You can't automate a factory with a guesser
Why the scarcest thing in AI isn't intelligence
The most valuable European AI companies of the next decade will be built on top of industrial infrastructure like factories, mines, energy grids or laboratories. By founders who have spent their careers inside these industries.
Here is why.
The market shift
There’s a version of the AI future that goes like this: foundation models keep getting better, Big Tech keeps telling everyone they are the ones owning pretty much all of this frontier – and eventually they actually do. In this future, the only winning move is (or would have been) building or investing at the model layer.
We believe this version is wrong. Not because the models aren’t impressive (they very much are). But because this future confuses intelligence with real-world competence.
The SaaS playbook used to be straightforward: find an inefficient workflow, build a tool for it, charge per seat per month. The tool was the product. And at least part of the moat was the engineering effort required to build it. That moat is gone. Just last week, Claude (my colleague, not the model) and I had a 90-minute deep dive with one of our portfolio CTOs, and the reality is: what used to take engineering teams months is now prototyped over a weekend and built in a week. Code itself has become a commodity.
But not just the building has changed, the “using” has changed too. The interfaces the software industry has spent decades designing (dashboards, forms, buttons, dropdowns) are about to become irrelevant when the next generation of users are agents executing API calls. When your user is no longer human, and your product is no longer hard to build, what exactly is your moat?
For a lot of rather generic SaaS companies, the honest answer probably is: not much. CRMs (if not evolving) are turned into passive databases while the actual sales execution moves into the AI layer. Support tickets get resolved end-to-end without a human in the loop. And “ask the data” commonly replaces dashboards. We’re not saying the incumbents are disappearing overnight. But you can watch their moats eroding and $300 billion in SaaS market cap vaporizing in real time.
What replaces these companies isn’t just more software, but a different kind of company entirely. One that doesn’t merely sell access to a tool but delivers a measurable work outcome. And the addressable market for “solved problems” is categorically larger than the addressable market for software tools. The global labor market (the thing AI-native companies now directly compete with) is $50 trillion. That’s a fundamentally different playing field than IT budgets.
A lot of the current discussions in our industry converge on this basic insight: value is moving from the tool layer to the intelligence layer, and the TAM is expanding from IT budgets into labor budgets.
But here is where we believe the conversation goes wrong.
Big Tech agrees – and that’s the problem
OpenAI explicitly validated this thesis a few weeks ago when it launched Frontier, an enterprise platform it describes as “an intelligence layer that stitches together disparate systems and data within an organization,” claiming: if you connect all your data (your CRMs, data warehouses, ticketing tools, etc.) OpenAI’s agents will operate with “shared business context.”
And as if to put more emphasis behind this, the company just announced “Frontier Alliances” with McKinsey, BCG, Accenture, and Capgemini, because (in their own words) “the limiting factor for seeing value from AI in enterprises isn’t model intelligence – it’s how agents are built and run in their organizations.” OpenAI confirms they need consulting firms with “deep industry, functional, and domain expertise” to bridge the gap between what their models can do and what enterprises can actually deploy.
The most valuable AI company in the world is telling you that intelligence isn’t the bottleneck, domain expertise is. You could read this charitably – after all, they are not the first enterprise platform to build a consulting ecosystem (just think about Salesforce or SAP). But the fact that OpenAI chose this as a major strategic move, rather than just a channel play, tells you something about where they see the gap.
The catch is: this whole development is primarily focused on organizational context that lives inside a company’s digital tools (think Slack threads, wiki pages, onboarding docs etc.). This is obviously a massive market. And seemingly all foundational model companies, as well as the ServiceNows, Notions, and Gleans of the world are racing to own this layer.
But this is also the easier version of the context problem.
The context that can’t be copied
The harder version of the context problem (and the best-performing companies in our portfolio share this pattern) is industry context. The proprietary decision traces that live in the physical and regulatory infrastructure of entire sectors. Factory PLCs. Mining operations. Laboratory notebooks. Construction site workflows. Energy grid dispatch. Logistics routing under real-world constraints.
None of this context lives on the internet. None of this lives in Notion. It barely lives in databases. Most of this lives in the heads and hands of people who’ve spent their careers operating in these environments, and in the physical systems they’ve built and maintained.
That distinction is everything, and it creates what we call the Trust Ceiling: a hard limit on how far generic AI can reach into the physical world without being grounded in verified, domain-specific context. This Trust Ceiling exists because of three properties that make industry context fundamentally different from organizational context.
Industry context isn’t digitized. While organizational context is messy but machine-readable (it already lives in digital tools), industry context lives in analog systems, physical processes, and human judgement that was never written down. The decision logic of a factory floor is encoded in PLC configurations, operator intuition, and maintenance schedules refined through decades of trial and error. You can’t scrape it. You can’t API-call it. You earn access through presence, expertise, and trust.
The cost of getting it wrong can be catastrophic. Hallucinate a sales summary, and someone has a bad meeting. Hallucinate a control parameter on a production line, and you destroy equipment – or worse. In regulated environments, wrong answers can mean sanctions, environmental damage or even loss of life. This is precisely why making AI reliable in physical environments requires deep domain-specific engineering: deterministic guardrails, verification layers, and safety architectures that only people who truly understand the operational context can design. Generic platforms can’t provide this.
You can’t automate a factory with a guesser.
And the data compounds differently too. Every deployment generates operational traces that feed back and make the next deployment smarter (think edge cases, failure modes, real-world performance data). None of this data you can purchase or synthesize. It’s proprietary, context-specific and it compounds over time. The longer you operate, the harder it is for anyone to catch up.
