"Union work is human work. It's the social contact, it's being there for people, it's helping them live face to face with the workers in the factories"
ABVV-Metaal believes in the power of human connection. But they face a practical problem: union delegates working on factory floors needed quick access to complex employment law information to help members with immediate questions.
Traditional resources - thick booklets, detailed websites, or waiting for expert advice - simply weren't fast enough when a worker needed to know their rights about leave entitlement, dismissal procedures, or work-life balance options. Union delegates, who are elected representatives still working their regular jobs, found themselves struggling to locate specific information quickly whilst trying to support members effectively.
The challenge was particularly acute in Belgium's multilingual environment, where delegates needed to provide accurate information in Dutch, French, and sometimes other languages, all while navigating complex employment laws that vary by sector and individual circumstances.
The solution
The union developed "Metal Bot" - a chatbot that provides instant access to sector-specific employment law information. Rather than replacing human contact, the system was designed specifically to enhance the work of union delegates by giving them rapid access to accurate information.
The chatbot uses OpenAI's GPT-4o model, trained on the union's own legal documents and guidance materials that were previously only available in hard-to-navigate booklets or buried on websites. "We made this specifically for union delegates that are elected and that are still in the factories doing the work there," the union representative explains.
Key features include automatic language detection and translation, downloadable conversation transcripts, and clear disclaimers about the bot's limitations. Each response includes a standard message directing users to contact union representatives for personalised advice or complex cases.
The system covers multiple sectors with different datasets, allowing users to select their industry or input their company name to access the most relevant information. Questions can be asked in any language, with the bot responding accurately in the user's preferred language.
Key benefits
Immediate access to information: Delegates can get instant answers to common questions about leave, dismissal procedures, and employment rights
Multilingual support: Automatic translation allows delegates to help members regardless of language barriers
Enhanced accuracy: Information comes directly from official union legal documents rather than delegate memory or interpretation
Improved delegate confidence: Representatives feel more prepared when supporting members with complex legal questions
Scalable learning: Each interaction helps improve the system's responses for future users
Time savings: Quick information retrieval allows delegates to focus on relationship-building and advocacy rather than research
Common questions
How difficult is it to implement?
"It's actually quite easy for someone to make something like that," according to the union's experience. The system uses OpenAI's assistant interface, requiring prompt engineering, data preparation, and careful temperature settings (they found 0.35 works best for empathetic but accurate responses).
What skills are needed?
Basic understanding of AI prompting, data preparation skills to convert existing documents into machine-readable formats, and ongoing monitoring capabilities to review transcripts and improve responses.
What are the main challenges?
Complex calculations proved problematic - the system couldn't handle intricate dismissal notice calculations that depend on multiple variables and date ranges. The union solved this by directing users to their existing calculation tool for such queries. Language detection can sometimes struggle with very short questions due to similarities between Dutch and English.
How do you ensure quality and accuracy?
Extensive prompt engineering with 16 different instructions, including restrictions on creative writing, translations, or comparisons with other sectors. Regular transcript reviews help identify areas needing improvement. Legal advisors reviewed the system before launch.
What about data protection?
Clear disclaimers warn against sharing personal information. The union uses OpenAI's paid subscription (which doesn't use data for training), deletes transcripts after review, and had legal advisors draft comprehensive privacy policies. Users can access the service without providing any personal details.
Getting started
First steps
Identify your most common member enquiries and existing guidance materials
Choose appropriate AI platform and subscription level that doesn't use your data for training
Develop comprehensive prompts that reflect your organisation's values and limitations
Ensure prompts prevent the system from comparing your sector with others or providing outdated information
Create clear disclaimers about the system's capabilities and data handling
Don't assume one model fits all - test different AI models with your specific data
Resources needed
Existing legal/guidance documents in digital format
Staff time for prompt development and ongoing monitoring
Paid AI service subscription for data privacy compliance
Legal review of privacy policies and data handling procedures
Looking ahead
The union plans to continue improving the system through regular transcript analysis and expanding to cover additional sectors. They're particularly focused on enhancing language detection capabilities and refining responses based on real-world usage patterns.
The success has reinforced their view that AI should enhance rather than replace human union work: "Every system that we are trying to develop is made to enhance and help the workers, and primarily the union delegates that are on the ground."
Future developments may include better integration with existing union tools and expanded coverage of complex scenarios that currently require human expertise.
AI transparency statement: This case study was written collaboratively using Claude.ai, which summarised the transcript of a video recorded by Unions 21 with Tim De Grom and formatted into a defined template for AI case studies.
