Fabian Westerheide is a founding partner of the AI-focused venture capital investor AI.FUND and has been privately investing in AI companies through Asgard Capital since 2014. Westerheide provides strategic advice to public and private institutions in the area of AI and invites you to the Rise of AI conference in Berlin every year.
Most AI startups don’t fail because their technology is bad. They fail because Microsoft integrates their business model.
That’s exactly what we’re experiencing right now. The real development is not that foundation models become a commodity. The bigger shift is happening one level above: LLMs are becoming the infrastructure layer of existing platforms. This means that the added value also moves upwards – in orchestration, integration, data access, governance and distribution.
The major platforms are now integrating AI agents and workflow automation directly into their core products. As soon as agents become part of the standard software, an old platform rule applies: the platform displaces the additions.
Many AI startups are currently raising capital for products that are already on the roadmap at Microsoft, Salesforce or SAP.
Where repression occurs
“LLM plus a nice interface” is no longer enough
Many AI products fail not because of poor technology, but because they do not have their own right to the customer. As soon as the major platforms integrate a solid 80 percent solution natively, the product quickly becomes a nice-to-have. And it is precisely these budgets that disappear first in difficult market phases.
Horizontal copilots become interchangeable
Standard use cases such as summaries, meeting recaps or writing emails have long since become part of the platforms. Anyone who just wants to be “a little better” will end up competing against the next product announcement.
The biggest danger for many AI startups is therefore not OpenAI, but the next Copilot rollout.
Companies no longer buy demos
The new standard is: agents, tools, workflows and control.
Today, enterprise customers no longer just buy answers, but operational execution – including logs, roles, policies, monitoring, traceability and liability. Anyone who cannot deliver this will be permanently stuck in the pilot project. And it is precisely this pilotitis that is causing the silent death of many B2B AI startups.
That’s why entire waves of startups are disappearing within a few years, similar to products that were replaced by native platform functions.
Where real opportunities arise
The next big gap is not in the next chatbot.
It arises where platforms do not have automatic access: in real processes, proprietary data and operational responsibility.
Regulated markets
Areas such as the public sector, finance, healthcare or critical infrastructure do not buy prompt magic, but rather operational capability. Data sovereignty, auditability, compliance by design and clear responsibilities are crucial. This is precisely where European providers can develop real advantages because trust, legal certainty and local infrastructure suddenly become strategically relevant.
Vertical workflows with proprietary data
Machine data, production reality, supply chains, field service or proprietary document stocks create real barriers to entry. The difference between a feature and a company today is no longer “we use AI” but “we own the process”.
Infrastructure instead of a killer app
N8n is a good example of this. Not one big AI application, but the workflow backbone between tools, systems and agents. Such building blocks win when AI becomes an operating system. Because every agent architecture ultimately needs interfaces, rights concepts, monitoring and workflow orchestration. Whoever delivers this infrastructure sits at a robust point in the value chain.
Europe underestimates itself
Europe has long had more than just model debates.
Companies like Mistral, Celonis, DeepL, Helsing or Synthesia show where AI actually creates value: in processes, industrial reality, safety-critical systems and real enterprise applications.
That’s where money is made in the long term, not in the next generic AI wrapper.
The three crucial questions for 2026
1. Will I become a button in Microsoft, SAP or Salesforce in twelve months?
If the answer is yes, it requires something that cannot be copied: process power, data access or distribution.
2. What is my unfair advantage?
Data, regulatory access, existing customer relationships, hardware or edge access – anything that can’t be replicated in a quarter.
3. Am I building a product or operating a system?
Agents are not a feature. Agents are operations. If you want to survive in this market, you have to master topics such as rights, monitoring, quality assurance, failover and cost control.
This is exactly where the new budgets are created.
Conclusion: AI becomes OS
The shakeout in the AI market will not happen because AI is getting smaller. It comes because AI is increasingly becoming part of everything.
And as soon as AI becomes an operating system, it is not the loudest wrappers that survive, but rather the companies that own processes, control distribution or provide the infrastructure without which no agent economy can function stably.
