Grounding a first version in a glossary and error pages so non-technical users get clear answers without the risk of an agent taking actions.

Support questions split into two groups: unique problems that need a human, and the same questions asked again and again. The second group is automatable — carefully.

The design

• Explain, never act. The assistant answers from a curated glossary and error pages. It cannot change keys, limits or data — no tools, no actions.
• Short prompt, mostly prohibitions: answer only from the provided pages, link the source, say "I don't know" when they don't cover it.
• Redaction first. People paste API keys into chat; key-shaped strings are stripped before text reaches the model or the logs.
• Easy handoff. One click posts the conversation to the team channel when the answer isn't enough.

The lesson: the safest first version of an AI feature is read-only. Ground it in documents you control, forbid actions outright, and measure resolution rate and handoff quality before expanding what it's allowed to do.