At Rise of AI 2025 in Berlin, Jürgen Brinkmann, co-lead of the AI Center of Excellence at Windmöller & Hölscher (W&H), and Eileen Neumann of SVA System Vertrieb Alexander present WINNI, a penguin-branded AI assistant W&H built to help ordinary employees, not just engineers, adopt AI at work. Brinkmann describes W&H as a 150-year-old German machine-building company facing growing complexity from regulation, IT sprawl and dense process documentation that employees struggle to use. His team built a multi-agent system that translates complex process data, including JSON-based process diagrams, into plain answers, but found employees still avoided tools branded around GPT because they did not understand or trust it. Their answer was WINNI, a young, imperfect penguin that lives inside Microsoft Teams, never criticizes users for asking the same question repeatedly, and became something employees say they love rather than merely use. Neumann then explains that although WINNI does not qualify as high-risk under the EU AI Act, W&H still applies AI governance practices she calls LLMOps, covering prompt and context versioning, model comparison, hallucination detection and human oversight, supported by a dedicated governance tool.
Key takeaways
- Windmöller & Hölscher built a multi-agent system with a delegation agent to translate complex, diagram-heavy process documentation into plain-language answers.
- Employees resisted a tool branded around GPT because many did not understand the term, so the company built a mascot, WINNI, instead of a 'W&H GPT'.
- WINNI lives inside Microsoft Teams rather than a new app, was deliberately made a young, imperfect penguin, and employees describe loving it rather than merely using it.
- Neumann says WINNI falls outside the EU AI Act's high-risk category, since no one's life or livelihood depends on its output.
- W&H still applies AI governance practices Neumann calls LLMOps, including prompt and context versioning, hallucination detection and human-in-the-loop review, via a dedicated governance tool.
Detailed notes
A 150-year-old machine builder facing rising complexity
Brinkmann introduces Windmöller & Hölscher as a German Mittelstand company founded on steam-powered machines more than 150 years ago and now a global leader in machines that make packaging to protect goods and food, with more than 2,000 employees in Germany and over 1,500 more worldwide, and revenue that has doubled over the last decade. He describes rising complexity from regulation, technical change, global conflicts and IT systems, noting employees already juggle 15 to 16 different systems and resist being handed yet another one. He says the company had heavily documented its processes as diagrams, which he as a manager found reassuring but employees found unusable: 'who as a human wants to have a question answered by looking at a wallpaper of diagrams?'
A multi-agent system to make processes understandable
The team built a system, developed with SVA, that uses large language models to turn complex process data, including process diagrams stored as JSON files, into understandable answers, going beyond simple retrieval-augmented generation. It uses a multi-agent architecture with a delegation agent that routes a user's question, such as 'what is my next step' or 'is the software release already approved', to the specialized agent that holds the relevant process or database knowledge, and returns an answer through one interface.
Why the interface had to be a penguin, not a chatbot called GPT
Despite the system working technically, Brinkmann says employees still would not adopt it, and user-experience research and works council conversations showed that most people simply did not understand what 'GPT' meant and were wary of a tool branded that way. The team decided against building a separate 'W&H GPT' and instead created WINNI, an assistant that lives inside Microsoft Teams, the platform employees already use, rather than adding a new system. They deliberately made WINNI a young penguin so mistakes feel forgivable, designed her to always answer kindly even when asked the same question repeatedly, and made it acceptable for her to say she does not know. Brinkmann says employees now say 'I'm loving Winnie,' a response he contrasts with how people talk about tools like Copilot or SAP, and the company reinforced the mascot with branded cups, t-shirts and stickers to build internal AI literacy.
Where WINNI is headed next
Brinkmann outlines two next steps: giving WINNI to the more than 400 service technicians who travel over 250 days a year and rarely see their families, so they can access knowledge on the road, and extending WINNI to customers operating W&H's large industrial machines (some 15 meters long and 7 meters high), letting them ask questions like how the last shift went or how to improve it, drawing on industrial IoT data, a concept already shown at a trade fair.
Q&A highlights
Neumann explains that WINNI does not fall under the EU AI Act's high-risk category, since no one's life or livelihood depends on its output and the only strict requirement is transparency that it is an AI, not a person. She argues AI governance is still worth applying as best practice for complex production AI systems, describing an emerging discipline she calls LLMOps that extends MLOps practices like model, data and code versioning with new demands specific to large language models: tracking prompts and context windows as closely as model code, accounting for different models having their own 'personality', handling non-deterministic output, and adding human-in-the-loop review and hallucination detection on top of standard quality metrics such as ROUGE and BLEU. She describes a governance tool that sits as a layer over any infrastructure and lets teams define approval workflows and track models, metrics and risk against both the AI Act's risk catalog and internal policy. Asked how W&H measures whether change management around tools like WINNI actually succeeds, Neumann says it is largely qualitative: success shows when people start talking about and using a tool voluntarily rather than because they were told to, which also helps make the case for the next project.
Quotes
Who as a human wants to have a question answered by looking at a wallpaper of diagrams?
Jürgen Brinkmann
Questions answered in this talk
Why did Windmöller & Hölscher build a penguin mascot instead of a branded GPT tool for employees?
According to Jürgen Brinkmann at Rise of AI 2025, user research and works council discussions showed employees did not understand what 'GPT' meant, so the company built WINNI, a young penguin living inside Microsoft Teams, to make AI feel approachable rather than adding another unfamiliar system.
Does the EU AI Act classify an internal assistant like WINNI as high-risk?
Eileen Neumann explains at Rise of AI 2025 that WINNI does not count as high-risk under the EU AI Act because no one's life or livelihood depends on its output, though it still must be transparent about being an AI rather than a human.
What is LLMOps and how does it differ from MLOps?
Eileen Neumann describes LLMOps at Rise of AI 2025 as extending traditional MLOps practices like model, data and code versioning with tasks specific to large language models, such as tracking prompts and context windows, comparing different models' distinct behavior, and adding hallucination detection and human-in-the-loop review because outputs are non-deterministic.