There is a categorical difference between a visitor who is unhappy and a visitor who has just mentioned their lawyer. The first is an ordinary support conversation having a bad moment. The second has changed what the conversation is, and the cost of missing it is not a lower satisfaction score — it is a chargeback, a cancelled contract, or a public review written by someone who felt ignored.
The risk radar on the right rail of the agent console watches for the second kind. It scans visitor messages for structural escalation language and pins each hit to a list the agent has to deal with explicitly. The collapsed pill counts open flags and turns red the moment any high-severity one is outstanding.
Why this is not the objections list
The rail already has a panel for visitor concerns. It catches this is too expensive and it does not integrate with our stack — the ordinary friction of a sales or support conversation.
Putting manager escalations on that same list would understate them. An unaddressed objection costs a deal. An unaddressed risk flag costs a contract, a payment dispute or a public post. The two need different responses at different speeds, so they get different surfaces, and generic complaint phrasing is deliberately left to the objections panel.
The eight categories, and the five that are high
Every pattern carries a category and a severity. Five categories are high severity, reserved for cues that map directly to a supervisor-grade or legally-aware response.
- Legal — lawyers, solicitors, suing, the regulator, reporting you to a trade body.
- Chargeback — disputing the charge, calling the bank, reversing the payment.
- Escalate — asking for a supervisor, a manager, someone in charge.
- Repeat contact — the third time I have contacted you, still no resolution. High severity because a visitor who has been bounced around and is not acknowledged for it is the fastest route from annoyed to done.
- Deadline threat — an explicit if this is not fixed by Friday, with a stated consequence.
Three are medium severity. Churn, where the visitor signals exit — cancel my account, switching to a competitor. Review threats, where they signal going public. And abuse, meaning profanity or slurs aimed at the agent personally; that one is flagged less because it needs a supervisor and more so the agent can see it named on screen and pace themselves.
How a sentence becomes a flag
Messages are stripped of HTML, whitespace-normalised and split into sentences on terminal punctuation and newlines. Anything under eight characters is dropped — a slightly higher floor than the sibling panels use, because a fragment is not enough context to justify raising an alarm.
A surviving sentence becomes a flag if it matches at least one risk pattern, and it takes the worst severity across everything it matched. A sentence that hits both churn and legal is filed as legal. Text is truncated at two hundred and forty characters, and rows are deduplicated against both the current message and every flag already open on that visitor, so repeating a threat does not multiply the list.
The negation rule
Before a flag is raised the sentence is checked against a negation list, and a match rejects it. I do not want to cancel. I am not going to sue you. We will not need a manager.
This matters more than it looks. Those sentences contain the exact trigger words and mean the opposite, and they are common in precisely the conversations the radar is watching — a visitor reassuring you that they are not trying to make a scene is a normal thing to say while raising a complaint. A radar that flagged them would fire hardest on the people being most reasonable, and an agent who sees two false alarms stops trusting the panel for good.
Why nothing clears itself
Rows arrive open and there is no automatic acknowledgement and no automatic resolution. Not after a reply, not after a delay, not when the topic moves on. The agent taps Ack to say they have read it and are acting, or Resolved to say it is closed, or Skip if the extractor got it wrong. Reopen puts any of those back to open.
Nothing times out because a stale green tick on a real legal threat is worse than no panel at all. The whole value of the surface is that the state on screen reflects a decision a person made, so the only thing that can clear a flag is a person.
The flag that holds up your wrap-up
This panel is the one that reaches furthest into the wrap-up card, and it is the only one on the rail that can set the full warning state on the readiness signal.
Any open flag turns readiness to warn. It stays at warn while any high-severity flag is open, and also while one has been acknowledged but not resolved — reading a legal threat is not the same as dealing with it, so acknowledgement alone does not clear the warning. The flags also flow into the recap you copy, so the escalation is stated in whatever ticket or CRM record the chat ends up in rather than living only in the transcript.
Closing a chat with an open chargeback or legal flag is the precise failure this panel was built to prevent, which is why it is allowed to be the loudest thing on the rail.
What it is not
It is a list of regular expressions, not a judgement. It reads English structural phrasing, so a threat made politely, made in another language, or made in a way no pattern anticipated will not appear. An empty panel is not evidence that a chat is calm.
It also does not notify anyone. Nothing is emailed, no supervisor is paged, and no rule fires elsewhere in the product; the flag is a signal on one agent's screen. If your escalation process depends on someone other than the agent finding out, that has to be a real process — the radar makes the agent aware, and the agent still has to act.
State is per visitor and in memory. A transfer or acceptance drops it, so an incoming agent starts clean, which means the handover note has to carry the escalation. That is the correct place to trip over this: it is a strong argument for resolving or writing up a flag before you transfer rather than after.
What to measure
Do not track flag counts as a quality metric. It measures how bluntly your visitors phrase things, and it will move when you change your pricing page or run an ad campaign, not when your support gets better or worse.
The number worth watching is different: how many chats were closed while a high-severity flag was still unresolved. That is a process failure rather than a phrasing statistic, and it points at something you can fix — usually a missing escalation route rather than an agent who did not care. When you do find one, when a live chat turns legal covers what should have happened next.
Put it into practice
- Open a live chat and press Ctrl+Shift+X, or click the pill on the right rail, to expand the card.
- Check that your team knows a red pill means high severity, not just a raised count.
- Agree what Ack means for your team. It should mean a person is acting, not that the row has been read.
- Resolve or write up any open flag before transferring a chat, since the panel does not travel with it.
- Once a month, look for chats closed with an unresolved high-severity flag, and fix the route rather than the agent.
Risk flags are one specific reading of a conversation. The broader subject line sits on a different card: the topic badge your console puts on a chat explains how a chat is classified as an outage, a cancellation or a billing question, and where keyword matching fails.