Enterprise RPA meets LLMs: where European workflow automation is converging in 2026
The bot that breaks on a PDF layout change
The European mid-market spent most of the 2010s buying robotic process automation from American vendors. UiPath, founded in Bucharest in 2005, ended that decade as the largest of them and is still the largest by revenue. Automation Anywhere and Blue Prism filled out the top three. The promise was straightforward. A bot reads an invoice, types the values into the ERP, files the email under the right folder, and goes back to its queue. Procurement teams calculated the payback against a junior accountant's salary and the math worked.
The product reality was a different shape. A typical RPA bot at a German manufacturer in 2022 carried sixty to ninety rule branches that mapped specific PDF layouts to specific field positions. When SAP pushed a quarterly UI update, half of the bots had to be re-recorded. When a supplier changed its invoice template, the corresponding bot stopped working and silently filed nothing for a week. Forrester's 2023 survey of enterprise RPA users put the average time spent on bot maintenance at thirty percent of total automation effort, which collapses the original payback math.
What changed the conversation was the arrival of language models that can parse an unstructured document, extract the fields without rule branches, and route the output. Suddenly the bot does not need ninety rules for ninety layouts. It needs one prompt, a few-shot example set, and a fallback. The economics shifted.
What the European stack actually looks like now
The interesting part of the European market is who is positioned for that shift and who is not. UiPath has spent the last two years bolting AI agents onto its existing orchestration layer. The product makes sense but the legacy weight is real. Customers running the classic Studio bots cannot rip-and-replace overnight, and the licence model carries forward.
The companies built after the LLM moment look different. n8n, founded in Berlin in 2019 by Jan Oberhauser, is a self-hostable workflow engine designed around code nodes, AI nodes, and an explicit fair-code licence that lets customers run the platform on their own infrastructure. The €156M Series C in 2025 was not the size that mattered. The structural fact that mattered was n8n's growing presence inside companies that have a hard requirement to keep automation data on EU-controlled infrastructure. Make.com, the Czech-founded ex-Integromat now part of Celonis, sits in a similar architectural neighbourhood with a stronger SMB lean.
Camunda, the German workflow company that grew out of the Camunda BPM open-source project, took a different path. Its BPMN-modelled processes were already used inside European banks and insurers for compliance-critical workflows where every step had to be auditable. The 2024 release of Camunda 8.5 added native LLM task workers, which lets a process step call a model the way it would call a microservice. The reason this matters in Frankfurt and Munich rather than San Francisco is that a Camunda process for a SEPA payment dispute already passes a BaFin audit. Adding an LLM step to that process keeps it inside that audit perimeter.
Celonis, the Munich-headquartered process mining company that bought Make and Symbio in 2023, is the European company best positioned to industrialise the convergence. Process mining tells you which workflow steps are actually slow or error-prone. The automation layer then targets those specific steps with a workflow rebuild. The combination, when it works, gives an operator something the previous decade of RPA never delivered: a measured before-and-after on each automated step rather than a pile of bots running on faith.
What evaluation actually looks like in 2026
The European buyer trying to choose between these tools faces a set of questions the procurement teams of 2020 did not have to ask. Where does the model inference happen, and under whose jurisdiction. Whether the workflow steps that call a model log the prompt, the output, and the human reviewer in a way that satisfies the EU AI Act's transparency and record-keeping obligations under Article 12 and Article 26 for high-risk uses. Whether the orchestration layer can prove that a specific automated decision was reviewed by a human before it affected a customer.
The vendors that handle these questions cleanly tend to have one structural property in common. They run their own EU infrastructure, they document their model providers explicitly, and they expose audit logs as a first-class feature rather than an enterprise tier upsell. Camunda, n8n, and Celonis all clear this bar with caveats. The American vendors largely do not, and the answer "our EU customers run us on AWS Frankfurt" is not the same answer as "we are an EU-incorporated company operating to EU criteria."
The procurement teams that ask these questions before signing a renewal save themselves the conversation they would otherwise have with their DPO eighteen months in. The teams that do not ask discover, usually during an enforcement action that names a different company in a similar position, that the answer they assumed was true was never tested. The European workflow market is consolidating around the vendors who treat this as the first question rather than the last one.