What Does It Mean for AI to Be Working?
Mon 27 Jul 2026
Across the UK, Germany, and France, this month’s AI in Practice coverage is asking a practical question: once AI reaches production, how do organisations know it is actually working?
Availability is only part of the answer. An AI system can be online and still produce inaccurate answers, rely on outdated information, create unexpected costs, or take actions that are difficult to trace.
This week’s regional perspectives look at that problem from three angles. In the UK, the focus is on behaviour, cost, risk, and business outcomes. In Germany, the question is whether an AI-generated answer can be traced back to reliable and current information. In France, the discussion extends to the infrastructure underneath AI and the degree of trust, control, and choice organisations retain over the systems they depend on.
UK: A Healthy System Can Still Be Wrong
The UK perspective starts with a problem traditional monitoring was not built to solve.
Infrastructure dashboards can tell teams whether a service is available, how quickly it responds and whether hardware is operating normally. They cannot always tell them whether an AI-generated answer is useful, correct, or safe.
Insights gathered through DevOps Live at Tech Show London highlighted the wider set of signals organisations may need once AI enters production. These include latency, GPU utilisation, cost per query, hallucination rates, retrieval quality, model drift, human corrections, and security findings.
What became apparent is that AI does not always fail visibly. A model can continue generating convincing responses while the quality of those responses deteriorates. An agent can complete a task while taking an action that introduces an unacceptable security risk.
Monitoring therefore becomes a question of behaviour as much as availability.
Germany: Can You Prove Where the Answer Came From?
The German perspective takes the question deeper into the information behind an AI response.
A position paper from Germany’s KI Bundesverband has argued that knowledge graphs can help organisations connect information held across systems and make the sources behind an AI-generated answer easier to identify.
That would allow a business to establish what an AI system said, as well as which information contributed to the answer, when that information was current, and who was responsible for it.
The approach does not remove the underlying data problem. Knowledge graphs still need to be maintained, and information still has to be updated and corrected. Responsibility for quality and currency still has to sit somewhere within the organisation.
The bigger point is simpler: AI observability does not stop at the model if organisations cannot see the knowledge behind its answers.
France: Trust Extends to the Infrastructure Underneath AI
The French perspective broadens the question again.
An analysis of Europe’s emerging sovereign technology landscape argues that digital sovereignty is now operational rather than ideological. Organisations are weighing data sensitivity, contractual dependence, reversibility, regulatory risk, and chain of custody alongside performance and price when choosing technology infrastructure.
In France, that discussion includes trusted cloud requirements and SecNumCloud. Across Europe, it also runs through initiatives and legislation such as GAIA-X, the EU AI Act, the Data Act and NIS2.
The implication for AI is straightforward. Proving a system works cannot be separated entirely from the infrastructure on which it runs.
An organisation may be satisfied with an AI system’s outputs and response times while still needing to know where its data is processed, who controls access, how dependencies are governed and whether it can move away from a provider if circumstances change.
That makes infrastructure choice part of the governance question.
AI Observability Needs a Wider Definition
The UK, German, and French perspectives test AI performance at different layers.
The UK asks whether the system is behaving as intended, Germany asks whether its answers can be traced back to reliable information, and France asks whether the infrastructure underneath it provides enough trust and control.
Some things do not change; however, conventional monitoring still matters. Availability, latency, and infrastructure performance remain essential, but they are not enough on their own.
As AI becomes embedded in business processes, organisations need to know not only whether a system is running, but whether its outputs remain useful, where those outputs came from, what they cost, what risks they introduced, and whether the infrastructure underneath them can be governed with confidence.
Deployment proves an AI system can run; what is harder to prove is whether it should keep running.
