The first question most Geschäftsführer ask when considering enterprise AI is "should we build our own models?" It is the wrong question. The right question is: "Where does our competitive advantage actually lie — in the model or in the workflow?"
For the vast majority of DACH Mittelstand companies, the answer is clear. The model is a commodity. The workflow integration is the asset. And yet, a remarkable number of organisations spend six to twelve months and €200K+ building custom models when they could have reached production in eight weeks with a commercial model and custom integration.
The build-vs-buy framing is outdated
The traditional build-vs-buy decision assumes you are choosing between two comparable paths to the same outcome. Build takes longer and costs more, but you own the result. Buy is faster and cheaper, but you depend on a vendor.
For enterprise AI in 2026, this framing misses the point entirely. The "build" is not one thing — it is three:
- The model — the AI that performs classification, generation, extraction, or prediction
- The integration — how the model connects to your data, processes, and systems
- The operating model — how humans and AI work together in the modified workflow
Most companies that "build" are building all three from scratch. Most companies that succeed in production have bought component one and built components two and three.
Why custom models rarely make sense for the Mittelstand
Custom model development — fine-tuning foundation models on proprietary data or training models from scratch — has a specific and narrow set of use cases where it is justified:
- You have a genuinely proprietary dataset that creates competitive advantage (rare)
- Commercial models cannot achieve the required accuracy for your specific domain (increasingly rare)
- Regulatory requirements mandate that models are trained and hosted on-premise (sometimes true, often overstated)
For a €50M industrial supplier deploying its first AI workflow — say, classifying incoming orders by product category and urgency — none of these conditions apply. A commercial language model with well-designed prompts and proper integration will outperform a custom model that took six months to train, because it reaches production faster, iterates faster, and costs a fraction.
The numbers are stark:
- Custom model: €100–500K, 6–12 months to production, ongoing retraining costs
- Commercial model with custom integration: €20–80K, 6–12 weeks to production, API costs that scale with usage
The cost difference is significant, but the time difference is critical. In the time it takes to build a custom model, a company using commercial models can deploy three to four production workflows, learn from real-world usage, and build the operational capability for Level 2 integration.
Where the real value lives
The value in enterprise AI is almost never in the model. It is in three places:
Workflow design. Which process gets AI? How is it scoped? What constitutes a good output? These decisions determine 80% of the ROI. Get the workflow wrong and no model — custom or commercial — will save the initiative.
Data integration. Getting the right data to the model and the model's output back into your systems is the hard engineering problem. This is specific to your organisation: your ERP, your document management, your approval workflows, your compliance requirements. No vendor can do this for you. This is where you build.
Operating model. How does the team work with AI output? What gets auto-approved? What gets human review? How are exceptions handled? This is organisational design, not technology — and it determines whether the deployment sticks or gets abandoned within 60 days.
Custom model training is none of these things. It is a technology investment that solves a problem most Mittelstand companies do not have.
The integration layer is your moat
If you accept that the model is a commodity — and for most use cases it is — then your competitive advantage comes from how well you integrate it into your specific workflows, with your specific data, under your specific compliance constraints.
This integration layer is inherently custom. No two organisations have the same ERP configuration, the same approval processes, or the same compliance requirements. Building this layer well requires deep understanding of the workflow, the data landscape, and the operating model. It is the kind of work that the Accelerator engagement is designed for.
The integration layer is also what makes switching costs real. A well-integrated AI workflow that is woven into daily operations is sticky — not because of the model (which can be swapped), but because of the integration (which represents real engineering and organisational investment).
When building does make sense
There are legitimate scenarios for custom model development, and they should not be dismissed:
Highly specialised domains. A medical device manufacturer classifying defect types from microscopic images may genuinely need a custom vision model trained on proprietary defect data. A general-purpose model will not have seen enough examples of their specific defect taxonomy.
Performance-critical applications. When the difference between 94% and 98% accuracy has significant financial impact — e.g., fraud detection in payment processing — custom models trained on proprietary transaction data may justify the investment.
Regulatory mandates. Some industries and some interpretations of the EU AI Act require models to be hosted on-premise with full auditability of training data. In these cases, custom models may be a compliance requirement rather than a technology choice.
Even in these scenarios, the recommendation within the AI Operating System methodology is to start with a commercial model, prove the workflow works, and then invest in a custom model where the proven workflow demands it. Prove value first, optimise second.
A decision framework
For any AI initiative, ask these four questions in order:
Can a commercial model achieve adequate accuracy for this workflow? Test it. Most organisations assume they need custom models without actually testing commercial alternatives. Run a pilot with a commercial model. Measure accuracy against your definition of "good output." If it passes, stop deliberating.
Does the time-to-production difference matter? If a custom model takes 9 months and a commercial model takes 9 weeks, what is the cost of the 7-month delay? Include the cost of delayed operating leverage, not just the development cost.
Where is your engineering capacity better spent? Every hour spent on model training is an hour not spent on workflow integration, data pipelines, and operating model design — the things that actually determine whether the deployment succeeds.
What is the switching cost if you start commercial? Almost always low. Well-designed integrations abstract the model behind an interface. Swapping from one commercial model to another — or from commercial to custom — is a weeks-long effort, not a months-long rebuild. For a detailed framework on evaluating vendors and managing lock-in risk, see AI Vendor Selection.
The Mittelstand advantage
The Mittelstand's typical constraint — limited AI talent, pragmatic budgets, fast decision-making — is actually an advantage in the build-vs-buy decision. Companies that cannot afford a nine-month custom model project are forced into the faster, more pragmatic path: buy models, build integration, get to production, learn from real usage.
This path produces more operating leverage per euro invested than the custom-model path. And it builds the organisational capability — the six dimensions — that makes Level 2 and Level 3 achievable.
To discuss whether your next AI initiative should build, buy, or — most likely — do both in the right proportions, book a Fit Call.
This article is part of the AI in Operations series by Andreas Anding. For the full methodology, see The AI Operating System.