For the past few years of my VC career, AI in Europe has been something you hid in the back of the system: a recommendation engine in a telco stack, a fraud model in a bank, a proof-of-concept chatbot that nobody really owned. Enterprises bought a few point solutions, hired a small data team and ticked the “AI” box. It was treated as seasoning, not as the way people actually interacted with tools.

That is now changing in a very simple way: people would rather talk than click. In 2024, European AI startups raised over $13 billion, and roughly a fifth of all VC funding in the region went into AI. A lot of that money is not going into new models; it’s going into products that sit in front of existing systems and let you say, “show me this, fix that, do this next”, in your own language. At the same time, the EU AI Act has entered into force, which means that by 2026 every Member State has to run at least one AI regulatory sandbox for “high-risk” systems in areas like healthcare, employment or critical infrastructure. So we’re not just going to have agents; we’re going to have agents that live under supervision, in hospitals, schools, banks and public services.

When I talk to founders now, they are not pitching generic copilots. They are building “Her-style” agents that can see across tools, understand the local language and regulations, and are safe enough that a CIO, a regulator and a works council can all sleep at night. The interface to work becomes a conversation, not a menu of icons, and Europe’s mix of regulation, domain depth and industrial customers makes it a surprisingly good place to build that.

The last decade of marketplaces came in waves. First the B2C giants (Uber, Airbnb, etc.) then a rush of B2B marketplaces for everything from freight to industrial inputs. Most ran into the same brick wall: expensive sales on both sides, heavy manual onboarding and compliance, long cycles to reach liquidity, and unit economics that stopped making sense as soon as growth slowed. A lot of those models were “right idea, wrong friction cost”.

What changes with AI isn’t just better search; it’s the workflow. The very tasks that made B2B marketplace economics painful: qualification, onboarding, KYC/AML, fraud checks, support, reconciliations, are the ones that AI can now automate. And crucially, the cultural shift has started: large tech companies are already asking employees to use AI as part of their day-to-day. The question is how quickly traditional industries follow, because once that expectation takes hold, automated workflows move from innovation to the new standard.

That’s why I see “AI in Europe” less as a contest over foundation models and more as an opportunity to own the workflow layer. Europe’s real edge sits in enterprise complexity, from financial infrastructure to manufacturing to regulated operations. And the horizontal AI that can orchestrate these systems is still wide open: products that plug into CRMs, ERPs, core banking, payments and data warehouses. In that world, marketplace liquidity can be unlocked far faster and cheaper because AI strips out the friction, and suddenly it’s no longer a brute-force, headcount-driven exercise.

For a long time, “AI” in the market meant two extremes: glossy chatbots on one side and dense research models stuck in Jupyter notebooks on the other. The serious science lived inside pharma R&D, climate labs or physics departments and almost never made it into products that normal people or SMEs could actually touch. It was easy to think of “AI & ML” as consumer interfaces or enterprise copilots, with scientific computing sitting off to the side.

That boundary is already blurring. We now have AI systems supporting almost every stage of the research process, from hypothesis generation and literature triage to data analysis and experimental design, and they are being used by working scientists, not just AI labs. You see it in climate modelling, in protein design, in materials optimisation. Companies like Lila are raising hundreds of millions to build “AI science factories”: automated labs guided by domain-specific models and valued north of a billion dollars. The interesting part for me is what happens next: those capabilities don’t stay in the lab. They leak out as APIs and tools that sit under climate software, energy-optimisation products, even consumer apps that help you understand how efficient your home, your training plan or your physiology really are.

If you look at what we’ve called “AI agents” over the last two years, most of it has been UI. Clever wrappers sitting on top of someone else’s stack: they demo well, they live in the browser, and then they get parked in a side channel because nobody trusts them with real money or real risk. From a capital point of view, we’ve basically funded thousands of small experiments at the edge.

That window is closing. In 2024, AI startups pulled in roughly a fifth of all European VC funding, which tells you that investors increasingly see AI as a core infrastructure layer, not an add-on. Regulation is catching up fast: the EU AI Act is live, and by 2026 every Member State has to operate an AI sandbox for “high-risk” systems in areas like finance, employment and critical services. That forces agents out of shadow IT and into governed environments where CROs, CISOs and regulators have an opinion.

In parallel, financial institutions are quietly leaning in. Around three-quarters of UK financial firms already use AI somewhere in their stack and another ~10% plan to within three years. The interesting shift is not another chatbot; it’s agents that sit across trading, risk, procurement or finance systems, read the logs and documents, and start proposing or triggering actions. In security specifically, this looks like an autonomous SOC: agents continuously ingesting telemetry across endpoints, cloud, identity and apps, triaging alerts, auto-resolving low-risk incidents, and escalating only the real edge cases to humans, under hard guardrails. Most of the thin, single-skill agents launched in the last 24 months will either be acquired as features or disappear. The ones that matter will look more like infrastructure: deeply integrated, auditable, with proper guardrails and kill switches.