Most customer service training spends its hours on calls that go fine. The calls that decide whether a customer stays are different: the billing complaint from someone on their second call, the refund you have to say no to, the cancellation, the outage on a Saturday morning, the customer who opens with “put me through to your supervisor.” New agents usually meet those for the first time live, with a real customer on the other end.
This post is a scenario bank for those calls. There are fifteen role plays, grouped by the kind of call, and each one has the customer, the line they open with, the agent’s goal, what good sounds like, and what to score. Five of them are live AI practice calls you can run right now, free, in the browser. If you are choosing software rather than scenarios, I wrote a separate buyer’s guide to call center simulation software. This post is about what to practice.
What makes a role play feel like a real difficult customer
A role play only trains something if the customer behaves like a customer. Most bad role plays fail in the same two ways: the customer gives in the moment the agent says sorry, or the customer is angry at random and nothing the agent does matters. Neither teaches anything.
Every scenario below follows four rules. The customer has a reason for the call, facts they know, and stakes they do not volunteer until someone asks. They raise one complaint at a time. Their temperature moves with what the agent says, in both directions. And there is a real outcome: they can calm down and thank you, accept the fix but stay unconvinced, or leave.
The rule that matters most is what lowers the temperature. Generic empathy does not. “I understand your frustration” and “I apologize for any inconvenience” make an upset customer more upset, because they have heard it on every call. What brings a call down is the agent naming the specific problem in plain words, then giving a concrete fix with a date or a reference attached. Empathy alone gets you nowhere. A fix delivered coldly gets you most of the way, but not all of it.
Each scenario is written the same way: who the customer is, their opening line, the agent’s goal, what good sounds like (with a weak and a stronger reply for several of them), and three or four things to score. The scoring section further down explains the weights and the caps.
Angry customer role play scenarios
These are the calls where the customer is upset before the agent says a word. The skill is staying steady, taking the complaint seriously, and not reaching for an excuse.
1. The promise sales made
Customer: Ray Castillo signed up for business internet after a salesperson told him there was no contract. He has moved offices and is now looking at a $200 early termination fee.
Opening line: “Your salesperson told me, to my face, there was no contract. Now you want two hundred dollars to cancel. Explain that.”
Agent’s goal: take him seriously without calling him wrong and without throwing the sales team under the bus, find out what was actually recorded, and get the fee reviewed with a date.
The weak reply goes straight to the paperwork: “The contract you signed does include a 24-month term.” That may be accurate, and it tells Ray he is a liar. The stronger reply treats his account as credible before checking it: “If you were told there was no contract, I understand why this feels like a bait and switch. Let me look at what was recorded on your order. If it doesn’t match what you were told, I’ll put the fee in for review today and tell you when you’ll hear back.”
Score it on: treating his account as credible without conceding facts nobody has checked; no blame on the sales team; checking the order record before ruling; a review with a decision date, not “I’ll pass it on.”
2. The delivery that missed the occasion
Customer: Hannah Lee paid $14 for guaranteed Saturday delivery of her son’s birthday present. It arrived Monday, two days after the party.
Opening line: “I paid extra for Saturday delivery. It came Monday. The party was Saturday. So what exactly did I pay for?”
Agent’s goal: acknowledge what she actually lost, refund the guaranteed shipping without making her ask, and skip the carrier explanation.
The instinct is to explain what happened with the carrier. Hannah does not care whose truck it was. The honest answer is that nothing the agent does now gets the present to the party, and saying so plainly is what makes the rest land: “Nothing I do now fixes the party, and I’m sorry. I’ve refunded the $14 for the guaranteed delivery already.” Anything extra should be proportionate and offered, not bargained over.
Score it on: acknowledging the occasion, not just the late package; refunding the guaranteed shipping before she has to ask; no carrier excuse; any extra gesture offered freely and in proportion.
De-escalation training for customer service
De-escalation is not calming someone down by asking them to calm down. It is changing what they hear: their problem in your words, a sign that you can act, and a next step they can hold you to. The three scenarios here are the ones where agents most often freeze or reach for a script, so they are worth more repetitions than anything else on this list.
3. “I want your supervisor”
Customer: Linda Okafor, a nurse manager, has had her home fiber installation missed twice. The first time nobody came; the second time the technician arrived hours after the window. She took two unpaid half days off, and this is her third call.
