What Didn't Pencil Yesterday Pencils Today

Scene one, before AI. An industrial company has a genuinely good idea: a learning platform for complex technical knowledge. Skilled workers are scarce everywhere, demand is validated, the first customers are paying. And still the venture is wound down after eighteen months. Not because the market said no. Because high-quality technical content was so expensive to produce that the unit economics never worked. Every new module cost more than it could ever earn.

Scene two, today. A manufacturer with millions of installed devices turns its after-sales function into a business. Service used to be a cost center: every diagnosis, every piece of advice, every repair referral needed a human. Now AI handles diagnosis and self-help, agents run the operations, and the cost center becomes revenue.

The idea barely changed between the two scenes. The economics did.

The misunderstanding: AI as an efficiency programme

When people talk about AI in companies today, it's almost always about the same thing: faster processes, less manual work, better texts, quicker analyses. All true. And long since standard.

Anyone in 2026 who advertises that they use AI in their own operations is advertising running water. Every consultancy, every agency, every service provider has that slide. Efficiency is table stakes, not an edge. It makes what exists cheaper. It invents nothing new.

The real shift lies elsewhere. AI doesn't only change how fast a company works. It changes which business models are worth building at all.

The shift: marginal costs that collapse

Many business models never failed on the idea, but on a single number: the cost per customer interaction. Every diagnosis, every individual consultation, every set of instructions, every case in the long tail needed a human. Beyond a certain point it simply didn't pay. So it stayed a cost center — or never got built.

Those are exactly the marginal costs AI is breaking open. Diagnosis, advice, service, guidance, documentation: what yesterday needed an expert per case scales today. Business models that were economic nonsense five years ago are suddenly profitable.

Research on failed ventures shows how big that lever is. In CB Insights' well-known analysis of more than 400 failed VC-backed companies, 70 percent ultimately ran out of capital. Underneath lie the real causes — and one of them is unsustainable unit economics, at 19 percent. That is precisely where AI bites. Not on the idea. On the math.

Who benefits — in theory

Now it gets interesting. Because the new economics are available to everyone. Every startup can use the same models, build the same agents, buy the same tools.

What isn't available to everyone is the asset you have to apply the new economics to: the installed base. Customer access. Decades of domain knowledge. Service data. Trust within an industry.

All of that sits with established companies. With the manufacturer that has millions of devices in the field. With the Mittelstand company that has known its niche for three generations. With the group whose service organisation knows every customer by name.

In theory, that's the best starting position in decades. The asset is there, the new economics are there. You just have to bring them together.

Why it doesn't happen in practice

And that's exactly where it catches. Inside the existing structure, asset and new economics rarely meet.

IT governance demands months of approvals for every new tool. Legacy systems can't be connected to modern agents. The service organisation sees the new model as a threat to its own existence — and is right. Budget sits in business units whose targets hang on the existing business. And every brand risk is hedged until nothing of the new thing is left.

Nobody meant any harm. The structure is built for exactly this: protecting what exists. For the core, that's right. For an AI business model, it's fatal.

Practitioners estimate that 80 to 90 percent of corporate innovation labs fail. AI doesn't improve that rate if the structure stays the same. You just get innovation theater with better tools.

The insight: the asset is inside. The freedom is outside.

The equation is simple. The established company's asset. Times the new economics of AI. Times the freedom to bring the two together without the old structure.

If any of the three is missing, the result is zero. A startup has freedom and AI, but no asset. The established company has the asset and access to AI, but no freedom. Only when all three come together does the lever appear.

Which means: the AI business model is built outside. In its own company. With its own tech stack, its own data architecture, its own processes. With agents in operations from day one, instead of retrofitting them years later into an organisation that was never built for them. And with access to the parent's asset — contractually agreed, not begged for.

A guardrail, before anyone takes the wrong turn

Freedom here doesn't mean around the rules. The AI Act applies. Data protection applies. Product liability applies. Whoever builds outside to dodge compliance is building on sand.

The opposite is true. On a greenfield site, an AI business model can be built cleanly from the start — compliance by design instead of retrofitting into legacy systems nobody fully understands anymore. Building outside isn't the way around the rules. It's the easier way to comply with them from day one.

What it changes in the venture

There's a second effect almost nobody talks about. A venture that runs lean on AI agents from the start needs fewer people per euro of revenue. Its break-even moves forward. It needs less capital to reach profitability.

That changes the risk for everyone involved. Less money on the line, faster clarity on whether the business carries. For an established company that wants to build its next business, that's not a technical detail. It's the answer to the question behind every investment decision: how much do I have to risk before I know whether it works?

The zoom-out: Europe's second chance

Europe largely missed the first wave of the platform economy. The big consumer platforms were built elsewhere. This wave is different. In AI for industry, service and B2B, the winner isn't automatically whoever has the fastest models. It's whoever has the asset they're applied to.

And a substantial part of that asset sits in European companies. In installed bases, in domain knowledge, in the customer relationships of industry and the Mittelstand. The market is growing fast: autonomous AI is expected to rise from around $8 billion in 2025 to $250–264 billion by the mid-2030s.

The question isn't whether this value gets created. It's whether the companies that own the asset capture it themselves — or watch others attack their asset from outside.

The principle: apply the lever where the core has run out of levers

What would a system look like that actually uses this window? It would have four properties.

First: AI in-house is mandatory, not strategy. Efficiency in operations is the entry ticket. Turning it into a differentiator confuses the tool with the business.

Second: look for the asset, not the technology. The right question isn't "where can we use AI?" but: "which of our assets suddenly becomes a business under the new economics?"

Third: the AI business model is built outside. Its own company, its own stack, its own processes, agents from day one — with regulated access to the asset and governance that protects the parent without suffocating the new.

Fourth: built cleanly, from the start. Compliance by design isn't a brake; it's a competitive advantage over everyone who will have to retrofit later.

The circle closes

Europe too rarely turns technology into companies. AI sharpens that question — and opens a window at the same time. Ideas that failed on the math yesterday pencil today. The learning platform from the first scene would be a different business now. The service from the second scene was unthinkable yesterday.

The asset sits with established companies. The economics sit with everyone. The only question is who brings them together — and where.

AI isn't a label. It's the lever that starts exactly where the core has run out of levers.

Sources

CB Insights, "Why Startups Fail" (more than 400 failed VC-backed companies; 70% ran out of capital, 19% unsustainable unit economics) · Capgemini / CB Insights, corporate innovation lab failure rate (80–90%, practitioner estimate) · Fortune Business Insights / Research Nester / GM Insights, autonomous AI market (~$8B 2025 → $250–264B 2034–35)