In June 2026, INXM emerged from stealth mode with a €5.7 million pre-seed round – led by Cherry Ventures and Redstone, with participation from Angel Invest and Linden Capital. The Berlin-based startup is building “Orchestrator,” an AI process execution engine for enterprise and Mittelstand operations, based on a new architectural principle: Compiled AI. In this model, AI first designs the optimal process; actual execution is then handled by deterministic software – auditable, repeatable, and free of token costs at runtime.

INXM was founded by a team rooted in aerospace and enterprise software: CEO Alex Oelling previously served as Chief Digital Officer at Isar Aerospace and Volocopter and before that founded Sensorberg. In this interview, he explains why conventional LLM approaches reach their limits in complex business processes, how Compiled AI works in practice – and why experience from regulated, safety-critical industries matters for industrial AI.

INXM emerged from stealth mode with a €5.7 million pre-seed round. Many AI tools regenerate on every run — with the well-known risks: high cost, hallucinations, unpredictable results. How does INXM solve this problem — and why is it so critical specifically for industrial processes?

The core problem is mathematical. An LLM with 95% accuracy drops below 50% correct execution after 14 steps — after 175 steps, the probability of success is 0.01%. For a business process with many sub-steps that runs hundreds of times, that's unacceptable. Then there's the cost factor: a standard SAP transaction requires 100,000 tokens or more with a conventional LLM approach. Our compiled execution achieves the same transaction with 1,000 to 1,250 tokens — and zero tokens at runtime. That's a factor of 80 on a process that runs thousands of times a day.

Hallucinations are particularly dangerous. An AI system that's occasionally wrong doesn't reduce operational risk. It adds a new one. INXM doesn't solve this with better models, but with a shift in architecture: Compiled AI. The AI designs the process. A human or AI agent reviews it, and deterministic software handles execution.

With “Compiled AI,” AI first designs a process that is then executed deterministically — following a fixed sequence. How does this work in practice, and where are the limits of this approach?

The principle is simple: let the probabilistic system build the optimal path. But don't let it become the execution engine itself. It starts with a prompt in natural language. From that, the planner creates a typed, executable plan. The system then checks schema, security, and governance against live interfaces. In the healing loop, a dry run follows with no side effects. If the system detects a deviation — a field in SAP, an API version, or a status value — it recompiles the affected sub-step and proposes the corrected plan for approval. The AI doesn't improvise at runtime; it repairs the plan without giving up the determinism guarantee.

The result runs with no further token cost, no variance, and is fully auditable: Run 1 is identical to run 1,000. Economically, that means 80 to 100 times fewer tokens; the break-even point is 17 transactions. The compilation phase remains subject to LLM context-window limitations, so very complex problems have to be broken down. Even if models become more reliable, the cost structure remains decisive for us. Right now, thinking and execution both happen inside the LLM. We separate them. It stands to reason that this becomes the standard going forward.

Your Orchestrator is meant to connect existing systems like ERP, PLM, MES, and QMS. What technical and organizational prerequisites do companies need to bring for that integration to actually succeed?

Technically, the hurdle is manageable. The INXM Orchestrator is built on the open MCP standard, already connects to many enterprise systems out of the box, and doesn't replace existing systems — it connects them. All credentials stay with the customer. SAP remains the system of record for master data, the PLM for design states. The Orchestrator never becomes the system of record.

The real hurdle is organizational. Companies need three things: first, a data strategy — structured and unstructured data need to be consolidated, cleaned, and equipped with clear access rights. Second, process experts, not primarily software developers. Third, the willingness to build internal champions. INXM doesn't compete with SAP. The Orchestrator is what makes SAP investments usable in an AI world.

Consistent and verifiable results are especially important in industry and regulated sectors. How do you make decisions and process steps traceable — and when does human intervention remain indispensable?

Traceability isn't a feature add-on for us — it's an architectural principle. Every transaction is deterministic: same input, same output, every single time. Everything gets logged — what ran, where it stopped, why. That enables complete audit trails and rollback capability. When something fails, the Orchestrator stops at the point of failure, generates a proposed fix, and hands control back to the responsible person.

Human-in-the-loop isn't a workaround — it's a design principle. Human intervention remains essential wherever decisions fall outside the defined rule set, regulatory compliance requires an explicit sign-off, or the system can't heal itself. The system actively recognizes these situations and routes tasks to the relevant departments.

Which workflows do you see as particularly well suited for AI automation — and where would caution still be warranted today?

The simple formula: high volume, low interpretive latitude, high consequences for errors. That's the sweet spot. Cross-system processes that are manually coordinated today are particularly well suited: engineering release and change management in PLM systems, invoice processing and three-way matching in finance, procurement workflows from supplier to SAP, and BOM analysis in manufacturing. These are critical processes that nobody really wants to do and where reliability isn't negotiable. Our existing test use cases also show that strategic competitive analysis and much more can be mapped well.

Caution is warranted wherever clear success criteria are missing: decisions without a solid data basis, creative work with an open-ended outcome space, processes without existing system integration, or situations where accountability for the outcome isn't clearly assigned.

Your founding team brings experience from aerospace, deep tech, and enterprise software. What lessons from that background carry over to industrial AI — and what doesn't?

Four founders, four industries — and all four of us stumbled onto the same problem, just from different directions. With a rocket, there's no margin for error: you can't retroactively fix a failed launch. With an air taxi, every step has to be certifiable. In manufacturing, reliability at scale is what counts. I was Chief Digital Officer at Isar Aerospace and Volocopter and before that founded Sensorberg. Our CTO Matthias Kainer brings more than 25 years of software experience, CPO Jesper Bylund comes from design roles at Volocopter and n8n, and CSO Kamil Klüber brings over 20 years of engineering and manufacturing experience, including nearly seven years at Siemens.

What carries over above all is this: we don't tolerate ambiguity at critical points. A system that sometimes works and sometimes doesn't is a risk — for a rocket launch just as much as for a ten-million-euro SAP posting. What also carries over is thinking in terms of requirements rather than solutions: Who writes the requirements? Who's accountable? Who signs off? What doesn't carry over is the sense of time in regulated industries: certifications take years. Enterprise software has to learn faster.

INXM was founded in the Berlin-Brandenburg capital region and targets European industrial companies. What strengths does the region offer for this approach? What's needed for this to become internationally successful enterprise AI?

Berlin-Brandenburg isn't a random location for us: this is where our roots and network are. The connections that now turn into partnerships, customer relationships, and hiring pipelines were built here. We're building something that doesn't exist yet in this form — and that requires an environment bringing together industry, research, and capital. The region gives us access to investors, decision-makers, and events where real conversations happen. Our first partnerships came from personal encounters.

On top of that, there's deep-tech research such as the DLR aerospace site in Adlershof, industrial settlement like the Tesla Gigafactory in Grünheide, Germany's most active startup ecosystem, proximity to federal policymaking, and strong connections to southern Germany's industrial region. That also shapes expectations: European industrial customers demand data sovereignty, compliance by design, and works-council readiness. INXM wants to become the infrastructure layer that European industry runs on. Invisible when everything works. Indispensable when it matters most.

Thanks for the great conversation.

Note: This interview was originally conducted in German and later translated into English language.