Governance and control
Escalation rules
The triggers that send a conversation to a person, how rules per type of question combine with a measured confidence threshold, where each escalation goes, and what the customer hears.
On this page
An agent is only as trustworthy as its judgement about when to stop. Escalation rules decide when a conversation goes to a person on your team, which person, and what the customer hears while it happens. Your team sets the rules with Telonic during implementation and can change them afterwards. Whatever else is set, a customer who asks for a person is handed to one.
How the agent decides to hand over
On every turn of a conversation, the agent checks whether any of your triggers has fired. Some triggers are about the request itself: a type of question your team always wants a person to handle, or a request outside the agent's limits. Some are about the customer: they ask for a person, they are becoming frustrated, or they show signs of being vulnerable. And some are about the agent: it is not confident enough in the answer it would give.
The last of these works in two layers. For each type of question, your team sets a rule. That rule is combined with a confidence threshold (the level of certainty below which the agent hands over rather than answers), measured for the model before deployment. Rules come from your policy, and the threshold comes from testing.
Trigger types
| Trigger | What fires it | Example | How it is set |
|---|---|---|---|
| Customer asks for a person | Any request for a person, in any wording or language | "أبي أكلم موظف" ("I want to speak to someone") | Honoured in every deployment |
| Type of question | A rule your team sets for that type of question | Any question about why a claim was declined goes to the claims team | Your team, per workflow |
| Confidence | The agent is not confident enough in a consequential answer | A coverage question that the policy wording does not clearly answer | Measured before deployment, per model |
| Frustration or distress | Rising frustration or distress, detected from the customer's tone and language | A passenger whose rebooked flight has been cancelled again | Your team sets which signals trigger escalation |
| Vulnerable customer | Signs that the customer may be vulnerable, configured during implementation | A policyholder who mentions a recent bereavement | Your team, with Telonic, during implementation |
| Named accounts or tiers | Customers you always want a person to handle | Top-tier loyalty members; a broker's key accounts | Your team, from the data in your systems |
| Outside scope or limits | A topic out of scope, or an action or value the agent is not permitted | A request to waive a cancellation fee outside policy | Your scope and limits |
| Request the agent cannot complete | A system it needs is unavailable, or the answer is not in your sources | The reservation system does not respond | Applies in every deployment |
Every escalation is logged with the trigger that fired, so your team can see why each conversation went to a person. See Audit trail and decision records.
Rules per type of question, combined with a confidence threshold
For each type of question in a workflow, your team chooses what the agent does.
| Rule | What the agent does | Example |
|---|---|---|
| Handles | Answers or completes the request within your limits | Claim status |
| Handles when confident | Answers when its confidence is above the threshold, and hands over when it is not | A coverage question answered from the policy wording |
| Prepares and hands over | Gathers the facts, then passes the conversation to a person | A refund request |
| Straight to a person | Recognises the request and routes it without attempting it | A complaint already with the regulator |
The confidence threshold is measured by testing the model against conversations from your industry before deployment, so it reflects how that model actually performs on your kind of questions. When a model changes, the threshold is measured again as part of testing the new model. See Testing before every release.
The effect is that the agent's own uncertainty is a trigger in its own right. On a question where a wrong answer would matter, a less certain answer goes to a person rather than to the customer.
When the customer asks for a person
The request is honoured every time, in any wording and any language, and the agent does not try to talk the customer out of it. It hands the conversation over with a brief, so the customer does not repeat themselves.
If nobody is available, for example outside your team's hours, the agent says so plainly, tells the customer when and how they will hear from your team, and records that as a commitment. See Commitments and promise tracking.
Where each escalation goes
Each trigger has a destination: a named person, a team, a queue, or an on-call number. The destination can differ by time of day, so the same trigger can go to the claims queue at 11am and to the on-call handler at 11pm.
The handover arrives where your team already works: your contact centre platform (the system that routes calls and messages to your team), your CRM (the system that holds your customer records), or the Telonic console (the web application your team uses). On a call, the agent transfers the call itself. Where your team is not available, the destination can be a queue for the next morning with a callback booked for the customer, configured during implementation. See Handover to your team.
What the customer hears
The customer is told, in their language, that they are being passed to a colleague, and why, in plain terms. They are told what happens next. They are not asked to repeat anything, because the person picking up has the full conversation and a brief.
Where the trigger is distress or frustration, the agent acknowledges what the customer has said before it hands over. Out of hours, it gives a clear next step rather than an open-ended wait: "Our claims team is available from 8am. I have passed them everything you have told me, and a handler will call you on this number tomorrow morning."
The technical detail
Escalation rules are defined per workflow. Each rule has a trigger, a destination for each time window, and the message the customer hears. Your team changes rules in the console, and each change is tested and approved before release. See How changes go live.
Sentiment (how the customer is feeling) is assessed continuously from the customer's tone and language during the conversation. On a call, the transfer uses SIP (the standard protocol business phone systems use to connect calls): SIP REFER (the standard instruction that passes a call on to another destination) by default, or a bridged transfer where the brief must be spoken to your colleague before the caller is connected. The call carries a conversation reference in a SIP header (a labelled field sent with the call), so your system can open the right record. See Voice: transfers, callbacks and overflow.
In practice
Rimal Air, a Gulf airline, uses the agent for booking changes and flight status on voice and WhatsApp.
- Its customer care team sets the rules: refund requests are prepared and handed over; complaints go straight to the complaints team; any passenger who mentions a medical condition or travelling with a person who needs assistance goes to the special assistance team.
- Top-tier loyalty members are a named tier. Their conversations go to the premium desk whenever they ask for a change.
- One evening, a flight to Cairo is cancelled. The agent rebooks hundreds of passengers onto options their fares permit.
- Nadia, whose rebooked flight is then delayed overnight, becomes increasingly frustrated. The agent acknowledges it, tells her it is passing her to a colleague who can help with a hotel, and hands her to the duty team with a brief.
- At 2am, a passenger asks to speak to someone about a refund. The refund desk is closed, so the agent records his request, tells him the team will call him after 8am, and books the callback.
- The next morning, the customer care manager filters the night's conversations by trigger and sees how many went to a person, and why.
What your team controls
- The rule for each type of question in each workflow.
- Which signals of frustration, distress or vulnerability trigger escalation.
- The named accounts and tiers that always go to a person.
- The destination for each trigger, by time of day.
- What the customer hears when a conversation is handed over.
Related
- Decision boundariesGovernance and control
- Handover to your teamAgents
- The controls your team holdsGovernance and control
- Sentiment and frustrationQuality and analytics
- Voice: transfers, callbacks and overflowChannels
Product names and logos are trademarks of their owners. Their mention shows systems Telonic connects to and does not imply partnership or endorsement.