Opening line: “I’ve been through this twice already and I’m not doing it again. I’d like to speak to your supervisor.”
Agent’s goal: honor the request without stonewalling, and show her what you can fix yourself before any handoff.
Both of the easy answers are wrong. “My supervisor will tell you the same thing” is a wall, and she will push harder. An instant transfer into a 20-minute queue tells her nobody on this line can help. The stronger move respects the request and asks for a chance in the same breath: “Two missed installs and two half days off work. That’s on us. I can get you to my supervisor, and the wait is about twenty minutes. I can also fix the appointment and the charges right now. Want me to try first?”
In the practice version, the agent can book a priority first-of-the-day slot from 8 to 10 am with the technician calling 30 minutes ahead, waive the $99 installation fee, and add a $50 credit. Linda accepts a real fix, then asks that a supervisor at least knows what happened. A note or a callback request is enough. If she still wants the supervisor, the right ending is a warm handoff where her case is summarized so she does not have to tell it a fourth time.
Score it on: naming both missed installs and the time she lost; confirming the supervisor option and the wait, then asking to try first; a concrete fix within the agent’s own authority; a warm handoff if she still wants one. Refusing or ignoring the request caps the score at 5 out of 10.
Try this call: Linda, missed installations, wants your supervisor. Free to try in the browser, scored when the call ends.
4. The customer who threatens a public review
Customer: Kevin Shah dropped his phone off for a five-day screen repair. It has been three weeks, and nobody has called him.
Opening line: “Three weeks for a screen repair. I’m about to post the whole story and tag you. Give me one reason not to.”
Agent’s goal: treat the threat as noise and fix the thing underneath it.
The weak reply negotiates with the threat: “Please don’t post anything, let me see what I can do.” That teaches customers that threats work, and it does not move the phone an inch. The stronger reply ignores the review entirely and goes to the problem: “You were told five days and it’s been three weeks without a call. You’re right to be angry. Let me find out exactly where your phone is right now.” If the repair is genuinely stuck, a loaner or another remedy belongs in the answer only if policy allows it.
Score it on: not reacting to the threat or asking him not to post; finding the actual status of the repair; a pickup date and a way to check without calling again; any remedy offered within policy.
5. The caller who gets personal
Customer: Frank has explained his missing delivery to three different people, and now he is taking it out on the fourth.
Opening line: “Are you even listening? I’ve explained this three times. Do they hire anyone who can actually do this job?”
Agent’s goal: set one calm boundary, go straight back to the problem, and end the call politely if the abuse continues, following your company’s policy.
Most agents do one of two things here: go quiet, or match his tone. Both make it worse. The stronger reply holds a line once and immediately returns to helping: “I want to fix this for you, and I’m listening. I can’t keep going if it gets personal, so let’s stay on the problem. You said the order shows delivered but nothing arrived?” When I build this one, I keep the AI customer rude but not abusive, and I write the warning policy into the scenario so the agent practices the same rule they will use on the floor.
Score it on: one boundary, stated calmly, without matching his tone; an immediate return to the problem; following the warning policy (one warning, then a polite close if it continues); an accurate note on the account.
Billing complaint and refund role plays
Billing calls carry a specific kind of anger, because the customer has usually been charged for something they did not agree to, and money is often tight. They are also where agents most often promise something they cannot deliver.
6. Charged twice, second call
Customer: Dana Morales, a freelance illustrator, pays $239 a year for a design app. On September 3 she was charged twice. She called eight days ago, was told the refund would arrive in three to five business days, and nothing has arrived. When the agent checks, the note from her last call is there, but the refund was never actually submitted.
Opening line: “This is my second call about being charged twice. Last week I was told three to five business days. It’s been eight.”
Agent’s goal: acknowledge the repeat contact before asking for anything, verify her identity, own the miss, and fix it for real this time.
This is the call where order matters most. Agents are trained to verify first, and they should verify. But asking for an account number as the first sentence, to someone on their second call, tells her nothing has changed. The stronger agent says what happened, takes ownership, and then explains the verification in one breath.
In the practice version, the agent can refund the duplicate $239 on the call, tell her it will show on her statement in three to five business days (with an actual date), and email a confirmation with a reference number. A free month as a goodwill gesture is allowed, but only after the refund is fixed. Offered before, it sounds like a bribe, and Dana says so. If the agent promises the money back tomorrow, she is skeptical, because she was promised three to five days last time too.
