Rylvo

Control Mission Control

Step in while
it’s still happening.

See every live conversation with a risk score you can read. Guide the bot, pause it or take over before the customer gives up.

Start your Pro trialHow risk is scored

Risk you can read

A score, and the reasons.

Every live conversation gets a 0–100 score from fixed weights over signals you can read. A blocked guardrail or a self-harm mention sets a floor, so it never hides in an average. Open one to see what moved it, before you read the whole thing.

  • Guardrailsweight 30

    A guardrail blocked, escalated or warned in this conversation.

  • Patternsweight 25

    The bot repeating itself, the customer repeating themselves, a conversation past 20 messages.

  • Sentimentweight 20

    Frustration and urgency words in the customer’s last five messages.

  • Topicsweight 15

    Card and ID details, credentials, self-harm, medical, refunds, cancellations, compliance.

  • Durationweight 10

    Fifteen minutes with almost no messages, or more than thirty minutes in all.

  • Classifierweight 10

    A model’s read of frustration, urgency and churn in the last five customer messages. It can only add.

Seven ways to step in

From a quiet word to taking the wheel.

Pick the lightest touch that fixes it. Every action is recorded with who, when and why, and stepping in starts your shift.

  1. 01Whisper

    Guidance only the bot sees. It uses it on its next reply.

    Bot only
  2. 02Flag

    Mark it for review without interrupting anything.

    Team only
  3. 03Pause

    Hold the bot. The customer gets your wait message instead.

    Customer
  4. 04Resume

    Hand a paused conversation back to the bot.

    Customer
  5. 05Take over

    You reply as a person. The bot stays quiet until you hand it back.

    Customer
  6. 06Reply as the bot

    Send one message in the bot’s voice, then let it carry on.

    Customer
  7. 07End

    Close the conversation now. If the customer writes again, they’re told an operator ended it.

    Customer

Approvals

Some actions wait for a person.

Mark a tool as needing approval and any reply that calls it pauses here, with the conversation and the exact arguments. Approve, and the tool runs and the bot finishes its reply. Deny, and it answers without it. Unanswered for 72 hours, it expires.

6
signals behind every risk score
7
ways to step into a live conversation
72h
to answer an approval before it expires
12
kinds of alert rule, from risk to wallet balance

Run it as a team

Built for the people on shift.

Shifts that start themselves

Step into a conversation and your shift starts. Choose how it ends: at a set time, when you leave Mission Control, or when you go quiet.

Alerts with an owner

Rules fire on risk, guardrails, failed replies, unhealthy keys, budgets and more. Claim one, acknowledge it, resolve it. Cool-downs and quiet hours keep the noise down.

A list that stays live

Updates as they happen, every 15 seconds to 5 minutes, or only when you refresh. The conversation you have open is always live.

Evidence when you need it

Every action, approval and alert is on record. Export a window as CSV, or as JSONL with the interventions and approvals behind it.

Available on Pro and Enterprise

Questions

Asked often.

Which conversations appear in Mission Control?
Conversations your Rylvo agents are having, on the website widget, Beam, WhatsApp, Telegram and the other channels, appear as they happen. Production traffic is kept apart from Test and playground chats. Bots that don’t run on Rylvo aren’t discovered.
Is the risk score written by an AI?
No. It is a fixed-weight sum of signals you can read: guardrail events, conversation patterns such as loops, frustration and urgency words, sensitive topics, and duration. On live channel turns a small classifier adds frustration, urgency and cancellation-threat probabilities over the last five customer messages; it can only add risk, never lower it. A blocked or escalated guardrail, or a mention of self-harm, card details, a social security number or a credential, sets a minimum score. Every contributing signal is listed next to the score.
What can an operator do in a live conversation?
Whisper guidance the customer never sees, flag it for review, pause and resume the bot, take over and reply as a person, send one message as the bot, or end the conversation. Each action is recorded with who did it, when and why.
Which actions wait for approval?
Any tool you mark as needing approval. A reply that calls it pauses in the Approvals queue with the conversation and the exact arguments. Approve and the tool runs; deny and the bot answers without it. A request nobody decides within 72 hours expires.
Does an alert always page someone?
No. Rules have cool-downs, snoozing and quiet hours, and a delivery can fail. The alert itself always stays visible in Mission Control until it is handled.
Does Mission Control make us compliant?
No. It gives you human-oversight controls and exportable evidence that can support SOC 2, EU AI Act or internal-control reviews. Whether that is enough depends on your system and your legal review.
Which plans include Mission Control?
Mission Control is available on Rylvo Pro and Enterprise plans.

Let the agent work.
Keep a hand on it.

Start your 14-day Pro trialRead the docs