Case 01 · AI agent
Product Compliance Agent
The agent answers regulatory questions using a defined set of source documents. Its current coverage includes Spain, Italy, and Germany, along with EU requirements for data minimization.
Each confirmed answer links to the relevant document version and section. If the sources do not cover a country or situation, the agent reports the gap. Its validation rules reject confirmed answers without citations.
The business problem
Take a guest data management system for hotels. One of its jobs is legal compliance: collecting guest data and sending it to the police. “Collect and send” means something different in every country. One country asks for one set of fields, another asks for a different set. Some require an electronic signature, some don’t.
The time-consuming part is explaining why each field is required. Say the police fine a hotel for collecting too much data: the passport number is enough, and there’s no need to ask for citizenship. The hotel goes to the software provider. Someone at the provider has to find the relevant provisions: another part of the same law says you must not collect data on the country’s own citizens, only on foreign guests — and to find out whether a guest is foreign, you have to ask for citizenship. That resolves the question. The next day, another hotel raises a similar issue in a different country.
Each question takes hours to resolve, and the whole company depends on one person who understands the requirements.
Sales and support face a similar problem. The product is complex, with many features, integrations, technical specifications, and country-specific requirements. Even with documentation available to everyone, colleagues turn to the same expert because it’s faster than searching for an answer themselves. In a sales meeting, this can undermine trust: a client asks about an integration, and the salesperson has never heard of the service.
The same kind of agent could work here: same principle, different document set — product documentation, specifications, integration descriptions, pricing, and release history instead of legislation. The rule about sources stays the same — the answer says which document it came from, and when there is nothing in the document set, the agent says so. Sales teams have a particular need for this: a confident wrong answer to a client costs more than “Let me check and get back to you.” Two main scenarios:
- Help me prepare for a client meeting: find the relevant services and explain how they work.
- Check whether our product supports what this client has requested.
Next steps
A source library maintained by qualified professionals. Web search can surface sources of uneven quality. Regulatory work needs more reliable source selection. The source library should track document versions and effective dates, with a qualified lawyer responsible for its content.
Query logs. If you save what people ask about, after a month you can see what is unclear to users, where they get stuck again and again, and which parts of the documentation are missing. These logs could also help the people responsible for sales training and documentation decide what to improve.
I built something similar in the early days of GPT assistants: a sales chatbot for online stores. It worked, but the token costs were too high, so we didn’t launch it. Since then the economics have changed and the problem is still there. I want to keep building tools that connect business needs with content and AI.