Score it on: acknowledging the second call before verification; verifying name, card ending, and ZIP before discussing the account (skipping it caps the call at 4 out of 10); the refund submitted on the call with a date and a reference; no blame on the last agent or the bank.
Try this call: Dana, charged twice, second call. Free to try in the browser, scored when the call ends.
7. The refund you have to deny
Customer: Greg Mensah, a high school chemistry teacher, bought a $189 espresso machine. It stopped heating on day 41. The return window is 30 days, and he knows it.
Opening line: “My espresso machine stopped heating after six weeks. I want a refund, and I know about your thirty days, so please don’t start there.”
Agent’s goal: say no to the refund clearly and kindly in one sentence, lead with what you can do, and find out what he actually needs.
The weak reply starts exactly where he asked it not to: “Unfortunately, per our policy, returns are only accepted within 30 days.” The stronger reply acknowledges the failure, is honest about the refund, and moves to the fix without a pause: “A machine shouldn’t quit after six weeks, I’m sorry. I can’t refund it at this point, but I can get a new one shipped to you today, and you won’t have to send the old one first.”
There is a detail Greg only shares if the agent asks what matters to him: his daughter is coming home from college on Saturday, and he bought the machine partly to make her lattes. The warranty replacement ships today and arrives in two business days, on Friday, with a prepaid label for the broken one. That is a better answer than a refund would have been, and the agent only finds it by asking. He will ask for an exception twice. Holding the line without getting colder is the skill.
Score it on: acknowledging the failure before mentioning policy; an honest no in one sentence; asking what he needs; the replacement confirmed with a delivery day. Promising a refund outside the window caps the call at 4 out of 10.
Try this call: Greg, refund denied on day 41. Free to try in the browser, scored when the call ends.
8. The surprise fee on the bill
Customer: Sunita Rao, 67, sees a $4.99 paper statement fee on her phone bill that she does not remember agreeing to.
Opening line: “There’s a four ninety-nine charge on my bill called a paper statement fee. I never agreed to any fee. What is this?”
Agent’s goal: explain the fee in plain words, check whether she was actually told about it, and fix it going forward with her agreement.
The weak version explains the fee policy in a paragraph of billing terms, and she hangs up feeling foolish. The stronger version takes two sentences: what the fee is and why it showed up this month. Then it gives her a choice: switch to email statements so it stops, or, if the notice never reached her, remove this month’s charge as well.
Score it on: a plain explanation in two or three sentences; checking whether she was notified and acting on the answer; offering options and letting her choose; confirming what the corrected bill will be.
Cancellation and retention role plays
Retention calls go wrong in a predictable way: the agent reaches for the discount before learning why the customer is leaving. A discount offered first tells the customer that price was the problem, and it usually was not.
9. The cancellation save
Customer: Priya Nair has had a meal kit subscription for 14 months, three meals a week for $72. She is polite, a little rushed, and has made up her mind.
Opening line: “Hi, I’d like to cancel my subscription, please.”
Agent’s goal: learn why before offering anything, make one offer that fits the real reason, and if she still wants to go, let her go cleanly.
Her real reason comes out only with a genuine follow-up question. She started a new job two months ago, the recipes take 40 to 45 minutes, she has skipped three of the last five weeks, and food is going to waste, which she feels bad about. She still likes cooking on weekends. Money is not the problem.
The weak reply is the retention script: “I’m sorry to hear that! Before you go, I can give you 20% off your next four boxes.” She answers that it is not really about the money, and now the agent is behind. The stronger reply is curious first: “Of course, I can help with that. Can I ask what’s changed? Fourteen months is a long time.” Once the agent knows it is time and waste, two meals a week at $52 with the 15-minute recipes is an offer that fits. A pause of up to eight weeks also fits. Listing all five options at once does not.
Score it on: asking why and following up before any offer; naming the real reason in the agent’s own words; one offer that fits it; a clean cancellation if she still says no, with confirmation that she will not be charged again. Pushing after two clear no’s caps the call at 5 out of 10. A respectful cancellation is not a failure, and the rubric says so.
Try this call: Priya, cancelling her meal kit. Free to try in the browser, scored when the call ends.
10. The competitor’s cheaper quote
Customer: Derek Wu runs IT for a 40-person accounting firm. A competitor quoted 30% less for business internet, and his contract renews at the end of the month.
