Grounding policy
Answers come from your knowledge base or don't come at all. Retrieval below the confidence threshold triggers a refusal and a ticket offer. Every answer carries citations.
Governance
Grounded and guard-railed, by design.
Most AI features ship with guardrails as a footnote. Ours are the product. Here is the whole rulebook — public, versioned, and audited monthly.
The seven guardrails
Confirm gates
Tickets, cost estimates, and any outbound action require an explicit user-approved summary. The concierge never fabricates ticket IDs, SLAs, prices, or client names.
Privacy at capture
A plain-language notice before any email or phone number is collected; OTP verification for updates; PII redaction on uploaded images and documents; configurable retention; exportable transcripts.
Audit trail
Every conversation logged, searchable, exportable. Managed-tier clients get monthly transcript reviews with findings and fixes.
Cost governance
Per-tenant token budgets, aggressive caching, small-model routing. Your bill is flat; our LLM costs are fenced.
Escalation discipline
Severity detection (outage language → High), SLA-aware timers, on-call paging rules you control. Humans are a first-class feature.
Transparency
The concierge always discloses it's an AI. We publish a monthly Tuning Report: intake accuracy, refusal rate, fixes shipped.
Latest Tuning Report → (first report publishes after launch)