How to turn lost calls into AI roleplay scenarios

You turn a lost call into a roleplay scenario in five steps: find the moment the call turned, group those moments by what the buyer said, write one buyer for each group, score the rehearsal on the same standard as a real call, and check the next week of real calls before you keep the scenario. Start with three or four scenarios, not thirty. A scenario that no real call has tested is a guess about what to teach.

Oleg KulakovCEO, Spellit8 min read

Most advice on AI roleplay starts at the other end. You pick a scenario from a catalog (the angry customer, the price haggle, the gatekeeper) and the team practices it. That works as a warm-up. It does not tell you whether the scenario is the one costing you deals this month, and nobody on the team can say why it was chosen.

The calls you lost already contain the answer. Each one has a point where the buyer said something and the conversation went somewhere you did not want it to go. Collect those points, and the question "what should we rehearse?" gets an answer you can defend on Monday morning. The rest of this article is the order in which to do it.

A lost call becomes useful only when you name the exact moment it turned

Open the recording, find the buyer's sentence after which the call changed direction, and write it down in one line with the time on the recording. That line, not the whole call, is the raw material for a scenario.

A default worth starting from is 20 lost calls from the last month. A lost call here means any call that ended without an agreed next step, or a deal that went quiet afterward. Twenty is our suggestion rather than a measured number, and it is enough to see whether the same moment repeats.

If you have no time to listen to twenty recordings, this is the part to hand to software. Spellit Actions reads every call and, for deals that went quiet, opens into the exact moment the thread was dropped, with the recording behind each claim. You can do the same thing by hand with a spreadsheet: one row per call, the buyer's sentence, the timestamp, and what the rep said next.

Write the buyer's sentence in the buyer's words, then add what happened in the thirty seconds after it. The second part matters more than it looks, because the scenario will have to reproduce the pressure of that half minute, not just the sentence.

Group the moments by what the buyer said, and stop at four groups

Sort the twenty lines by the buyer's sentence, not by the rep's mistake. You will usually end up with a handful of piles, and the largest three or four are your scenarios. The rest go on a list for later.

Grouping by the buyer works better than grouping by rep behavior because the buyer's sentence is what repeats across people, markets and products. The wording changes. The meaning stays. In our data, the objections tagged on sales calls fall into sixteen families, and the first three carry about half of the total.

Family of objectionShare of tagged objections
Price, no budget, needs installments19.9%
Needs to think, not now15.5%
Distrust, too good to be true13.6%
Not interested, fine as it is5.9%

Those shares count objections buyers raised on calls at eight sales teams in five companies, over the 90 days to October 4, 2026, grouped by hand into families. They describe what buyers say, not why specific deals were lost, and your own pile will look different. The full table, and the reason a ranking from another company travels badly, are in the objections article. What carries over is the method: the pile is built from sentences, and the sentences decide the order. The second family matters beyond this slice: the authors of The JOLT Effect report, in their own study of 2.5 million sales calls, that between 40% and 60% of deals are lost to customer indecision.

Two cautions. If one pile holds more than half of your lost calls, split it by what the buyer wanted next, otherwise you will end up with one scenario trying to do four jobs. And if a pile has only one call in it, keep it on the list and do not build for it yet. One call is an anecdote.

A scenario card describes a buyer, not a script for the rep

For each group, write a card of five lines. The AI customer plays the card, and the rep decides what to say.

Line on the cardWhat goes in it
OpeningHow the buyer starts the call, and how warm they are
The sentenceThe buyer's line from your pile, rewritten so no real person or company can be recognized
What they push onThe follow-up question or silence that came after, which is where the pressure sits
What softens themThe one thing a rep can say or ask that moves the buyer forward
Pass conditionWhat has to be true at the end, in terms you could check in a transcript

Here is one card, built to show the format. It is illustrative, not a client case.

Opening: Friendly, asked two practical questions about setup, no hurry. The sentence: "This looks good. Let me think it over and get back to you." What they push on: After the rep offers to send the details, the buyer says "great, thanks" and starts to wrap up. What softens them: Being asked what, specifically, the thinking depends on. Pass condition: The rep asks what the decision depends on, and the call ends with a dated step.

The card does not contain the line the rep should say. The point of rehearsal is to find your own words for a moment you have already failed once. Spellit Academy plays this kind of buyer by voice: one who objects, hesitates, goes quiet and asks the awkward question, as many times as it takes and without a real customer on the line. The onboarding use case shows how the same rehearsal fits into a new hire's first weeks.

A second card, shorter, for the distrust family. The sentence: "It sounds too good to be true. What's the catch?" What they push on: Every extra claim the rep adds makes the buyer more guarded. What softens them: One concrete detail about how it works, followed by a question about what would convince them. Pass condition: The rep stops adding claims and asks what proof the buyer needs.

Score the rehearsal against the same standard as a real call

A rehearsal is only worth comparing with live calls if both are judged by the same list of behaviors. Take the scoring items you already use on real calls, keep the ones that apply to the scenario, and use them to mark the practice.