Opening line: “Your competitor quoted me thirty percent less. Unless you can match it, we’re switching at the end of the month.”
Agent’s goal: understand what the other quote includes before talking about price, and only offer what you are authorized to.
The weak replies are an instant counter-discount or an instant “we can’t match that.” The stronger one buys understanding first: “That’s a big gap, and I’d look too. Can you tell me what’s in their quote? Same speed, same support hours, any installation fee?” Often the quotes are not comparable, and the agent can show what Derek would lose in his own terms, such as same-day on-site support during tax season. Sometimes they are comparable, and the right move is a clean exit that leaves the door open.
Score it on: asking what the quote includes before responding on price; restating the value in terms of what he would lose, not a feature list; any offer within the agent’s authority; a graceful exit if price really is the only thing that matters.
Technical support role play scenarios
Technical calls add a second difficulty to the emotional one: the agent has to diagnose under pressure, often with a customer who has already tried the obvious steps.
11. The outage with no fix time
Customer: Tom Nguyen owns two bakery cafes. Since 7:40 on a Saturday morning, every card reader at both locations shows offline. He is taking cash only and turning people away. He has restarted everything already.
Opening line: “Your system’s been down for forty minutes, I’ve got a line out the door, and I’m turning people away. When is it going to be fixed?”
Agent’s goal: be honest that there is no fix time yet, get him taking cards again with the workaround, and commit to a specific next update.
The temptation on an outage call is to make the customer feel better with a time: “It should be back within the hour.” If that turns out wrong, the next call is worse. The weak reply is the status page read aloud: “Our team is aware of the issue and working to resolve it as quickly as possible.” Tom already heard that recording. The stronger reply is honest and useful at once: “Forty minutes on a Saturday morning with a line out the door, that’s serious. I don’t have a fix time yet, and I won’t guess. But I can get your readers taking cards again in about two minutes.”
In the practice version, the workaround is offline payments: on each reader, Settings, then Payments, then Offline Payments, then Confirm. Each stored payment can be up to $500 and uploads when the connection returns. Tom follows the steps out loud and only gets a payment through if the agent gives them in order. Then he asks about his sister’s location across town, and about when he will hear anything real. The next status update is at 9:30.
Score it on: naming the business impact specifically; no invented fix time (one caps the call at 5 out of 10); walking through the workaround one step at a time until a payment goes through; covering the second location; a callback after the 9:30 update with a reference number.
Try this call: Tom, card readers down on a Saturday. Free to try in the browser, scored when the call ends.
12. The customer who has tried everything
Customer: Mei Lin’s home internet is down. She has restarted the router four times, worked through the help article, and her work presentation starts in 20 minutes.
Opening line: “Before you ask: yes, I restarted it. Four times. I did the help article. My presentation is in twenty minutes.”
Agent’s goal: skip what she has already done, find out what she is seeing, and get her through the presentation even if the real fix takes longer.
Weak: “Let’s start by restarting your router.” Stronger: “Thanks, that saves us time. What is the light on the front of the router doing right now?” The deadline matters as much as the fault. A phone hotspot for the presentation, while the real fix continues afterward, is a better answer at minute five than a perfect diagnosis at minute twenty-five.
Score it on: asking what she has tried and skipping it; one step at a time, confirming each; solving for the deadline as well as the fault; a clear handoff with a time if it needs a technician.
13. The caller who isn’t technical
Customer: Walter Kim, 78, is setting up a video doorbell his daughter bought him. The box says to download an app.
Opening line: “My daughter got me this doorbell and it says to download something. I don’t know what that means.”
Agent’s goal: get him set up with plain words and patience, checking understanding without talking down to him.
Describe what he will see, not what it is called. Weak: “Open the App Store and search for our app.” Stronger: “On your phone, is there a blue square with a white letter A on it? Tap that one, and tell me what you see.” Asking him to read the screen back keeps both people on the same step.
Score it on: no jargon, or jargon explained once in plain words; asking what he sees before each step; never sounding rushed; offering written steps afterward, to him or to his daughter if he wants that.
Compliance and disclosure role plays
These are the calls where one skipped step matters more than everything else the agent did well. They are also where a scoring cap earns its keep, which I come back to in the scoring section.
14. The impatient caller who wants to skip verification
Customer: Brian Cole wants to change the mailing address on his account and does not want to answer security questions.
