Sponsored by
Industrial AI Doesn’t Fail at the Model.
It Fails on the Plant Floor.
In Today’s Issue:
🏭 Why industrial AI fails on the plant floor, not in the model
🧠 What “context” really means once mistakes get dangerous
🤖 Copilots vs. closed-loop agents, and where humans stay in control
🛠️ The Celanese test: when a deployment is more than a demo
📉 The 2028 prediction Geir will stand behind
✨ And more intelligence from the industry…
Dear Readers,
Most of us buy things made in factories every day: phones, clothes, the coffee maker, and so forth, but without ever setting foot on a plant floor. What you’d find there isn’t just machines: it’s people who can tell you, in minutes, what’s actually possible, what it costs, and how to keep your order moving without blowing up someone else’s. That judgment, seeing and reading the floor in real time, managing trade-offs no spreadsheet fully captures, is precisely the thing AI is now being asked to absorb. This interview is a sober take from someone who believes the upside is real, but who’s just as clear-eyed about how much of that human judgment still has to stay in the loop.
Every few weeks the headlines promise that AI is about to run the world’s factories, refineries, and power grids. Then you walk onto an actual plant floor and the story goes quiet: impressive pilots, multiplying dashboards, and almost none of it surviving to the night shift. This week, we talked to Geir Engdahl, co-founder and CTO, AI of Cognite, the Oslo-based industrial software company behind Cognite Data Fusion and the Cognite Atlas AI agent workspace. He is refreshingly blunt about why so much industrial AI stalls between a slick demo and daily use, what context really means once a wrong answer becomes a safety incident, and exactly where autonomous agents belong on the plant floor, and where they do not. It is one of the sharpest, least hype-driven conversations on industrial AI I’ve had all year.
Geir Engdahl, co-founder and CTO, AI of Cognite. (Cognite)
All the best,
Kim Isenberg
In Conversation: Geir Engdahl, Co-Founder and CTO, AI at Cognite
Superintelligence: Everyone says industrial AI is huge — but where do projects actually fail in practice? If foundation models are improving so fast, why are manufacturing and energy deployments still so hard to scale? What are the real bottlenecks: data quality, integration, latency, trust, organizational incentives, or something else?
Geir: Industrial AI doesn’t fail because AI lacks capability. It fails when powerful models aren’t connected to the realities of the plant floor.
There’s still a persistent myth that once you have the data and a strong model, scale is inevitable. The graveyard of industrial AI is full of impressive pilots that never made it into daily operations.
The real bottlenecks aren’t theoretical. They’re brutally practical and they’re not a secret. Vendors know about them. They just don’t fix them, because fixing them is hard to package in a sales cycle:
Data that exists but can’t be trusted in real time
Systems that were never meant to talk to each other
Models that can’t run reliably in production environments
And a trust gap because if operators can’t understand or rely on the output, they simply ignore it
Foundation models are very capable given the right context, but they don’t fix broken pipelines or disconnected workflows. You can have the best model in the world, and it still won’t matter if it’s not embedded where decisions actually get made.
Most teams don’t fail at building something that works, they fail at making it impossible to ignore. That gap between “it works” and “it’s used” isn’t a detail. It’s huge. Literally 100x harder to do production at scale than PoC.
Lindy remembers what you forgot to write down.
Lindy sorts your inbox, drafts replies in your voice, and preps you before every meeting over iMessage. 400,000+ professionals use it. Try it free.
Superintelligence: What does “context” really mean in industrial AI — beyond the buzzword? You argue that scaling AI in heavy industry requires more than just bigger models. Concretely, what kinds of context does an Industrial Knowledge Graph add that a powerful time-series model like NV-Tesseract would otherwise miss?
Geir: “Context” is where most industrial AI conversations quietly fall apart. Everyone says they have it. Almost no one actually does. And here’s the test: if your AI team can’t tell you what a specific asset failure costs the business in dollars and cascading downtime, you don’t have context. You have a dashboard with a machine learning sticker on it.
A time-series model can scream “anomaly” all day long but it has zero idea what it’s looking at. It doesn’t know if it’s flagging a nuisance fluctuation or the early warning sign of a multimillion-dollar failure. It doesn’t know what’s connected, what’s critical, or what’s about to cascade.
It’s not intelligence, it’s pattern-matching dressed up as insight.
