Contact center AI: intent routing, an assistant that resolves simple requests, proactive messages, translation and quality at scale
Everything on this page obeys the same single switch as the rest of the case AI: the case AI policy (Settings · Cases · AI policy). With the policy off, nothing here runs and nothing here is written. There is no second switch.
Everything on this page obeys the same single switch as the rest of the case AI: the case AI policy (Settings · Cases · AI policy). With the policy off, nothing here runs and nothing here is written. There is no second switch.
Routing by intent The conversational assistant already lets you register intents (a name, some examples, a few keywords). Until now the routing rules could not see them: a rule could match channel, queue, keyword, language, region, product, account tier, SLA target and severity, but not "the customer wants to cancel". Now they can. In an assignment rule, pick one or more intents and, if you want, a minimum confidence.
Two things worth knowing before you write the rule. The intent is worked out from the subject, the description and the last thing the customer wrote, so it can change in the middle of a conversation. And a case whose intent we do not know, or whose confidence is below the minimum, does NOT match the rule; it falls through to the next rule, exactly like the other conditions of the routing engine. Sending a case to the retention team on a guess is worse than leaving it in the general queue.
The intent panel shows the volume per intent and how long each one takes to resolve. That is information for whoever writes the rule, not an automatic decision: choosing an owner by historical average would need a sample per person that most teams do not have.
An assistant that resolves simple requests end to end The conversational bot can now do six things in the CRM instead of only talking: check the status of a case, list the customer's cases, check the support coverage, search the service catalog, reopen a case and schedule a callback. Each one is configured on its own: off, ask for confirmation (an agent confirms in the case), or do it automatically. Reading tasks are automatic out of the box; writing tasks ask for confirmation.
The rule that matters is identity. The customer typing an email address in a chat window proves nothing: if that were enough, any visitor could read anybody's case. A request only reaches a case when the person was identified by the customer portal (a login link), by a channel that proves possession (the WhatsApp number, the email address already linked to the contact), or by an agent acting with their own permissions. Otherwise the assistant explains that it needs the portal login.
Everything the assistant does, and everything it refuses to do and why, is recorded in a trail that cannot be edited. Writing counts against the same daily AI action cap as the rest of the product; when the cap is reached the assistant does not go quiet, it goes back to proposing.
Messages the assistant starts, that a person approves Two situations produce a message the customer did not ask for: the resolution deadline is at risk or has passed, or the customer wrote and has had no news from us for more than a day. A third one is closing: after a case is resolved, the assistant can write the recap the customer reads.
All three land in the same approval queue on the case. Nothing is sent until a person approves it, and there is no setting that changes that. The text is editable before sending: approving something you cannot correct is either blind faith or a button nobody uses. Sending goes out through the channel the case came from when there is one (respecting the 24-hour window and the approved template), and by email when there is not.
Translation inside the conversation When the customer writes in another language, you can translate their message into yours, and your draft into theirs, without leaving the case and without pasting the customer's text into an outside tool. The translation is always shown marked as machine translation and never replaces the original: what the customer wrote is what counts, and it is the original a dispute depends on.
The customer language comes from the contact record when it is filled in, otherwise from what we detect in the conversation, otherwise the screen asks you. Text already in the target language is not translated.
Quality at scale The review of a single reply already existed. What did not exist was quality the way a contact center means it: your own rubric, a sample per period, and a score per agent.
Create a rubric with the criteria your operation promises and a weight for each; a technical support team weighs "grounded in facts", a collections team weighs "compliance". Set a sample size and, if you want, a weekly or monthly cadence. Each round picks the cases deterministically and spreads them across agents, so a run of the same period twice reviews the same cases and does not pay for the model again.
The screen always shows how many cases were reviewed out of how many were resolved. A score from a sample presented without the size of the sample is a polite way of lying. And these scores only start a conversation: nothing here changes routing, targets or pay.
Customer tone over time The case AI already measured the tone of a conversation, but it overwrote the number every time it re-read the case: the product could say how a customer is, never that they are getting worse. Now every measurement is kept, so the panel can show the curve per agent and per account, and the case owner is warned when the tone of a conversation drops sharply or stays consistently negative. It costs no extra AI: it reads the analysis that already exists.
With fewer than two measurements the answer is "not enough data", never "stable". They are different answers.