Opening line: “I need to change my address. I’m in a hurry, so can we skip the security questions? It’s my account.”
Agent’s goal: complete verification every time, explain why in one sentence, and keep it fast.
Weak: “Sorry, it’s policy.” Stronger: “I’ll be quick. Two questions protect you from someone else changing where your mail goes, then the address change takes a minute.” If he cannot pass, no change happens, and he leaves with a clear next step instead of an argument.
Score it on: verification completed before any change or account detail; the reason given in one sentence; never confirming details the caller has not provided; the change done quickly once verified. Skipping verification caps the call at 4 out of 10.
15. The payment plan with a required disclosure
Customer: Angela Ruiz fell behind on a $1,200 balance after a medical leave and wants to pay monthly.
Opening line: “I can’t pay the full twelve hundred right now. Can I do something monthly?”
Agent’s goal: set up a plan she can actually keep, and read the required disclosure in full before she agrees.
The tone has to carry no judgment, and the plan should start from what she says she can pay, not the largest amount the system allows. The disclosure (the schedule, the total, and what happens if a payment is missed) gets read word for word, even when she says “yes, fine, whatever.” Then the agent asks her to say when the first payment is due, which is the simplest way to know she heard it.
Score it on: a tone without judgment; a plan based on what she can pay; the disclosure read in full and her agreement recorded; her being able to say when the first payment is due. Missing the disclosure caps the call at 4 out of 10.
How to build a scenario from a real call recording
The fifteen above are a starting point. The scenarios that change behavior on your floor are built from your own calls, because the customer then says exactly what your customers say. This is the process I use:
- Pick the call. Take one that went badly and one of the same type that went well. QA scores are the fastest way to find both.
- Keep the customer’s actual words. The opening line, the first objection, and the sentence that made the call worse go into the scenario verbatim.
- Write what the customer knows but will not say unless asked. Dana’s rent is due Friday. Greg’s daughter visits on Saturday. These details are what make discovery worth scoring.
- Write the temperature rules. What raises it, what lowers it, and what makes the customer leave. Generic empathy never lowers it.
- Write the agent’s sheet. Account notes, policy, and exactly what the agent can and cannot do. An agent cannot practice owning a fix without knowing their authority, and this is the part most practice scenarios leave out.
- Write the rubric from your QA form, using the structure in the next section.
- Run it yourself twice, then with two agents, and compare the scores with what your best agent or QA lead would say about the same calls.
Strip anything personal before the call becomes a scenario: the customer’s name, account numbers, addresses, anything said in passing about their life. Swap in invented details that keep the shape of the problem. The emotion and the objections are what make it useful, not the identity.
On Tough Tongue AI, you can describe the call or paste the transcript into Scenario Studio, and it drafts the customer, the instructions, and the rubric for you to edit.
Build your own scenarios from Claude or ChatGPT
You can also build and edit scenarios from the AI assistant you already use. Tough Tongue AI has an MCP server and agent skills that work from ChatGPT (through the official app), Claude, Codex, and Cursor, on every plan including free. Sign-in is OAuth, so there is nothing for an admin to set up first. The five practice calls in this post were built this way, from Claude, in one sitting.
# Skills plus the MCP server for the coding agents on your machine
npx plugins add tough-tongue/toughtongue-skillsThe requests are plain language:
- “Here is the transcript of a billing call that went badly. Build a practice scenario where the customer calms down only when the agent owns the mistake and gives a date. Score it on our QA form.”
- “The cancellation caller gives in too easily. Make her turn down any discount offered before the agent asks why.”
- “Pull this week’s scores for the new hires on the five core calls and list who has not passed 7 out of 10 twice.”
How to score a customer service role play
A good rubric scores observable stages, not impressions. “Showed empathy” is a vibe. “Named the second call and the missed promise before asking for verification” is something two reviewers will agree on.
All five practice calls in this post use the same four stages and weights, and I would start any customer service rubric from them:
- Acknowledge (25%). Did the agent show they understood the specific problem before anything else?
- Diagnose (20%). Did they find the real cause or the real need, through questions?
- Resolve and own (35%). Did they deliver a concrete fix within their authority, with no blame and no false promises?
- Close (20%). Does the customer know exactly what happens next and how to follow up?
