Skip to main content
Support quality & coaching

De-escalation call scorecard for upset customers

When a call with an upset customer is scored, a line like "showed empathy" gets two different answers from two QA specialists, and a transcript does not show tone of voice at all. This article breaks de-escalation into four observable stages, gives an illustrative six-criterion form, separates "not applicable" from "unable to assess", and covers the role of an authority map, an illustrative escalation matrix, typical mistakes and how to discuss the result with the agent step by step.

September 29, 20268 min read

Short answer

You cannot evaluate a call with an upset customer by asking "was the agent calm?" — the question is subjective and cannot be checked reliably from the call's text. To measure de-escalation, break it into observable behaviours: did the agent acknowledge the problem, clarify what happened, state clearly what they could do, and hand the issue over properly when it was beyond their authority?

These four behaviours can be checked the same way on every call. They measure what the agent did regardless of the customer's mood, which makes them fair: a customer may still be unhappy at the end of the call even though the agent did everything right.

Why an "empathy" criterion fails

Many evaluation forms include a line such as "the agent showed empathy". The problem is not what the line means but how it is checked. Two QA specialists review the same call and answer it differently, because each understands empathy through their own experience.

"Empathy" is also often tied to tone and intonation. A system working from the transcript cannot measure tone. Google's Quality AI guidance says so plainly: transcript-based evaluation does not perform acoustic analysis, so questions about tone of voice do not belong there.

The fix is to turn the criterion into behaviour visible in words. Instead of "showed empathy": "the agent restated the customer's problem in their own words and acknowledged the concern".

The four stages of de-escalation

  1. AcknowledgeThe agent shows they heard the complaint: "I understand — the order is three days late and you haven't heard anything about it." They do not argue or immediately blame someone else.
  2. ClarifyThey ask questions that establish what happened: order number, when, what the customer was told. They isolate what the customer actually wants — a refund, a new date or simply an explanation.
  3. Explain what is possibleThey say clearly what they can and cannot do, without unconfirmed promises: not "it will definitely arrive tomorrow" but "the order goes to the courier tomorrow and we'll text you the delivery time".
  4. Hand off or closeIf it is beyond their authority, they hand it to the person responsible and tell the customer who will contact them and when. If it is within it, they wrap up the next step.

Illustrative example: a de-escalation form

The form below is an illustrative example. Answer options for each criterion are yes, partial, no and not applicable. "Not applicable" means there was no complaint in the call, or that stage was not needed.

  1. 1. The complaint was acknowledgedYes: the agent restated the problem in the customer's words or openly acknowledged it. No: the agent went straight to explaining the rules, or argued.
  2. 2. The real request was clarifiedYes: the agent asked or confirmed what the customer wanted. Partial: the question was asked, but the answer did not shape the solution.
  3. 3. Blame was not shiftedNo: phrases like "that's the courier's fault" or "I don't know, it's another department" were said without offering a solution.
  4. 4. What is possible was explained preciselyYes: what will be done, when and by whom was stated. No: a vague or unconfirmed promise was made.
  5. 5. The handoff was done properlyNot applicable: within the agent's authority. Yes: the customer was told who would contact them and when.
  6. 6. A closing confirmation was obtainedYes: the agent briefly repeated what was agreed and asked whether the customer had any other questions.

"Not applicable" versus "unable to assess"

A de-escalation form only fully applies to calls where there was a complaint. That is why two statuses must be kept apart:

  • "Not applicable" — the customer was not upset, no handoff was needed. The criterion is left out of the score and does not count against the agent
  • "Unable to assess" — the recording is incomplete, the audio dropped, the speakers were mixed up. This is also left out of the score, but reported separately
  • "No" — the behaviour was needed and did not happen. Only this status lowers the score

Without the distinction, an agent with no upset customers gets an artificially high or artificially low score, depending on which rule was chosen.

A form is incomplete without an authority map

To check "the agent said what they could do", you have to write down what the agent can actually do. Otherwise the same answer will look right to one reviewer and wrong to another.

  • Which compensation the agent can offer on their own authority
  • In which cases escalation to the next tier or a supervisor is mandatory
  • Who approves a refund, a date change or an exception
  • Which timeframes can be promised to the customer, and which cannot

That list is both an instruction for agents and the benchmark for evaluation.

Illustrative escalation matrix

To check the fifth criterion — "the handoff was done properly" — fairly, you have to write down in advance when a handoff is mandatory. The matrix below is an illustrative example; each company sets its own thresholds and owners.

  1. Within the agent's authorityInformation about a late delivery, a new date, a standard apology coupon. No handoff needed; if one was made, it was an extra step.
  2. To the shift leadA compensation request above the standard, a third complaint from the same customer, a repeat problem the agent cannot solve.
  3. To the owning departmentA refund, a change to contract terms, investigating a technical fault — the agent tells the customer which team and how long a reply will take.
  4. Urgent human reviewA legal claim, a safety risk, threats or abuse. These need a written company rule, and the evaluation must not rest on an automatic result.

Without a matrix, agents either hand off everything or nothing. Either way, line five of the de-escalation form turns into an argument.

Typical mistakes

  • Reading out the rule without hearing the complaint: "According to our policy..."
  • Confusing an apology with a promise: "we're sorry, this will never happen again"
  • Trying to silence the customer: "calm down or I'll end the call" — apart from abuse, for which the company's rule should be written down
  • Saying nothing to the customer during a transfer — they tell the new agent everything again
  • Closing the call as "solved" when nothing was solved

Discussing the result with the agent

A call with an upset customer is stressful for the agent too. The feedback conversation should take that into account.

  1. Show the good firstPoint out the criteria marked yes. The agent should know what they did right in a difficult call.
  2. Pick one behaviourDo not discuss every "no" at once. Choose the one with the most impact — usually acknowledgement or precise explanation.
  3. Show the line itselfRead the sentence in the transcript together and write an alternative phrasing together.
  4. Separate the result from the customer's moodThe customer may still be unhappy at the end. If the agent's behaviour was right, that is not their mistake.

Limits

  • A transcript does not reliably show tone of voice, shouting or intonation
  • The customer's mood at the end of the call is not a measure of the agent's quality
  • A transcript does not prove that a step the agent promised was later carried out
  • Abuse, threats and safety incidents need a separate company rule and human review
  • An AI evaluation should not be the sole basis for disciplinary decisions

What Vexvon Audio Analyzer offers

  • You write the stages above into your own call standard — as text, in your own words
  • On every call each stage is judged as met, partial, missed or not applicable; a stage that did not apply does not count against the agent
  • Each judgement comes with a comment and the transcript lines behind it — ready to open in a feedback conversation
  • Each call gets recommendations and a short summary

More: Vexvon Audio Analyzer. Winning a customer back after a bad review is a separate topic, covered in negative review recovery calls.

First step

Pick 10 complaint calls from last month and have two people score them independently with the six criteria above. The criteria where their answers differ are the first places to fix in your form. This article belongs to the support quality and coaching section. To try it on your own recordings, get in touch.

Further reading on this topic: customer sentiment vs agent performance, call transfer quality, call center coaching plan.

Live demo

Ready? Let's start

See Vexvon live in a 10-minute demo.

  • A scenario built for your business
  • A live sample call
  • A tour of the platform
Get a demoorBook a meeting

Your details are used only for the demo and to get in touch.