The right question is not “which AI”, it is “which problem”
Over the last few months, almost every German utility has received AI proposals from hyperscalers, software vendors and consultancies. Azure OpenAI, AWS Bedrock, Google Vertex AI, open models such as Llama or Mistral, SAP-embedded AI or vertical platforms for energy. Each comes with its own narrative, expected ROI and licence model.
The problem is that few of these conversations start with the right question. Before choosing a technology, a German utility should answer three others: which processes it actually wants to improve, which data it can legally use (GDPR, BSI, KRITIS) and how much control it needs over the model itself. Only then does the tooling decision make sense.
Being AI-agnostic means exactly that: putting these questions before the stack.
What AI-agnostic means at Principal33
It is not marketing neutrality. It is an operating model with three practical principles:
-
- Use-case-based selection, not contract-based. A closed model such as GPT-4 might be optimal to summarise regulatory letters; an open-source model deployed on-prem might be the only viable option for KRITIS-bound metering data. We decide case by case, not by catalogue.
- Portability by design. When we build an AI pipeline for a utility, we keep orchestration, data layer and model layer cleanly separated. If the provider changes tomorrow —because of price, regulation or performance— the client does not have to rewrite the solution end-to-end.
- Transparency on cost and lock-in. AI projects are not just monthly licences. There are inference costs, initialisation costs, egress costs and switching costs. We make them explicit upfront, not after go-live.
Why this matters in German utilities
German utilities operate in an environment where the cost of choosing the wrong technology is high. Three factors amplify the risk:
-
- Moving regulation. Bundesnetzagentur, BSI and KRITIS requirements evolve every year. An AI solution trained today on public cloud data may have to migrate to a restricted environment tomorrow. If the architecture is welded to a single vendor, that migration is painful.
- Data sovereignty. Metering data, customer data and market data (MaKo, EDIFACT, GPKE) carry strict constraints. Not every model is allowed to process them and not every hosting setup is compliant. An AI-agnostic partner knows which stack passes which filter.
- Long investment cycles. A SAP IS-U system or a utility CRM typically lives 8–15 years. Any AI that connects to them must survive several model generations without forcing complete rewrites.
What AI-agnostic is NOT
The term is being used very loosely lately, so it is worth being explicit:
-
- It is not “we use whatever AI the client asks for”. In some use cases we explicitly recommend not to apply AI, or recommend a single option as the right one. AI-agnostic includes saying no.
- It is not “we have no partners”. We work with Microsoft, AWS and specialised vendors. The difference is that we are not obliged to push a single stack to meet a sales quota.
- It is not “total abstraction”. There is no magic layer that makes all models interchangeable. What does exist is architectural discipline that minimises switching cost.
How we apply this with German utilities
Real engagements with German utilities follow a stable pattern:
-
- Joint technical-regulatory discovery between senior architects and utility-domain experts (SAP IS-U, MaKo, billing).
- Cross-evaluation of at least two stacks per use case, scored against cost, latency, compliance and portability.
- Short pilots (4–8 weeks) with metrics agreed beforehand.
- Industrialisation only if metrics are met. Otherwise the learning is documented and the case is closed without scaling.
The teams executing these projects are senior, German-speaking and nearshore DACH, which reduces regulatory misunderstandings and accelerates coordination with IT, regulatory and business inside the utility.
Conclusion
For a German utility starting its AI journey, the biggest risk is not picking the “wrong” model. It is locking the architecture in a way that prevents future change. An AI-agnostic partner protects that flexibility and turns it into a competitive advantage over the full life of the system.