This is the step that people skip, and it removes the main benefit. If the practice is graded on a made-up form, a score of 85 in training and a score of 60 on the live call cannot be set side by side, and nobody learns whether the training moved anything. If both use the same items, the gap between them is a result you can read.

The scoring items should come from the pass condition on the card, plus whatever your standard already requires for that stage of the call. A role-play rubric lists what a practice call is judged on and how many points each item carries. In Academy, the verdict after a role-play follows the standard the real calls are scored on, and for each weak area it shows the sentence that was said and a better one to use instead.

Resist adding items that exist only for the roleplay, such as "sounded confident". They cannot be checked on a live call, so they cannot be used to decide whether a scenario worked.

Keep the scenario only while real calls still show the moment

The scenario is a hypothesis: if people rehearse this moment, fewer real calls will lose it. Test that on real calls, not on practice scores.

Run each scenario for a week or two, then go back to the live calls and look for the same moment. For the card above, that means checking how many calls where the buyer said "let me think it over" ended with a dated step. Compare the share before and after, and compare it for the people who rehearsed against those who did not, if you have both groups.

In Spellit, Spellit Academy gives the rep an AI customer to practice on before the next live call. Spellit Actions gives each rep one to three tasks per week, so the rehearsal should match one of them, not add a fourth.

Three outcomes are normal. The moment still shows up in live calls: rewrite the card, because the buyer you wrote is easier than the one on the phone. The moment is gone: retire the scenario and take the next pile. The moment is gone because the deals stopped reaching that point: check before you celebrate. Practice scores and live calls can drift apart, and we looked at what the evidence does and does not support in our piece on AI roleplay and ramp time. Published practice gains deserve the same caution: in a 2025 trial at two UK medical schools, an AI voice patient raised self-rated communication skill, but the outcome was how students rated their own skill after one session.

What to do next

Pick 20 calls from the last month that ended without an agreed next step. For each, write one line: the buyer's sentence and the timestamp. Sort the lines into piles by what the buyer said, take the three largest, and write a five-line card for each one. Do that before you open any roleplay tool; with the cards on one page, you can start rehearsing on the same day.

If you want to see it on your own recordings, we will run a session on ten of your calls and show which moments repeat. Book a demo.

Key points
  • Name the moment, not the call. A lost call is a recording, and a scenario needs the one sentence where it turned.
  • Group by what the buyer said, and stop at four groups. On our data, three families of objections carry 49% of everything buyers say out loud.
  • A scenario card describes a buyer: what they open with, what they push on and what makes them soften. It is not a script for the rep.
  • Score the rehearsal against the same standard the real calls are scored on, so practice and live results can be compared.
  • Keep a scenario only while real calls still show the moment. Retire it when they stop.
See the demo now. Then run it on yours.

A live demo of our products and a real conversation about the growth problem you need to solve.

Book a demo

FAQ

How many lost calls do I need before I write scenarios?

Twenty from the last month is a default worth starting from. It is our suggestion rather than a measured threshold. It is enough to see whether the same buyer sentence repeats, and few enough to review by hand in an afternoon. If you have fewer than ten, treat every pile as a hypothesis and write the card anyway, but plan to rebuild it after the next month of calls. If you have hundreds, do not read them all. Take the most recent twenty, build the first three scenarios, then use the rest to check whether those piles are also the largest in the full set. A call that ended without a next step counts as lost for this purpose, even if the deal is technically still open.

Should a scenario use a real recording or an invented buyer?

Build it from the recording and rewrite it. Keep the order of what the buyer said, the kind of pressure and the pause, because those make the scenario feel like the real call. Change everything that could identify a person: names, company, figures, the product details and the exact wording. A rep who rehearses against a close copy of a real customer's lines learns that one conversation. A rep who rehearses against the pattern learns the moment, and the moment is what will come back on a different call. Do not paste a transcript into a scenario unchanged, and do not quote customers by name in training material.

How many scenarios should a team run at the same time?

Three or four. The largest three families of objections in our data account for 49% of everything buyers raised, across 3,091 calls at five companies over the 90 days to October 4, 2026. Your own spread will differ, but the shape usually holds: a few moments account for most of the loss, and a long tail accounts for the rest. Running more than four at once spreads the practice thin, and nobody can tell which one moved a number. Retire a scenario when live calls stop showing its moment, and bring in the next pile. The size of the set matters less than the habit of replacing it.

How do I know a scenario worked?

Look at live calls, not at practice scores. Pick the moment on the card, count the calls where it appears and the share that ended well, before the rehearsal and after, using the same scoring item each time. If the people who practiced and the people who did not are both in the data, compare them as well. Treat small samples as preliminary: ten calls before and ten after can move by chance. Practice scores tell you whether someone learned to handle the scenario. They do not tell you whether the scenario was the right one to teach. For longer-run effects such as ramp time, compare against earlier cohorts, as we describe in the ramp-time article.