How SLAtech handles training data, hallucination, citation grounding, bias monitoring, human-in-the-loop, visitor consent, misuse and the right to explanation. No waffle — every principle is paired with the concrete technical practice that enforces it.
Tenant content is never used to fine-tune or train any model SLAtech operates. RAG retrieves from tenant content at query time; the LLM sees the retrieval result as runtime context, not training signal. This is a contractual commitment in the DPA and a technical control in the ingestion pipeline.
Every bot response carries a confidence + factuality + hallucination flag visible in the admin Inbox. Bot output that scores below the configured threshold routes to a human-handoff fallback ("Let me connect you with a human") rather than guessing. Medical and Legal verticals add domain-specific safety constraints (UPL safeguard, opt-in identifier tokenisation).
We document which language model serves each pipeline (a lighter model for translation seeders, a higher-capability model for production bot answers, internal options for vertical-routing experiments). Model swaps are communicated via the changelog. Enterprise customers can request a pinned-model configuration.
Bot response payload ships QuerySource[] with { sourceUrl, snippet, score }. The widget renders the snippet as an "according to" hover-card so visitors can verify the citation before acting on the answer. Citation-rich responses also let AI scrapers (Perplexity, ChatGPT) ground their downstream citations in the actual quoted text rather than paraphrasing.
A per-vertical eval scoreboard runs quarterly across multiple platforms (SLAtech vs Intercom Fin vs Tidio Lyro vs Chatbase) — see /en/eval/. Scores are auditable; raw transcripts ship on request. Regulated-vertical responses additionally undergo LLM-as-Judge cross-check at 100% coverage.
Medical and Legal verticals route every substantive clinical / legal question to a human (Legal's UPL safeguard, Medical's "I'd rather have a clinician confirm" fallback). Sales and Hospitality verticals route low-confidence answers to human-agent live chat. The bot's role is to qualify and triage, not to replace expert judgement.
The web widget requests minimum visitor data needed for the conversation. Lead capture is opt-in. Visitor IDs are anonymous fingerprints unless the visitor explicitly identifies themselves. PII collected during lead capture is encrypted at rest and accessible only to tenant-authorised users.
SLAtech declines to operate bots designed for deceptive impersonation (a bot pretending to be human is permitted only where clearly disclosed), for regulated-advice without a licensed professional in the loop (clinical diagnosis, legal advice, financial advice), or for surveillance of visitors without consent. We reserve the right to terminate contracts in flagrant breach of this posture.
Visitors interacting with a SLAtech-powered bot can request the reasoning chain behind any answer (which source URLs grounded the response, what confidence score the LLM-as-Judge assigned). Tenant admins can audit any conversation in the admin Inbox. The right to explanation is operationalised, not just stated.
Email the founder. Public ethics statements evolve through scrutiny, not despite it.
Four self-serve tools for evaluating SLAtech (or any AI chatbot vendor) without a sales call: