined with contextual business data.
“You need to know: Was there a work order? When was it executed? Where is this equipment in the plant schematic? What’s upstream and downstream?” Jason explains. “All of those things provide context that allows you to reason.”
What Cultural Barriers Block Industrial AI Scale?
Two outdated beliefs block industrial AI from scaling: that industrial data belongs on-site and should never leave, and that copying data is inherently bad. Both beliefs originated when storage costs were prohibitive. Those conditions do not apply. Jason Schern of Cognite notes that storage now costs pennies per gigabyte, and the real cost in an AI environment is compute, not storage.
Two outdated paradigms are blocking progress:
- “Industrial data belongs on-site and should stay on-site.”
- “Copying data is bad.”
Both stem from an era when storage costs were prohibitive. “Storage costs are pennies on the gigabyte now,” Jason says. “What kills you is compute; when you try to find relevant data and connect it.”
The solution? Copy the data. Optimize it for different use cases. “There’s almost zero cost in making that copy, and in the copying, you’re optimizing things in a way that dramatically lowers compute costs, which in the age of AI is everything.”
How Did Aker BP Cut RCA Time from Weeks to Hours?
Aker BP used AI augmentation to compress one aspect of root cause analysis from six weeks to less than six hours. Root cause analysis is the process of identifying the underlying cause of an operational incident. Because traditional RCA takes months, most companies can only investigate a fraction of incidents that warrant review. The Aker BP result means the same team can now work through a backlog of analyses that previously went unresolved.
Aker BP’s work with root cause analysis (RCA) illustrates AI’s real potential. Traditionally, RCA activities take months, meaning companies can only investigate a fraction of incidents that warrant analysis.
Using AI augmentation, one aspect of the RCA process was compressed from six weeks to less than six hours.
“The company has not been able to do all the RCAs they should be doing,” Jason points out. “With the same people, I can now attack that backlog of RCAs that I would never get to and get the business benefit of those resolutions.”
Is It Too Early or Too Late to Act on Industrial AI?
It is both too early and too late at the same time, according to Jason Schern of Cognite. Industrial AI is evolving fast enough that organizations feel behind, while the technology is still maturing enough that it feels unready. Schern’s position is that this ambiguity should not slow action, because value captured now has an outsized impact compared to equivalent value captured a year later.
Jason’s advice for organizations feeling paralyzed by AI’s rapid evolution: “You’re going to feel like you’re too early and too late at the same time. And it’s true. You are both.”
Things are evolving so quickly that it feels like catch-up mode (too late), but also like the technology isn’t quite ready (too early).
“Don’t let that ambiguity slow you down,” Jason warns. “Value now is much more important than potential value in the future. The value you capture at this moment has an outsized impact versus equivalent value achieved a year later.”
The companies winning at industrial AI aren’t finding magical use cases. They’re building data architectures that allow small incremental improvements to scale across all assets rapidly.
As Jason puts it: “If you don’t have a plan to handle the data properly, that whole data ops conversation fueling AI is the most relevant conversation to value scaling.”
The question isn’t whether AI will transform industrial operations. It’s whether your data architecture will let you scale that transformation beyond the first site.
Sponsored by Cognite
About the author
This article was written by Greg Orloff, Industry Executive, IIoT World. Greg previously served as the CEO of Tangent Company, inventor of the Watercycle™, the only commercial residential direct potable reuse system in the country.
FAQ
1. Why do industrial AI projects fail to scale across sites?
A large upstream oil company in Asia built 140 AI applications across 11 offshore assets, but zero were deployed beyond the original site where each was built. According to Jason Schern, Field CTO at Cognite, the cause was fragmented data architecture. When AI solutions are built on context-poor data foundations, they cannot be transferred from one asset to another without losing accuracy and relevance.
2. How does data architecture affect industrial AI ROI?
Data architecture determines whether an AI solution built for one industrial site can scale to others. Jason Schern of Cognite explains that industrial time-series data carries almost no built-in context: no equipment relationships, no work order history, no plant schematic position. Without that context layered in, AI models cannot reason across assets, and incremental gains at one site never become measurable business value across the operation.
3. What is the difference between AI value at one site vs. scaled AI value?
An AI solution that delivers value at one site but cannot be replicated is what Jason Schern of Cognite calls a “bubble.” Bubble value is incremental and unmeasurable at the business level. Scaled value, by contrast, occurs when the same solution is deployed rapidly across multiple assets. Schern’s position is that small improvements scaled across all assets outperform large, isolated wins at a single location.
4. How did a U.S. midstream operator achieve a 15% reduction in energy consumption using AI?
A midstream operator in the United States achieved a 15% reduction in energy consumption by combining operational data with financial forecasting, energy cost predictions, and equipment relationship data. The operator also developed virtual flow meters that eliminated laboratory testing by using analytics from pressure, temperature, and flow rates combined with contextual business data. The gain came from adding context to the data, not from upgrading sensors.
Related Reading
Scaling AI at the Energy Edge: Why Pilots Succeed and Deployments Stall
