“It could genuinely jeopardize your business”: n8n’s argument for model-agnostic AI

      When asked about his company's position in the AI hierarchy, Jan Oberhauser analogized it to a car. The models, he explained to an audience at the Raise Summit in Paris this week, serve as the engine. The other components—the vehicle, the roads, and the regulations enabling the engine to actually function—are what he is focused on developing. In a year where the engine garners significant attention, this perspective is rather unconventional, and that’s precisely the point.

      Oberhauser founded n8n in 2019, positioning it as the orchestration layer for AI, or as he has referred to it in other contexts, the “Excel of AI.” The platform operates between the AI models and a company's existing systems, integrating large language models, deterministic code, and human approval into workflows that can withstand real-world application. Notably, the platform does not create its own models.

      This model-agnostic approach is not a recent marketing tactic; it is fundamental to how the company offers its software. n8n utilizes a Sustainable Use License, a "fair-code" model that the company developed and adopted in 2022 after transitioning from Apache 2.0 with a Commons Clause. The software is open-source and can be self-hosted, although commercial use is limited.

      According to Oberhauser, the design objective was to ensure that users can retain ownership of their data, self-host their solutions, and avoid vendor lock-in regardless of external influences. He argued that this freedom is particularly appealing to large organizations.

      In the past year, this notion transitioned from theory to tangible demand. As OpenAI, Anthropic, and DeepSeek launched new releases, companies rushed to integrate models into their operations, creating a demand for platforms that could connect any of these models to their existing systems.

      Currently, n8n boasts over 1,400 enterprise customers and approximately 1.7 million monthly active builders, with notable clients including Meta, Vodafone, and Mercedes-Benz.

      An interesting point, as Oberhauser highlighted, is that most customers rarely change their models. They desire the option to do so. He noted that switching can disrupt operations, as different models from various vendors—or even newer versions of the same model—can behave in ways that necessitate thorough evaluations before trust can be established.

      However, there are valid reasons for switching: scalability, latency, quality, or the ability to fine-tune an open-source model for cost and speed advantages. Most companies, he stated, are currently focused on building the infrastructure to facilitate future transitions, rather than implementing migrations immediately.

      Sovereignty has accelerated discussions around this issue. Oberhauser mentioned that boards are increasingly concerned about what would happen if a provider raises prices, gets acquired, or loses access in certain regions—a dependency that could “literally kill your company,” as he put it.

      The emergence of standardized connections for linking models to tools has made switching seem less like a research endeavor and more like a necessity, with European clients particularly emphasizing control over where their data and models reside.

      He suggested that the ability to transition “reasonably fast” from one provider to an open-source alternative is now acting as a form of insurance.

      This perspective is proving to be commercially beneficial. In May 2026, SAP invested in n8n at a reported $5.2 billion valuation and agreed to integrate its visual workflow canvas into Joule Studio, SAP’s environment for building agents.

      Meanwhile, Mercedes-Benz has implemented n8n as a global automation platform, partly attracted by its self-hosted and cloud-agnostic deployment that maintains sensitive data on its own systems.

      When discussing the shift from grassroots adoption to enterprise sales, Oberhauser described a deliberate strategy: distribute the product widely, allow workloads to accumulate, and then engage with large organizations when they require single sign-on, enterprise support, and other features. The more challenging message he aims to convey relates to return on investment.

      He referenced research indicating that only about 5% of corporate AI initiatives yield real value, mirroring MIT’s well-known “GenAI Divide” report, which found that 95% of enterprise pilots produced no measurable profit-and-loss impact.

      According to him, the successful companies view AI as one element of a broader system, integrating it with deterministic logic that is fast, affordable, and dependable, along with a final human evaluation. This approach is decidedly unglamorous.

      In a market still captivated by the engine, the company focusing on the car, the roads, and the regulations is betting that this less glamorous aspect is where the real profits will ultimately emerge.