AI Assistant & Knowledge Base

The Case for AI in Hospitality Communication

A 30-bed surf camp in Ericeira receives roughly 80–120 guest messages per day during peak season. Of these, 40–60% are routine questions with predictable answers: 'What time is check-in?', 'How do I get from the airport?', 'What should I bring for surf lessons?', 'Is breakfast included?', 'What's the Wi-Fi password?' Every staff member answers these same questions dozens of times per week, crafting slightly different responses each time, sometimes giving inconsistent information because the check-in time changed last month and not everyone got the memo.

Traditional knowledge base tools like Zendesk Guide or Intercom Articles solve part of this problem — they let you publish help articles that guests can search. But guests don't want to search a help centre. They want to ask a question on WhatsApp and get an immediate answer. The gap between a static help centre and real-time conversational messaging is where AI assistants add transformative value.

Artidal's AI Assistant bridges this gap by combining a maintained knowledge base with multi-provider large language models that understand natural-language questions, retrieve relevant knowledge base articles, and generate contextually appropriate responses — all within the unified inbox, across every channel, with confidence scoring that determines whether the answer is sent automatically or presented to an agent for review.

Multi-Provider AI Architecture

The AI layer supports multiple foundation model providers — Anthropic's Claude, OpenAI's GPT, and Google's Gemini — with per-organization configuration. Operators choose their preferred provider based on cost, speed, language support, or compliance requirements. A European retreat centre concerned about data sovereignty can select a provider with EU data processing guarantees, while a budget-conscious surf camp can choose the most cost-effective model for their query volume.

Each provider is abstracted behind a unified interface, so switching providers doesn't require reconfiguring the knowledge base, retraining workflows, or adjusting confidence thresholds. The system maintains provider-level analytics — response accuracy, latency, cost per query — so operators can make informed decisions about which model delivers the best results for their specific question patterns.

The AI doesn't operate as a black box. Every suggested response includes a confidence score (0–100%) and citations to the specific knowledge base articles used to generate the answer. Operators configure two thresholds: an auto-send threshold (e.g., 85%+) above which the AI sends the response directly to the guest, and a suggestion threshold (e.g., 60–84%) where the response is presented to the agent for one-click approval or editing. Below the suggestion threshold, the AI stays silent and the agent handles the query manually. This graduated approach builds trust incrementally — operators start with conservative thresholds and widen them as they see the AI perform accurately.

The Knowledge Base: Vector Embeddings and Branch Scoping

The knowledge base is not a simple FAQ list. Each article is processed into vector embeddings that enable semantic search — so when a guest asks 'Do you have veggie options at dinner?' the system matches it to the article about dietary accommodations, even though the guest didn't use the word 'dietary' or 'accommodations'. This semantic matching dramatically outperforms keyword search, which would fail on natural-language questions that don't use the exact terminology of the article titles.

Articles are scoped by branch (location), so a multi-location operator can maintain location-specific answers. The question 'How far is the beach?' returns a different answer for the Bali property than for the Portugal property, automatically based on the guest's booking. Shared articles (cancellation policy, company values, sustainability practices) are maintained at the organisation level and inherited by all branches, eliminating duplicate content management.

The knowledge base supports rich content — formatted text, images, embedded maps, PDF attachments — so articles can include visual directions, equipment checklists, and downloadable forms. Article analytics show which articles are most frequently retrieved, which have the highest AI confidence scores, and which generate the most follow-up questions (indicating the article may need improvement). A content gap analysis identifies questions the AI couldn't answer, highlighting topics that need new articles.

Version history and approval workflows ensure knowledge base accuracy in team environments. When a staff member updates the check-in time or adds a new activity, the change can require manager approval before going live. This prevents well-meaning but incorrect edits from propagating through AI responses before being caught.

AI-Suggested Replies in the Inbox

Inside the unified inbox, every incoming message triggers an AI analysis. The system retrieves relevant knowledge base articles, generates a suggested reply, and presents it to the agent inline — below the guest's message, with the confidence score and source citations visible. The agent can send the suggestion as-is (one click), edit it before sending, dismiss it, or flag the underlying article for review.

For new staff, this is transformative. Instead of spending weeks learning the property's policies, procedures, and local knowledge before being able to handle guest queries independently, a new team member can start responding to messages on day one with AI-generated suggestions guiding their replies. The AI acts as a real-time training tool, showing the new agent what the correct answer is, where it came from, and how it should be phrased.

Over time, the AI learns from agent interactions. When agents consistently edit a suggestion in the same way (e.g., always adding 'Please arrive 10 minutes early' to surf lesson answers), the system flags this pattern and suggests a knowledge base article update. This feedback loop continuously improves answer quality without requiring explicit training or curation.

Why Generic AI Chatbots Fall Short

Off-the-shelf AI chatbot platforms — Tidio, Drift, Chatfuel — provide conversational AI but lack the deep integration with hospitality data that makes responses genuinely useful. When a guest asks 'Can I change my booking to next week?', a generic chatbot can only say 'Please contact us to discuss changes.' Artidal's AI assistant can check the guest's booking, verify availability for the requested dates, apply cancellation and modification policies, and respond with a specific, actionable answer — or route to an agent with all the context pre-loaded.

The knowledge base architecture also distinguishes Artidal from tools like Notion or Confluence used as informal knowledge stores. Those tools require staff to search, find, read, and synthesise information before responding. Artidal's knowledge base is purpose-built for AI retrieval — structured for vector embedding, scoped by branch, versioned for accuracy, and measured for effectiveness. It's not documentation for staff to read; it's a data source for AI to query on behalf of both guests and staff.

For operators who've tried and abandoned chatbots, the confidence-score approach addresses the core failure mode: embarrassing wrong answers. By setting conservative auto-send thresholds initially and reviewing AI performance through the analytics dashboard, operators maintain full control while gradually offloading routine queries to automation.