Skip to main content

Nox Capabilities

Audience: Management · Sales · QA · Support · Engineering · Where in app: Nox chat / Voice · Plan availability: Nox plans (Pro+)

This is the honest map of what Nox can answer today, what it can do partially, and what's not built yet. Use it to set customer expectations and to know whether a thumbs-down is a bug or an unbuilt feature. Honesty about limits is part of the Nox brand — Nox is built to say "I can't do that yet, here's why" rather than fabricate.

Legend: ✅ DONE (shippable today) · 🟡 PARTIAL (data exists but a tool/aggregation gap) · 🔴 FUTURE (needs net-new infrastructure).

The object matrix​

#ObjectWhat Nox queriesStatus
1A specific personPerson + PersonProfile + UserSession + chats/calls✅
2A specific companyPersons mapped to that company🔴 (no normalized account entity)
3A cohortPersonProfile + GHL stage + behavior🟡 (behavior ✅, CRM-stage join ❌)
4A conversationChat/call transcript + ExtractedFacts✅
5A booking / outcomeReverse-lookup booking → person → journey✅ attribution; 🔴 outcomes
6A page / behavior patternAggregate sessions by page/behavior🟡
7The pipeline as a wholeGHL state × on-site × time🔴
8PredictionsPattern-match across historical persons🔴
9RecommendationsReasoning over all the above🔴
10ActionsExecute against AI agent + GHL + calendar🟡 (see below)

Version packaging​

VersionScopeState
Nox LiteObjects 1, 4, 5 — person / conversation / booking-attribution✅ shippable today
Nox ProAdds 3, 6, 7 — cohorts, page/behavior, pipeline🟡 ~70%
Nox FullAdds 8, 9, 10 — predictions, recommendations, full agentic actions🔴 not started

Note on object 10 (Actions). The KB matrix originally marked actions 🔴 ("all read-side"). Since then the action layer shipped: Nox can place or schedule outbound AI calls (knox_place_call, scheduled_at up to 30 days out), draft-and-send customer follow-ups with review (knox_draft_followup / knox_send_followup), enrich GHL contacts (knox_enrich_contact), and set return-visit watches/auto-engage (knox_visitor_watch). On Nox funnel beta tenants it can also tag contacts, move pipeline cards, trigger workflows, book meetings, and mark leads Won (knox_tag_contact / knox_move_pipeline / knox_trigger_workflow / knox_book_meeting / knox_mark_customer). So actions are now 🟡 — a defined set of approved actions works; full open-ended "just do anything in GHL/calendar" is still 🔴. See Tools and Nox / Knox Admin Assistant.

Note on object 7 (Pipeline). On funnel-beta tenants the knox_funnel tool (summary / priorities / stats / contact) plus GHL pipeline mapping and nightly reconciliation give a real pipeline read — treat object 7 as 🟡 there. Non-funnel tenants remain 🔴.

Person-centric (Nox's superpower)​

✅ DONE​

  • "Did John visit my site again after our meeting?"
  • "When did Emma last visit?" / "How many times has Raj come back this month?"
  • "What pages did John visit on his last session?" / "How long did Emma spend on the integrations page?"
  • "Walk me through everything John did last night."
  • "What questions did Sarah ask in her last conversation?"
  • "Did Sarah switch devices between sessions?"
  • "Show me John's full journey — every device, every visit, every conversation."

🟡 PARTIAL​

  • "Is John on the site right now?" — now covered by knox_is_visitor_live / knox_presence (was a gap in the original KB; treat as ✅-leaning, verify per case).
  • "Compare Sarah's first vs second visit" — data ✅, knox_visitor_profile returns a comparison field; no rich canned diff narrative.
  • "Did John watch the demo video to the end?" — viewport + idle proxies only; no explicit video-completion extractor.

🔴 FUTURE​

  • "Has anyone from ACME been on the site this week?" — no account entity.
  • "Show me everyone in my GHL pipeline who visited today" — no CRM-stage join.
  • "What was John about to ask before he left?" — speculative; no intent prediction.

Company-centric (B2B)​

Entire section gated. PersonProfile.company_name is free-text — needs an account_key (CRM company id or email domain) + a knox_account_search tool.

  • "Has anyone from BuildCo visited this week?" 🔴
  • "Which companies are showing buying-committee activity?" 🔴
  • "Walk me through ACME's full journey" 🔴

Cohort questions​

✅

  • "Of yesterday's 3 bookings, how many came from Google Ads vs organic?"
  • "Show me the conversion rate by source for the last 30 days."
  • "Show me everyone who visited last weekend." / "leads who visited multiple times today."