What an industrial knowledge graph does is inject reality back into the system:
What the asset actually is and how much it matters
How it connects to everything upstream and downstream
What’s broken before, how it failed, and what it cost
What “normal” really looks like in this plant, not in theory, but in practice
Without that, models don’t understand the system - they hallucinate relevance. With it, they start to think more like operators: weighing trade-offs, understanding consequences, and acting in context.
That’s the difference people underestimate. This isn’t about making models slightly smarter. It’s about dragging them out of abstraction and forcing them to operate in the real world. Because in industry, a model that doesn’t understand context isn’t just useless, it’s dangerous. It will confidently optimize for the wrong thing and the consequences won’t show up in a benchmark. They’ll show up on the plant floor.
Superintelligence: Are we moving from copilots to closed-loop industrial agents — and where do you draw the line? Cognite Atlas AI is positioned as a workspace for industrial agents. In your view, which decisions should agents be allowed to recommend, which should they automate, and where must humans remain firmly in control in factories, plants, and energy systems?
Geir: Yes, we’re moving from copilots to agents but there is a lot of confusion in the market about where that line is drawn. Some vendors are selling autonomy roadmaps they know their customers aren’t ready for, and that the technology can’t yet safely deliver. There needs to be more transparency there because the industry won’t self-correct if everyone keeps nodding along. The risks are too great in an industrial environment for there to be any ambiguity about what can and can’t be handled via autonomous agents.
Platforms like Cognite Atlas AI are a big step forward because agents finally have access to real operational context not just dashboards and isolated data. That makes them useful, however, it’s not trustworthy enough to hand over control. And in industrial environments, getting that line wrong isn’t a UX issue, it’s a safety incident, a shutdown, or worse. Therefore, the progression isn’t smooth or universal. It’s gated.
Here’s how it actually plays out:
Recommend aggressively: Agents should be everywhere when it comes to analysis: flagging root causes, proposing process changes, prioritizing maintenance. This is where they’re already outperforming humans in speed and pattern detection.
Automate surgically: Let agents take on repeatable, low-risk actions such as creating work orders, triggering standard workflows but only in tightly governed, fully auditable lanes. No black boxes and no silent failures.
Control reluctantly: Closed-loop automation should be the exception, not the goal. It only belongs in tightly bounded, well-understood scenarios where the downside is contained and the guardrails are explicit.
If the decision involves safety, ambiguity, or meaningful trade-offs, a human owns it.
What’s dangerous right now is the push to anthropomorphize agents: to treat them like operators instead of what they are: systems that can accelerate decisions, not own them.
The companies getting this right aren’t chasing autonomy. They’re building accountability: faster decisions, tighter feedback loops, and humans placed exactly where they need to be. They’re also the companies who pushed back on their vendors and said: show us where this actually works, not where it might work in your roadmap.
Because in industry, “closed-loop” isn’t a milestone. It’s a liability if you get there too early.
Superintelligence: What is the strongest evidence that this is more than a demo? For a customer like Celanese, what specific KPI moved — throughput, downtime, energy efficiency, waste, forecasting accuracy, or operator workload? And what would you say to skeptics who think industrial AI announcements still too often stop at pilot stage?
Geir: The strongest evidence is when AI changes how a plant is actually run, not just how it’s analyzed.
In the Celanese deployment, this isn’t theoretical:
Downtime drops because issues are caught earlier, before they escalate
Throughput improves because processes are tuned continuously, not reactively
Operator workload goes down because insight shows up in the flow of work, not buried in another tool
But candidly, those KPIs aren’t even the most important signal. The real proof is that the system doesn’t get turned off. It runs continuously, on live data, across shifts. It survives handoffs. It’s monitored, governed, and most importantly it’s used. That’s where most AI efforts quietly die: somewhere between the pilot and the night shift.
A demo is easy. A model that works once is easy. Even a pilot that shows lift is easy. What’s hard is building something that people rely on when things go wrong at 2 a.m.
To the skeptics: stop being polite about it. The reason AI announcements stop at pilot stage is that vendors and customers are both incentivized to let them. Vendors get case studies. Executives get board slides. And nobody has to be accountable for the production failure that comes eighteen months later. The whole ecosystem rewards the demo and punishes the deployment. That won’t change until customers start demanding production SLAs instead of pilot metrics—and vendors start putting real money on the line when systems fail.
