How-to questions, one by one
Customers opened a ticket for routine questions like “How do I add a recipe?” or “How do I update inventory?”, every time.
Ticket queue noise
Case study · Navi
Navi is an inventory management platform for restaurants. Chapeau! designed and built Chef Marco, an AI-driven virtual assistant embedded directly in Navi’s product: Claude-powered chat, live inventory lookups, and a model-agnostic backend built to extend.
Routine questions were indistinguishable from real problems. Every how-to question and every inventory lookup went through the same support queue as an actual issue, adding load without adding urgency.
Support meant
Customers opened a ticket for routine questions like “How do I add a recipe?” or “How do I update inventory?”, every time.
There was no way for a customer to get a factual answer about their own inventory without waiting on a person.
Navi already had a knowledge base, but it wasn’t comprehensive, and part of it lived in long training videos. Finding a specific answer meant already knowing where to look.
A canned chatbot window would not have moved the number. Chapeau! built Chef Marco on a backend that injects real Navi context into every message and gives the model tools to query live data, not just script responses.
Four pillars
A persistent chat button is built into Navi’s own nav bar. Every message is enriched with what Navi is before it reaches the model, so answers are grounded, not generic.
Tool-use lets Chef Marco query a customer’s own data directly, e.g. “List all my active wines”, scoped to only the restaurants that user is authorized to see.
Built to swap between Anthropic, Gemini or other providers without rebuilding, avoiding lock-in.
A help tier answers how-to questions like “How do I perform an inventory count?”, bound to Navi’s own documentation. A data tier answers factual questions like “How many inventory items do I have?”.
The first 13 questions are the start, not the ceiling. The infrastructure behind Chef Marco is built to take on more without a rebuild.
What it unlocked
The model integration, chat UI, conversation persistence, tool framework and error handling are built once, then reused by every assistant Navi adds after this one.
Each new use case added to Chef Marco increases how much support volume resolves itself.
If Chef Marco isn’t confident in an answer, it’s being built to say so and point the customer to a person by email or phone instead of guessing.