Concierge — Accountable AI Support: answers from the knowledge base, escalates when unsure, logs everything
Law, medical, and accounting practices don't want a chatbot that will answer anything. They want an AI front desk that answers from approved content, hands off when it can't afford to be wrong, and leaves a full audit trail.
NexAgent's own embeddable AI support product for regulated businesses (law, medical, accounting): it answers only from a knowledge base the tenant controls, auto-escalates uncertain questions to a human — with live takeover from a Telegram ops desk — and logs every conversation. One script tag embeds it in any website.
A generic chatbot is a liability in a regulated industry: it can invent legal advice, offer medical judgments, or promise fees that don't exist. Clinics and law firms need the opposite — answers must come from content we approved, anything uncertain must go to a person, and every word must be traceable after the fact. Concierge was designed around exactly those three constraints: answer from the knowledge base, escalate when unsure, log everything. It isn't a smarter chat box — it's an accountable AI front desk.
The challenge
Website inquiries in regulated industries carry real business value — prospects, bookings, quotes — but hallucination risk keeps these businesses away from generic AI support: one piece of wrong professional advice can mean complaints or legal exposure. Human-only support can't cover 24/7, so evening and weekend inquiries are simply lost.
Our approach
Put both 'whether to answer' and 'what to answer' inside a system of rules. Responses are generated only from the knowledge base each tenant maintains in the admin panel — content that isn't there doesn't get invented. When confidence is low, the conversation auto-escalates: the on-call person receives full context in a Telegram ops desk and can take over the live session with one tap. Every exchange, escalation, and takeover is logged. Deployment is deliberately minimal: a single script tag drops the isolated chat widget into any website without touching host-page styles.
Results
The product runs in production on a multi-tenant architecture — each business maintains its own knowledge base and escalation rules. All tenant conversations are fully auditable; the AI absorbs after-hours inquiries and uncertain questions get human follow-up, so website leads stop leaking. The same product powers the support component in our CPA firm case study.
Three product red lines
Concierge's differentiation isn't the model — it's the constraints:
- Knowledge-base-only answers: responses are generated from tenant-approved content; if it's not in the base, the bot says so.
- Escalate when unsure: low confidence triggers a human handoff — never a forced answer.
- Everything logged: every exchange, escalation, and takeover is recorded and available for compliance review.
The Telegram ops desk: live human takeover
Escalation isn't an email nobody reads — it's a push to the on-call person's Telegram:
- The operator instantly sees the conversation context and the visitor's question.
- One tap takes over the same session, replying as a human in real time.
- Takeover and release are both logged, keeping the AI/human responsibility boundary explicit.
'Hand it to a human when the AI can't answer' only works if a human can actually take over in seconds. The ops desk exists for exactly that.
One-line embed, multi-tenant management
For the customer, integration cost is near zero: paste one script tag and the chat widget renders in isolated styles — it neither pollutes nor gets polluted by the host page. On our side, a multi-tenant admin manages every customer's knowledge base, escalation rules, and on-call roster, with a single published port facing the internet to keep the ops surface minimal.
Technical breakdown
- Conversation engine: Python FastAPI with retrieval-augmented generation scoped to the tenant knowledge base.
- Admin: multi-tenant CMS where tenants self-manage knowledge bases, escalation rules, and on-call staff.
- Embed: vanilla-JS chat widget in isolated styles (Shadow DOM), installed with one script tag on any site.
- Ops channel: Telegram bot pushes escalation tickets and supports live takeover and release.
- Deployment: five-container Docker Compose stack, one published port, automated nightly backups.
Reusable lessons
- Regulated-industry customers don't buy 'the AI is smart' — they buy 'the AI has boundaries.' The red lines are the selling point.
- Escalation speed decides the product: the ops desk must enable takeover in seconds, or 'transfer to human' is theater.
- Embedded widgets need hard style isolation, or every customer website becomes a bespoke compatibility incident.
Want results like these?
Tell us about your business and we'll scope a proof of concept.