Two of these components are not Google products. OpenWebUI and LiteLLM are open-source software that runs on Google Cloud infrastructure as Cloud Run workloads. The tool layer is a central registry on the Model Context Protocol with more than 20 MCP servers packaged in Artifact Registry, including a custom integration with dpa (Germany's national news agency) available to every agent as a standard tool.

Extensibility was the single most important requirement, and it drove the choice of the Agent Development Kit as the framework. Where the question was operations rather than flexibility, FUNKE went the other way: running the agent runtime itself on Kubernetes was considered and set aside, because Gemini already provides the scaling, session memory, and logging the team would otherwise have had to operate. That mix is the point—Google-managed services where operations are better handed over, open components where FUNKE wants to shape the behavior itself.

This architecture allows Paul and his team to "right-size" intelligence spending by running each task on the smallest model capable of completing it rather than the most capable model available. Many jobs given to AI have become too simple for the largest frontier models: questions about internal policies, lookups in business documents, and queries against research archives, for example.

These tasks make up the bulk of the platform's traffic and run on fast, low-latency Gemini Flash models. Since routing is managed by the central gateway, each agent can be equipped with the most efficient engine for its purpose and updated when a more cost-effective or advanced option becomes available.

At the opposite end sits the Deep Research Agent: a five-stage pipeline that runs up to 25 searches across FUNKE's own archives and licensed sources. It uses stronger reasoning models for the analytical stages and Flash for the high-volume ones. FUNKE estimates that a piece of research that would have once occupied an editor for hours now runs as a 9.7-minute session on average.

Nothing about FUNKE AI is compulsory, which makes the usage figures the clearest verdict on it. Around half of the company actively uses the platform each month, with users coming from every part of the business, including not only newsrooms but also accounting, human resources, legal, and controlling. More than 90% of them come back the following month. The target for the end of 2025 was 800 monthly active users, but the platform reached nearly double that number.

The architecture is evident in the traffic. From June to August 2026, 84.6% of all model outputs on the platform were generated on Google Cloud, totaling over 40 billion tokens, with the remainder running on a second cloud provider. Gemini models account for nearly 40% of the total, almost entirely the fast Flash family. The numbers show the benefits of right-sizing. Because FUNKE built on open standards, the composition remains its own to adjust—and today the large majority of everything runs on Google Cloud.

FUNKE has since modeled the productivity value the platform generates, arriving at a figure above €2 million annually—and the largest rollout is still ahead. Paul is deliberate about what that number is: not revenue, but headroom. How the organization uses the extra time is a management decision, whether that means increasing output or allocating more time to investigative work.

Rather than outsourcing the platform, the unit built it in-house, reasoning that a project like this is as much organizational as it is technical and that change is easier to manage from within. FUNKE runs learning programs on its own platform and maintains a network of AI ambassadors across the business, alongside dedicated AI functions in each division.

When asked what he would do differently with the knowledge he has today, Paul emphasizes speed over direction. Most decisions were correct, but some were made too slowly. This is a common pattern in complex technology projects. The direction was never in doubt.