Superintelligence: If industrial AI really works, what changes first: labor, planning, or the economics of operations? Over the next three to five years, where do you expect the biggest shift: predictive maintenance, process optimization, autonomous operations, workforce productivity, or entirely new ways of planning industrial systems? And which of those changes do you think the market is still underestimating?
Geir: When industrial AI actually works, the first thing that changes isn’t headcount, it’s how much wasted motion you can no longer hide. Workforce productivity moves first, but not in the way people expect. This isn’t about replacing operators, it’s about exposing how much of their time is burned hunting for data, reconciling conflicting signals, and second-guessing incomplete information. When that friction disappears, decision velocity spikes and so do expectations. Suddenly, “we didn’t have the data” stops being a valid excuse.
The next shift is process optimization but not as a one-off initiative. It becomes continuous, relentless, and a little uncomfortable. Instead of reacting to failures, you’re constantly tuning the system and margins stop being defended episodically and start being engineered every day.
Autonomy will come, but slower than the headlines suggest and in far tighter lanes than vendors admit.
What the market is massively underestimating is the compounding effect of small gains. A 1–2% improvement sounds trivial until it’s happening every day, across every asset, without backsliding and that’s when it stops being optimization and starts being a structural advantage.
And here’s the prediction I’ll stand behind: the industrial companies that don’t have AI-driven process optimization embedded in operations by 2028 won’t be acquired. They’ll be shut down. Their cost structure will simply be uncompetitive and not by a little, but by a margin that no amount of operational excellence or workforce restructuring can close. The window to act is shorter than anyone in a comfortable market position wants to admit.
Your creative brief is due Friday. Viktor wrote it Tuesday.
Tell him the campaign. Viktor pulls last quarter's performance from Meta and TikTok, scrapes competitor ads, drafts the brief, posts it for review. You edit, he ships the creative requests to your designer. Inside Slack.
The thread running through Geir’s answers
Industrial AI rarely fails because the model is too weak. It fails because a capable model is never wired into the messy reality of the plant: the data nobody fully trusts, the systems that were never built to talk to each other, and the operators who quietly switch it off the moment it stops being reliable.
The hard part is the last mile, not the model. Getting from “it works” to “it’s used” is, in Geir’s words, “Literally 100x harder to do production at scale than PoC.”
Context beats raw horsepower. A knowledge graph tells a model what an asset is, how it connects, and what a failure actually costs, so it stops being “pattern-matching dressed up as insight.”
Agents recommend, humans decide. Closed-loop autonomy stays the rare, tightly gated exception, because getting the line wrong is “a safety incident, a shutdown, or worse.”
The Optimistic Read
Industrial AI is past the question of whether the models are good enough. The real frontier now is operational discipline: getting trusted data, genuine context, and clear human accountability into the same place where decisions actually get made. Geir’s sharpest line is also his most demanding: by 2028, he argues, industrial companies without AI-driven process optimization embedded in operations “won’t be acquired. They’ll be shut down.” You don’t have to accept that exact timeline to take the point. The distance between a pilot that demos well and a system that survives the 2 a.m. night shift is where the next decade of industrial competitiveness gets decided.
About Our Guest
Co-Founder and Chief Technology Officer, AI at Cognite
Geir Engdahl is the co-founder and Chief Technology Officer, AI of Cognite, the industrial software company he helped start in 2016 in Oslo, Norway, with John Markus Lervik and backing from the Aker group. He leads Cognite’s AI technology strategy and helped build Cognite Data Fusion, the platform that turns scattered, untrusted industrial data into something models and engineers can actually use.
Before Cognite, he founded Snapsale, an AI-driven classifieds startup, and ran it as CEO and CTO until Schibsted acquired it in 2017. Earlier he was a senior software engineer at Google, building AI systems for ad targeting and conversion optimization. He holds a master’s degree in computational science from the University of Oslo.
In May 2025 he stepped into the CTO, AI role to, in Cognite’s words, “return to his disruptive founder roots” and push the company’s agentic AI strategy across any cloud and any large language model. That blend of frontier research and plant-floor pragmatism is exactly what makes this conversation worth your Sunday.
Sources:
🔗 Cognite Atlas AI: customer momentum and major release
🔗 Cognite and NVIDIA operationalize NV-Tesseract with Aker BP and Celanese
🔗 Cognite names co-founder Geir Engdahl CTO, AI