🟡​

  • "Everyone who watched the demo but didn't book" — via semantic search; accuracy depends on narrative quality.
  • "Which visitors viewed pricing 3+ times this month?" — needs an explicit page-frequency filter.

🔴​

  • "Of my 12 demo-scheduled opportunities, who's been back?" — needs CRM-stage join.
  • "Are SaaS founders converting differently than agency owners?" — no persona join.

Conversation-centric​

✅​

  • "Show me yesterday's longest conversation."
  • "Walk me through every conversation that ended in a booking yesterday."
  • "What budgets have leads mentioned this month?" / "Which competitors are coming up?"

🟡​

  • "What questions are people asking most this week?" — needs canned aggregation (knox_aggregate top_questions partially covers).
  • "Was the AI's response to Sarah good?" — quality_score heuristic exists but is coarse.

Booking + outcome​

✅​

  • "Of yesterday's 3 bookings, what was each person's full journey?"
  • "Which page did each booked lead spend the most time on before booking?"
  • "How many bookings were first-time vs returning?"

🔴​

  • "Of last week's 12 demos, how many showed up?" — no outcome field auto-populated.
  • "Which bookings turned into closed-won?" — needs CRM stage write-back.

Page + behavior​

✅​

  • "Where are people scrolling on the homepage?"
  • "Which CTA gets the most clicks on /pricing?" / "Average time on /enterprise?"

🟡​

  • "Why are people leaving /pricing after viewing the Scale tier?" — top_exit_pages ✅; element-level scroll-dwell needs a custom tool.

Pipeline / Predictions / Recommendations​

  • Pipeline 🔴→🟡 — on Nox funnel beta tenants, knox_funnel answers "what's in my funnel / who's hot / are we on pace" and the reconciliation queue compares behaviour stage vs mapped GHL pipeline location. Non-funnel tenants still lack the CRM-stage join (🔴).
  • Predictions (whole section) 🔴 — no model layer.
  • Recommendations 🔴→🟡 — the funnel daily brief now ships deterministic priority lists and executable offers (knox_brief_offers); free-form recommendation synthesis is still 🔴.

Agentic ("just do it")​

🟡 — the approved action set works (see object-10 note above). Still open:

  • "Slack me whenever a hot lead from my GHL pipeline visits" — lead-score notification path ✅; GHL-pipeline filter ❌.
  • Open-ended "do anything in GHL/calendar" 🔴.

Edge cases​

Existential / meta — ✅​

  • "What can you actually do?" / "How do you know all this?" / "Where does your data come from?"

Skeptical / frustrated​

  • "Just give me the answer, stop hedging" ✅
  • "I don't think that's right" 🟡 — feedback recorded; no self-correction loop.

Open-ended​

  • "What's interesting this week?" / "Surprise me." — 🟡→🔴: the insight engine now produces proactive insights for the brief, but free-form "surprise me" narrative is still limited.

Real-time / urgent​

  • "Who's on my site right now?" ✅ via knox_presence.
  • "What just happened?" — knox_action_feed answers "what happened" 🟡; no "interesting-threshold" trigger for an unprompted alert 🔴.

Citations & feedback​

Nox shows its work so admins can trust (and audit) it:

  • Citations. Knowledge-base answers cite chunks as [S1], [S2]. Clicking a citation reveals the source doc title, the matching chunk summary, its section_path, and a link to the full source (knox_knowledge_chunks). Data tools (session/visitor/CRM) are not cited — the data is the source. Hybrid answers cite only the knowledge portion.
  • Feedback. Each response can be marked 👍 / 👎 (with an optional comment), stored on knox_agent_conversations.feedback. It drives the Admin → Knox → Conversations audit view and an aggregate quality metric for super-admins. It does not retrain Nox or trigger self-correction.
  • What Nox must never claim: that it predicted an outcome, that an action was taken when it wasn't, that it remembered something it didn't, or that data exists from before the company joined.

How to answer with this matrix​

  1. ✅ → answer confidently.
  2. 🟡 → answer best-effort and disclose the caveat ("I can show approximate matches via semantic search, but I don't have a canned exact count").
  3. 🔴 → explain why and what's needed ("I'd need a CRM-stage join — not built — here's a workaround").