Industrial Scale Agentic AI in Hospitality and Restaurant Analytics
Built on Azure, Fabric, and Databricks
“Business intelligence tells a restaurant operator what happened last quarter. It cannot tell them that service complaints cluster on Friday closing shifts at three specific locations. Closing that gap takes LLMs reading reviews at the attribute level — and agents that turn the finding into a task for a shift manager and a reply to the guest who wrote it.”
Session Overview
A conversation with David Giard on Technology and Friends (Episode 895) about building an industrial-scale agentic AI application for the hospitality and restaurant industry at Long View Systems. We covered real-time review analytics, why traditional BI stops short of the insight operators actually need, the agentic framework behind the product, and what it takes to run it for thousands of concurrent users across North America with secure per-client data isolation. Full episode on YouTube.
The Problem Statement
Hospitality runs on thin margins, and performance varies enormously — between locations, between shifts, sometimes between weeks at the same location. Operators are drowning in guest feedback but can’t act on it fast enough to matter: by the time a quarterly report surfaces a service problem, the shift that caused it is months gone. The project targets real-time review analytics — understanding what guests are actually saying, at the granularity where an operator can do something about it.
Traditional BI reports on what already happened and aggregates the nuance away. A four-star review that praises the food and criticizes the wait time lands in the same bucket as a flat four-star. The interesting signal is inside the sentence, not in the star rating.
Architecture & Technical Stack
- Data & Models: Microsoft Azure, Microsoft Fabric, and Databricks carry ingestion and the analytics layer. LLMs and natural language processing extract attribute-level sentiment from review text — separating how guests feel about food quality from how they feel about service, cleanliness, or value, rather than collapsing them into one score.
- Agentic Orchestration: Multiple agents work past analysis into action. One set surfaces and explains the pattern; another generates actionable tasks for staff tied to the specific location and shift; another drafts personalized responses to individual guests based on what that guest actually wrote.
- Scale & Isolation: A containerized architecture supports 5,000–10,000 concurrent users across North America, with stable, secure, and isolated data handling per client — multi-tenancy where one operator’s review data and model context never cross into another’s.
Key Takeaways
- The insight lives below the star rating: Attribute-level sentiment is what separates this from a dashboard. Knowing food good / service slow at a named location on a named shift is actionable; knowing 4.1 average is not.
- Analysis without action is still a report: The agentic layer earns its keep by producing the staff task and the guest reply, not by producing a better chart. That’s the line between a BI refresh and an operational system.
- Multi-tenant isolation is an architecture decision, not a config setting: Serving thousands of concurrent users across many clients means designing containment in from the start — isolation of data, context, and execution per client.
- Enterprise scale is where agentic AI gets hard: Demos handle one review. Production handles the concurrency, the tenancy boundaries, and the cost curve of running LLM inference across every review, continuously.
At the time of recording the application was in MVP evaluation, with a full-feature release planned for later in 2026.