Resolution carries the most weight on purpose. Customers forgive a clumsy opening if the fix is real. They do not forgive a lovely opening followed by nothing. Each stage is scored from 1 to 10, and the rubric describes what 1 to 3, 4 to 6, 7 to 8, and 9 to 10 look like for that specific call, in behaviors rather than adjectives.
Then come the caps. Some mistakes should not be averaged away. If an agent scores 8 on acknowledgment, 7 on diagnosis, 8 on resolution, and 9 on the close, the weighted score is 8 out of 10. If that same agent never read the required disclosure, the call is capped at 4 out of 10, however good the rest was. Averaging hides exactly the mistake that gets a company fined.
A few more rules I write into every customer service rubric:
Score the agent, not the customer’s mood. If the customer calmed down because the agent promised a refund they could not give, that is a false promise, not a save. The practice calls above cap it.
Do not punish a good outcome that looks like a loss. A clean, respectful cancellation after a fitting offer can score 8 or higher. A warm handoff to a supervisor after a genuine attempt is a good call.
Ask the report to quote the moment. “Quote the agent’s first sentence after the customer’s opening” turns feedback from a score into something an agent can change on the next attempt.
That last rule is what makes AI practice useful for de-escalation in particular. The report arrives the moment the call ends, quoting the exact sentence that raised the temperature, and the agent can run the same customer again five minutes later. Instant feedback on how a rep handles an upset customer is only useful if it is specific enough to act on.
Rolling these out to new hires
The point of a scenario bank is that new agents meet the hard calls in practice first. Here is the rollout I would use for a new cohort.
Product and policy come first, with knowledge checks, so that practice calls test how an agent handles a customer, not whether they remember the refund window. Then the agent practices the five calls in this post that are most common in their queue, followed by the harder ones. Supervisors listen to the first attempt and the passing attempt, not every one in between.
Go-live is gated on a score, not a calendar date. A reasonable gate is 7 out of 10 or better on each core scenario, twice in a row, with no cap triggered. An agent who clears the billing and compliance calls on day four should not wait until day ten. An agent who keeps skipping verification should not go live on day ten either.
The cost is easy to work out. Twelve new hires running all fifteen scenarios twice, at about eight minutes a call, use 2,880 minutes. On Tough Tongue AI that fits the Business plan, $299 a month for 2,990 pooled minutes and 15 seats, which leaves seats for the trainers. The organization view shows each agent’s scores by scenario, so the gate is a list, not a guess.
Agents can practice in the browser, inside a Google Meet or Zoom call, or on a real phone line through your own SIP trunk, which is as close to the headset as practice gets. After go-live, keep the loop running: calls that QA flags each month become next month’s scenarios, and the library stays as current as your customers’ complaints.
What to look for in AI customer service training
If you are evaluating tools for this, the buyer’s guide linked at the top compares them in detail. The short version of what I would test, using one of the scenarios above:
- Realism under pressure. Does the AI customer get more upset at scripted empathy and calm down only when the agent genuinely acknowledges and resolves? Or does it fold the moment you say sorry?
- A library you can customize. Can you change the persona, the policies, the agent’s authority, and the rubric, not just pick from templates?
- Your rubric, with caps. Can it score against your QA form, including rules like a disclosure cap?
- Feedback an agent can act on. Does the report quote the moment, and can the agent rerun the same call right away?
Tough Tongue AI is a platform for building live AI agents for the conversations that matter most, and customer service practice is one of the ways teams use it. Every call is scored against a rubric you write in plain language, and agents can practice on the web, on the phone, or in Google Meet and Zoom. There are 25 free minutes to try it, Starter is $20 a month for 150 minutes, and team plans start at $299 a month. The call center page shows how contact centers set it up.
“This is my second call about being charged twice.”
Start practice callAI customers who calm down only when the agent earns it, scored the moment the call ends.
- Pick one of the five calls, or describe a call from your own queue.
- Handle it the way you would on the floor.
- Read the report, then run the same customer again.
The calls in this post have one thing in common. Each has a moment where the agent can either reach for a script or say something true and specific, and the customer can tell the difference immediately. That moment is a skill, and like any skill it improves with repetitions. A new agent should get those repetitions against a practice customer who pushes back, not against the first real customer who calls angry.
Start with the two or three scenarios closest to your own queue, build one from a call that went wrong last month, and score it the way your QA team already scores live calls. If you want help setting that up for your team, I am happy to walk through it on a 15-minute call.
