A concierge agent becomes useful when it can tell the difference between a stable question, a recommendation request and a moment that needs a person. That is an architecture and operating decision, not just a prompt-writing exercise.
What the published project demonstrates
NFI built Inndira's Boutique Hotel Concierge around three conversational paths: direct answers about policies, neighbourhoods and comparisons; a curated property search blended with live web lookup; and consent-first contact capture with CRM hand-off by webhook. The published case identifies a LangGraph agent architecture, a curated knowledge base and fallback across multiple model providers.
The case also documents a proprietary knowledge base of 79 hand-enriched boutique hotels, more than 480 room types and over 100 structured fields per property. It reports three curated picks rather than 200+ generic results, under two minutes from first message to personalised recommendations, and 99.9% uptime with fallback across four providers. Those are Inndira's published project facts, not generic concierge guarantees.
Give each question a source and a path
A proposed concierge design can start with a small routing map. Stable business facts may come from approved source material. Recommendations may require structured selection criteria and an explanation of why each option was chosen. Live information should be identified as live information, with a defined response when it is unavailable or cannot be verified.
- 01Define the jobs before the graph. Separate direct questions, discovery or recommendation requests, contact capture and escalation. Each job needs permitted actions, source boundaries and a completion condition.
- 02Curate before retrieving. Decide which material is approved, who updates it and what fields support a comparison. Retrieval quality cannot compensate for stale, ambiguous or unowned source information.
- 03Preserve the reason for a recommendation. Capture the relevant preferences, source evidence and selection criteria so a person can understand the result and correct it where needed.
Plan fallback as an operating behaviour
Fallback is more than selecting another provider. For each dependency, decide what can be retried, what needs a simpler response, what should be queued and when the agent should say that it cannot complete the request. Record enough context for a team member to continue without asking the visitor to repeat everything.
Before launch, test incomplete requests, conflicting preferences, unavailable sources, a failed hand-off and a request outside the concierge's scope. Monitor the paths that end in a hand-off as carefully as the paths that end in an answer. A reliable service has a useful next step even when one component is unavailable.
Treat consent and CRM hand-off as part of the journey
If a concierge collects contact details, agree the consent language, validation, data destination and ownership of follow-up before the interface is designed. Inndira's published implementation used a consent-first contact flow, email validation and CRM delivery by webhook. Another organisation should validate its own privacy, consent and retention requirements rather than copy that pattern unchanged.
The best concierge path is not the one that answers everything. It is the one that knows which source to use, what it can safely say and how to hand off the rest.
Build a smaller, inspectable first version
Start with the repeat questions and a narrow, approved source set. We can help map the conversation paths, evidence boundaries, integration hand-offs and review process for an agent that has a clear job in your operation.
