Monthly thinking room
How much control does AI need when voluntary use is not enough and models change constantly? This edition shows how companies organize commitment, permissions and review without tying themselves to individual vendors.
For working with your own AI assistant. How it works
We handpicked 38 podcast episodes released in April 2026 and distilled 114 insights from them. These ten findings bring together recurring experiences from the selected conversations.
Deutsche Welle runs a unified web interface with a fixed URL, behind which language models are swapped via API, and keeps a local model on its own servers as a fallback.
Despite its Office 365 environment, GUS Group chose Langdock over Copilot Studio because Langdock bundles all models in a closed, GDPR-compliant environment.
At Blackboat, Christoph Magnussen built a code-based layer prompt system in which changes to a central prompt are automatically carried over into around 100 flows, instead of adjusting each flow individually when the model changes.
Supported by these conversations
Denkwerk gave around 70 people in software development free access to AI in the IDE and is now managing the role change over 12 months with training and role clarity, because without an active change process nothing happened across the board.
At the Key Academy, all employees have their own Claude account and are required to ask Claude, for every task, how the agent would complete it.
GUS Group had management select 20 champions instead of relying on volunteers and gives each of them one person-day per week; KPMG requires all partners and directors to complete a Digital MBA.
Supported by these conversations
Thorsten Leimbach (Fraunhofer IAIS) assigns emails and recorded team meetings to standard products from Microsoft or Google and, for core business data, advises companies to use their own or open models and servers in their own basement.
Because of its promise regarding customer data, Finanzinformatik rules out US-hosted services such as ChatGPT and runs the SKI Pilot with open-source models and NVIDIA as a partner in its own data centers.
Harting selects the LLMs in its internal GPT by use case, country or security level; Christoph Magnussen transcribes highly confidential customer conversations locally with Mac Whisper and summarizes them with open-source models in LM Studio.
Supported by these conversations
With a subscription
With a subscription you read all 10 with their sources and download them as a text file.
The complete collection
38 conversations, ordered from newest to oldest.
Architecture before AI tools
Core thesis: AI as an invisible control layer built from Slack and Claude Code beats a collection of 200 SaaS tools, and a central prompt system displaces visual no-code flows. Relevant for decisions because the architecture question comes before any tool selection.
With a subscription
With a subscription you open all 38 conversations of this room with their insights, search the catalogue and download the files.
First edition. Future editions will document changes to the podcast catalogue, key findings and recommendations.
