Live supervision for every
AI conversation.
Watch every live bot conversation, score risk in real time, route the right alert to the right operator, intervene with a surgical action, resolve approvals inside an SLA, and keep an immutable audit trail — all from one control surface.
Regulation is catching up fast
EU AI Act Article 14 compliance is mandatory by August 2026. Mission Control captures the oversight evidence regulators expect.
Silent failures cost most
Risky conversations stall, escalate, or churn customers long before anyone notices. We surface them before they age.
Operators need leverage
Whisper, pause, takeover, approve — five surgical levers beat the blunt 'shut it off' button enterprises usually have.
How it works
Watch. Score. Route. Intervene. Resolve. Audit.
Six real-time layers run continuously. Every stage hands enough context to the next so operators act on signal, not noise.
The supervision pipeline
live · real-timeWatch
Every live conversation, on one screen
Score
Risk isn't a guess — it's a weighted signal
Route
Alert the right operator, not the whole room
Intervene
Five surgical actions, always with a rationale
Resolve
Time-boxed approval queues, never abandoned
Audit
Immutable trail — built for SOC 2 and the EU AI Act
Watch
Every live conversation, on one screen
Mission Control streams every active bot conversation in real time via Firestore subscriptions. Operators see message counts, duration, assigned bot, assigned operator, and current status without leaving the dashboard.
How it works
- Real-time onSnapshot streams for conversations, alerts, and approvals
- Status badges: active, paused, operator_takeover, escalated, ended
- Bot, workflow, and org filters for focused triage
Score
Risk isn't a guess — it's a weighted signal
A scoring engine continuously computes a conversation-level risk value by blending five independent signals. Each signal is surfaced to operators so they see why a conversation lit up, not just that it did.
How it works
- Guardrail (30%) — blocks, escalations, policy hits
- Pattern (25%) — loops, repetition, stalled negotiation
- Sentiment (20%) — frustration / urgency keywords
- Topic (15%) — PII, crisis, compliance-sensitive content
- Duration (10%) — extended or stalled sessions
Route
Alert the right operator, not the whole room
Configurable AlertRules decide when to escalate. Thresholds, cooldowns, and operator targets are tuneable per org. Fired alerts land in the dashboard and route via the channels you already use.
How it works
- 8 trigger types incl. risk spike, guardrail block, repeat failure
- Per-rule cooldown to prevent pager fatigue
- Operator routing by shift, role, or specialty tag
Intervene
Five surgical actions, always with a rationale
Operators act with the lightest possible tool. Whisper guides silently. Pause freezes. Takeover lets a human drive. Inject pushes a scripted message. Kill ends the session with a safe fallback. Everything that changes the bot's behaviour is rationale-logged.
How it works
- Whisper — invisible guidance to the bot, never shown to the user
- Pause — stop new messages until a human clears the session
- Takeover — operator drives the conversation directly
- Inject — scripted message sent as the bot
- Kill — end session with a safe fallback and audit entry
Resolve
Time-boxed approval queues, never abandoned
When a bot attempts a gated action, it pauses and fires an ApprovalRequest with an SLA-aware urgency. If no operator responds in time, the request auto-denies — so sessions never hang silently.
How it works
- Urgencies: immediate (15s), standard (120s), extended (900s)
- Pending queue surfaced in real time with elapsed timers
- Auto-deny on timeout — configurable per rule
- Every decision becomes a reviewable audit entry
Audit
Immutable trail — built for SOC 2 and the EU AI Act
Every intervention, approval, alert, and status change is written to an append-only OversightAudit log. The data model is purpose-built for EU AI Act Article 14 evidence and SOC 2 change-management requirements.
How it works
- Append-only: reads are open, writes go through enforced rules
- 16 audit action types across operator and system activity
- Exportable for compliance reviews and customer audits
- Ties back to the run trace for reproducibility
Risk scoring signals
Five inputs, one transparent score.
Guardrail
Blocks, policy triggers, and escalations from safety rules.
Pattern
Loop detection, repetition, stalled negotiation.
Sentiment
Frustration and urgency keywords in user messages.
Topic
PII, crisis, financial, or compliance-sensitive matter.
Duration
Extended or stalled sessions past org thresholds.
Intervention deck
The five actions operators use mid-session.
Whisper
Silent guidance to the bot mid-conversation.
rationale optionalInject
Push a scripted assistant message verbatim.
rationale requiredPause
Freeze further messages until a human clears it.
rationale optionalTakeover
Human operator drives the conversation.
rationale requiredKill
End session with a safe fallback message.
rationale requiredApproval SLAs
Time-boxed, never silently stalled.
Immediate
15sSafety-critical or financial actions.
Standard
120sRoutine gated actions needing a human check.
Extended
900sDeliberate reviews that can tolerate latency.
Approvals default to auto-deny on timeout so conversations never hang. Thresholds and fallbacks are configurable per rule.
Built for SOC 2 and the EU AI Act
Evidence, rationale, and immutability by default.
EU AI Act — Article 14
Live human oversight, ability to intervene or override, rationale capture, and audit trail — the exact evidence expected for the August 2026 deadline.
SOC 2 change management
Every intervention, approval, alert rule change, and operator shift is recorded against a user, timestamp, and conversation. Exportable for auditors.
Answers
What teams usually ask
Does Mission Control work with existing bots?
Yes. Any bot that routes through Rylvo automatically appears in the live feed. Risk scoring, alerts, and the audit log activate once Mission Control is enabled for the org.
Is every action logged?
Yes. Interventions, approvals, and alert acknowledgements are written to an append-only OversightAudit collection with 16 action types — ready to export for compliance or customer audits.
How does this map to the EU AI Act Article 14 deadline?
Article 14 demands ongoing, effective human oversight for high-risk AI. Mission Control provides the live feed, intervention controls, rationale capture, and evidence trail that article requires — built to satisfy the August 2026 deadline.
Can operators override the bot entirely?
Yes. Takeover transfers the conversation to a human operator who drives the response, while Kill ends the session with a safe fallback. Both require rationale and are logged.
Stand up oversight in a day
Put a human in the loop — without slowing the bot down.
Mission Control is available on Growth plans and up. Risk scoring, live feed, five intervention actions, approval SLAs, and the immutable audit log are all included.