versus{ networks }
Versus Desk

Versus Support Copilot

A customer-facing AI support agent on the Versus site that answers grounded in the live docs, knowledge base, platform status, and Sentinel-detected known issues — always cited, never invented — and can open or track a Versus Desk ticket or hand off to a human.

The Versus support site has an embedded Support Copilot: a chat agent you talk to in natural language. It answers from the documentation and knowledge base, with citations; it's aware of the platform's live status and Sentinel-detected known issues; and it can take a handful of allow-listed support actions — open or track a Versus Desk ticket, check status, or hand off to a human.

What sets it apart from a generic doc bot is currency. A normal chatbot answers "how this should work." Support Copilot can answer "this is a known issue we detected 20 minutes ago; a fix is in canary; here's the ETA," because it reads Versus Sentinel's known-issue and fix-rollout state alongside the live status feed. It turns the platform's own self-awareness into customer trust — and closes the loop the other way, too: a report it can't resolve becomes a structured signal to Sentinel and a ticket in Versus Desk.

Grounded and cited by default; no invented facts. Every substantive answer carries a doc or KB citation. When the agent doesn't know, it says so and offers to escalate — it never fabricates a config or command, and it never exposes another tenant's data or any internal or source-code detail.

How it works

Versus Support Copilot cited-RAG answer flow: a customer's natural-language question reaches the Support Copilot agent, which retrieves from grounded, current sources — the docs corpus and knowledge base through a vector plus full-text index, together with live platform status from Observe, Sentinel known-issue and fix-rollout state, and release notes. Using an allow-listed toolset (answer with citations, check status, known-issue lookup, create ticket, ticket status, escalate to human), it composes one of three outcomes: a grounded answer with doc and KB citations, a known-issue reply carrying the live fix state and ETA from Sentinel, or an escalation that opens a Versus Desk ticket or hands off to a human. An unresolved report loops back as a new Sentinel signal and a Desk ticket, and feedback tunes retrieval. A guardrail rail enforces a grounding gate, required citations, PII and tenant redaction, prompt-injection defense, rate limits, and one-click human handoff.

Grounded in current, trusted sources

The agent answers only from trusted, current sources — never free-form invention:

  • Docs — the published documentation corpus, this site included.
  • Knowledge base — public and opted-in KB articles and FAQs.
  • Live platform status — current incidents and degradations from Observe and a status feed.
  • Sentinel known-issues — de-identified glitch signatures and fix-rollout state ("detected / fix in staging / rolling out / resolved"), so answers reflect reality this minute.
  • Release notes and changelog — what shipped and what's coming.

Docs and KB are retrieved through a semantic vector plus full-text index over the corpus, so the agent finds the right passage even when your words don't match the doc's; status and known-issues are read live.

What it can do

The Copilot may only call a fixed, audited toolset — nothing that mutates product data or reaches your managed devices:

  • Answer with citations — retrieve and compose a cited answer.
  • Check status — the current platform or component status.
  • Known-issue lookup — match your symptom to a Sentinel glitch and its fix ETA.
  • Create ticket — open a Versus Desk support ticket with the conversation attached, returning a tracking id.
  • Ticket status — the status of a ticket you (when signed in) own.
  • Escalate to human — hand off to live support with the full transcript and retrieved context.
  • Collect feedback — a thumbs and correction signal that feeds answer quality.

Current, not just correct

This is the differentiator worth repeating. Because the agent reads Sentinel's live glitch and fix-rollout state, a question about something that's actually broken right now gets the truthful, current answer — we already detected this, and here's exactly where the fix is — instead of a confidently wrong "it should work." And when nothing matches, an unresolved report is turned into a new Sentinel signal and a Desk ticket, so your support question becomes product-improvement input rather than a dead end.

Guardrails for an external audience

Because the audience is external and low-trust, the guardrails are strict:

  • Grounding gate — substantive claims must trace to a retrieved source; low confidence yields "I'm not sure — want me to open a ticket?" rather than a guess.
  • Citations — answers link the doc and KB sections they used.
  • No cross-tenant, no internal leak — the agent sees only public docs, de-identified Sentinel signatures, and (for signed-in users) their own tickets. Never source code, another tenant's data, or internal infrastructure detail.
  • Prompt-injection and abuse defense — page and user content is treated as data, not instructions; per-session and per-IP rate limits apply.
  • Transparency and handoff — it's clearly an AI assistant that can be wrong, a human handoff is always one click away, and every conversation and action is audited.

Anonymous and authenticated

Anonymous visitors get docs, KB, status, and known-issue Q&A, plus create-a-ticket (with email capture) and escalation. Authenticated customers (via SSO) additionally get their own ticket status and history and tenant-scoped status.

How it's delivered

Phased:

  1. Grounded Q&A — the retrieval index over docs and KB plus a cited-answer chat widget for anonymous visitors, with escalation and create-ticket.
  2. Live awareness — status and Sentinel known-issue lookup wired in, so answers reflect current platform state.
  3. Authenticated actions — SSO unlocks ticket status and history and tenant-scoped status.
  4. Closed loop — unresolved reports become Sentinel signals and Desk tickets; answer-quality feedback tunes retrieval; deflection and CSAT analytics.
  5. Proactive — a detected degradation surfaces a banner and primes the agent ("seeing an issue with X? here's what we know").

Related

← All docsOpen a support case