The moats that actually hold
For organizational context, the moat question can be tricky. As one commenter on Evan Armstrong’s “Context is King“ piece put it: “what stops companies from taking their context to the cheapest vendor?” For wiki and workflow context, that can be a real objection. Migration is painful but often possible. The cheapest vendor eventually might win.
For industry context, two moats in particular operate at a fundamentally different intensity.
Network effects from deployment data. Every customer in a specific vertical generates proprietary operational data that feeds back into the product. A manufacturing AI company serving 50 customers should have a categorically better product than one serving only 5, just from encountering and solving 10x more domain-specific edge cases. Factory 50 benefits from everything that went wrong at the first 49. And this is fundamentally different from organizational context. Anyone can create a Notion workspace and start generating workflow data tomorrow. Try this with a factory deployment. The acquisition cost for relevant data is orders of magnitude higher, which means the network effect gap is orders of magnitude harder to close.
Brand and trust. In industrial and regulated environments, a trusted brand is the moat. When a pharmaceutical company selects an AI system for lab automation, or an energy company chooses a platform for grid management, the decision goes through not just procurement, but legal, compliance, and sometimes even a regulator. Switching cost isn’t mainly bound to data migration, but to recertification, revalidation, and reestablishment of trust with every stakeholder. And this in an organization that is often categorically risk-averse. Once you’re in, you’re in for years.
These moats create a flywheel. Network effects increase the quality of the product, which builds brand and trust to win the next customer, which again generates better network effects. Harder to start. But also much harder to stop.
This is why we believe the most valuable AI companies of the next decade won’t look like many of today’s AI startups. They’ll be vertical champions with deep industry context, hard-earned trust in regulated environments, and compounding advantages that grow with every customer.
Who can build this
You can’t build above the Trust Ceiling from the outside in, because the networks providing access to the data and design partners needed more often than not are personal, not institutional.
The people who can build these companies are industry insiders who are also exceptional technologists. We call them Expert Founders, or Nerds and Pioneers.
They have spent years embedded in the industry they are now building for. They have lived the problem. They know the workflows, the internal politics, the procurement cycles, the unwritten rules. And they have the trust networks to get design partners and first customers in months, not years.
And they also bring genuine technical depth (the ability to build deterministic systems on top of that context that can operate above the Trust Ceiling). Not API wrappers. Not chatbots with industry-specific system prompts. But actual engineering that a compliance officer or plant manager can sign off on.
This is a rare combination, and that rarity is the point. There are plenty of domain experts who understand the problem space deeply but who lack the technical ingenuity to build a product. And there are plenty of strong engineers who can build impressive systems but have no access to (or no understanding of) the data, the design partners, or the trust networks needed to deploy in these environments. The intersection (founders who have both) is exceptionally small.
One of our founders visited more than 350 factories and spent nine years selling to factory managers before building shop floor automation. His company grew revenue 5x last year with customers like ABB, Viessmann, Bosch and Continental. Another spent a decade as a senior leader in the global mining industry, including at the WEF, before building an AI solution now deployed at ArcelorMittal and FLSmidth. A third has a PhD from Oxford, did research at NASA JPL, and led ML teams at three companies before tackling confidential computing.
Every one of them had a career in their respective industries before they started their own company.
Why this points to Europe
Europe’s strongest card in the AI race isn’t its model layer or its venture ecosystem. It’s the proprietary industrial context that no other region can replicate at the same density.
This proprietary industrial data isn’t lying around waiting to be ingested by an API. It’s trapped in analog systems and guarded by organizations who have cultivated a culture of deep-seated caution about IP for decades. And that is exactly why Big Tech cannot touch it: unlocking it requires deep, localized trust. The AI companies that will win here are the ones built by founders who already speak the language of these incumbents. Teams that leverage established trust to solve complex, existential bottlenecks from the inside out.
And Europe still has the world’s densest concentration of advanced manufacturing, chemicals, energy infrastructure and precision engineering. The German Mittelstand alone represents countless world-leading companies in niche industrial categories, sitting on top of decades of hyper-specific, proprietary, physical-world data. These are the source of the proprietary context that makes vertical AI possible in the first place, and defensible in the second.
This density is compounded by an accelerating macro tailwind: a tangible mandate for European industrial sovereignty. Driven by the urgent need for supply chain resilience, energy independence and aggressive climate mandates, European industry is modernizing at an unprecedented scale.
But these advantages do have an expiration date. If European industry doesn’t move to digitize and defend these analog moats now, someone else (like a US or Chinese platform) will find a way in. The window is open, but it will close, and Europe’s hidden champions are facing a critical turning point: grasp the opportunity or risk that the industrial knowledge they’ve spent decades building becomes someone else’s competitive advantage. The question isn’t if this data will be unlocked, but whether our incumbents will actually trust European startup founders in time for this to be a European success story.
What this means for us
We built NAP around the conviction that the most valuable European AI companies of the next decade will be built by industry insiders who use AI as a tool to solve problems they’ve spent their careers understanding.
And while the discourse in our industry seems headed in a similar direction (context matters, SaaS is being repriced, the TAM is expanding into labor budgets) it’s missing the crucial nuance that not all context is created equal. Building vertical AI companies in physical-world industries requires a different kind of founder, a different kind of company, and a different geographic center than the one Silicon Valley is optimizing for.
This is not a frictionless path, and building in these environments demands patience from founders and investors alike. But at NAP, we’re backing the Nerds who know the factory, the mine, the grid, and the lab from the inside. And the Pioneers who are building the intelligence layer on top of the industries that make our physical world work.
In Europe.