The architecture behind the answer.
AnswerBnB is not one model call. A Context Engine identifies the property and reservation, a Knowledge Engine and retrieval step gather only that property's verified facts, and a Safe Answer Layer declines to answer, and escalates to your team, when the evidence does not support a reply. This page says what each part does and which parts are live today versus still being built out.
Six things it does before it replies
Each one is shown as an original AnswerBnB screen. Names, dates and facts in them are illustrative.
Nine parts, each labelled for what it is today
Green means live today. Amber means the architecture is in place and the fuller version is being built out. Nothing here is a claim about accuracy or guarantees.
Context Engine
Every conversation today is tied to a real organization and property before anything else happens — cross-property contamination is architecturally prevented, not just filtered in the UI. The fuller Context Resolution Engine (resolving reservation, guest, order, or ticket identity from channel and conversation metadata, and asking for clarification when confidence is insufficient) is designed and being built out incrementally.
Knowledge Engine
Properties support multiple knowledge sources — manual entries and crawled URLs — with real robots.txt-respecting crawling, HTML extraction, SHA-256 content deduplication, and chunk storage. Structured fact extraction with per-fact provenance, source authority ranking, and automatic conflict detection are the next layer being built on top of this foundation.
Retrieval
Retrieval is scoped to the resolved property before anything else — a hard filter, not a ranking preference — so a 100-property account never surfaces another property's knowledge.
AI Agent Engine
Today's pipeline classifies intent, retrieves grounded context, and generates a response for one role: a property's guest-communication assistant. A generalized agent model (identity, tools, permissions, channels, guardrails, reusable across property, commerce, support, and ISP roles) is architecture, not yet a shipped abstraction.
Safe Answer Layer
The current pipeline already refuses to answer when its grounded context and confidence don't support a response — it escalates instead of guessing. The fuller decision model (explicitly distinguishing Answer / Clarify / Escalate / Conflict / Action as separate outcomes, with separate source-confidence and answer-confidence scores) is the next iteration of this same layer.
Automation
The principle — an AI may only perform actions it has been explicitly authorized to take — is a design constraint from day one. A general trigger → condition → action automation/journey engine is on the roadmap, not yet built.
Human Escalation
When the AI isn't confident or a conversation calls for judgment, it escalates with the full conversation and grounding context attached, rather than guessing or going silent.
Auditability
Conversations, AI decisions, and escalations are recorded as real, queryable records — not just chat transcripts. Every API operation is scoped to the caller's organization and enforced server-side, independent of what the UI shows.
AI Provider Abstraction
The application never calls Groq (the current AI provider) directly outside one abstraction layer. Swapping or adding a provider is a change to that one layer, not a rewrite of the application.
What it does, and what we deliberately don't
What the AI is built to do
- Answer from the knowledge you provide for that property
- Read the conversation so far and the reservation from your PMS
- Cover each part of a multi-question message the knowledge supports
- Escalate with the reason when it cannot answer, and let you take over
What we deliberately don't do
- Never expose hidden chain-of-thought — decision transparency shows sources, authority, freshness, and outcome, not raw model reasoning
- Never let the AI silently pick between conflicting sources of information
- Never optimize for “always answer” over answering safely
- Never publish accuracy, uptime or performance numbers that have not been measured and verified
See the rest of the picture
Put your guest messages in good hands
Create an account, connect your PMS, and test answers on your own property before AI replies to a single guest.