European startups building agentic AI — systems that do more than chat, such as searching databases and automating workflows — have attracted €7.9bn this year, already beating the €7bn raised in 2024, according to Sifted data. The money is going into areas like legaltech (Legora), healthtech (Tandem Health), and fintech (Cleo), but the hard part for enterprises is not raising capital; it's making the agents work reliably inside messy real-world operations.
A big hurdle is turning unstructured data from Slack, email, and documents into a structured 'perception layer' that agents can reason with, says Emma Burrows of Rezonant. Her firm builds context graphs that map information into business categories like sales or finance, so agents know where to look. Lindsay Keim of N8n warns that waiting to clean up entire data stacks first is a trap; 'data cleaning purgatory' stalls many projects. Instead, platforms like N8n let users pull data out of legacy systems, run it through AI, and move it into newer systems without a full overhaul. For web-sourced data, Tavily's Rotem Weiss says an 'abstraction layer' is needed to turn the unpredictable internet into consistent context.
Governance is just as important. Burrows notes that configuring permissions for each agent workflow is the 'most difficult and important part' of what Rezonant does. Agents should learn routing patterns from human feedback over time, but high-stakes decisions still need a person. Weiss adds that agents retrieving web information should send only the minimum necessary context from company data, and strict auditability matters because a wrong fact can escalate if it's written straight into Salesforce. Keim sums it up: AI provides speed, scale, and data synthesis, but humans must supply judgement, context, and accountability.
Success is also being measured differently. Keim says the fastest returns come from targeting repetitive tasks rather than trying to automate everything at once — 'we rarely see a customer go from zero to fully agentic in one move.' The long-term value lies not just in doing things faster, but in agents learning patterns and gradually taking on more work while keeping human oversight in place.