Does AI roleplay training work for sales? The research supports practice with feedback, mostly in medicine. It does not yet show that AI roleplay raises sales results. We searched and found no controlled study of AI roleplay with salespeople. Eubrics sells AI roleplay software, so we show our work and our limits.

By Maxim Dsouza

McGaghie and colleagues (Academic Medicine, 2011) found that simulation practice with deliberate practice beat traditional clinical teaching, with an effect size of 0.71 across 14 studies. The authors call the number of reports small.

Cook and colleagues (JAMA, 2011) pooled 609 studies with 35,226 trainees. Simulation beat no training. Most studies were small, and two thirds had no separate comparison group.

Macnamara, Hambrick and Oswald (Psychological Science, 2014, corrected 2018) pooled 88 studies. Deliberate practice explained 14% of the differences in performance overall, and 1% in professions.

We found no peer-reviewed controlled study of AI roleplay with salespeople.

Eubrics’ case studies show before-and-after scores from three teams. They do not describe a control group.

The question

Simulation training in medicine

Deliberate practice and its critics

What carries over to sales

Where the evidence is thin or absent

Conditions for practice to work

Eubrics’ own data

FAQ and conclusion

We found no controlled study showing that AI roleplay training raises sales results. That is the short answer. Eubrics sells an AI sales roleplay platform, so we have a stake in it. We say so here, and we try to show our limits. The main problem is a gap. The best evidence comes from medicine and other fields. Almost none of it was built for selling.

Most answers online come from sellers too. On October 12, 2026, we searched “does AI roleplay actually work for sales training evidence.” Nine results came back. Five came from four companies that, by their own pages, sell AI sales or sales-enablement software: Outdoo, Highspot, Darwin AI and Hyperbound. We did not classify the other four. For a plain definition of the product, see what AI roleplays are. For the case teams make for it, see why AI sales training.

Yes, in medicine. Two large reviews from 2011 found that simulation training helped learners. Neither studied sales. Both are meta-analyses, which pool many studies. An effect size is a standard way to compare results. A bigger number means a bigger gain.

McGaghie, Issenberg, Cohen, Barsuk and Wayne compared simulation with deliberate practice to traditional clinical teaching (Academic Medicine, 2011). They searched MEDLINE, Web of Knowledge and PsycINFO for studies from 1990 to 2010. Of 3,742 articles, 14 qualified. The abstract reports an overall effect size of 0.71 (95% CI 0.65 to 0.76). The authors note that “the number of reports analyzed in this meta analysis is small.”

Cook, Hatala, Brydges and colleagues reviewed technology-enhanced simulation (JAMA, 2011). They pooled 609 studies with 35,226 trainees and compared simulation with no training. Effect sizes were 1.20 for knowledge, 1.09 for process skills, 0.81 for other behaviors with patients and 0.50 for effects on patients. Results varied widely from study to study. Two thirds of the studies were single-group before-and-after designs. The authors point to “important methodological limitations.” Higher-quality studies showed smaller effects. Of 910 outcomes, 690 were skills measured in a training setting. The review left out studies that used human patient actors.

Deliberate practice has strong supporters and strong critics. Ericsson, Krampe and Tesch-Römer introduced it in Psychological Review in 1993. They argued that differences among performers, even elite ones, are closely related to how much of this practice they have done. Ericsson later described it as individualized practice on tasks chosen by a teacher, with a clear goal and immediate, informative feedback (Perspectives on Psychological Science, 2016).

Macnamara, Hambrick and Oswald tested that claim (Psychological Science, 2014). They pooled 88 studies, 111 samples and 11,135 people. After a 2018 correction by the authors, deliberate practice explained 14% of the differences in performance overall. By field, it explained 24% for games, 23% for music, 20% for sports, 5% for education and 1% for professions. The professions group had seven effect sizes and 321 people. It covered computer programming, military piloting, soccer refereeing and insurance selling. The link was not statistically significant (r = .09, p = .377). The authors conclude that deliberate practice is “important, but not as important as has been argued.” The 2014 paper first reported 12% overall and under 1% for professions. The authors say the correction did not change their conclusion. We use the corrected figures.

Read this study carefully. It compared hours of practice with performance across people. It left out experiments that trained one group and not another. So it cannot say whether a training program works.

Both sides make fair points. Ericsson replied that the sports follow-up counted “a much broader and less defined concept” than his definition. Macnamara and colleagues answered that his evaluation had contradictions, omissions and errors. In 2019, Macnamara and Maitra repeated the 1993 violin study with 39 students. They report: “We did not replicate the core finding.” They also found an effect that was substantial but smaller. Both sides agree practice matters. They disagree on how much, and on what counts.

Our reading is that the method carries over better than the results do. This section is reasoning, not a measured finding. The ingredients are repetition, a clear standard and feedback. They appear in the medical studies above. Taylor, Russ-Eft and Chan pooled 117 studies of behavior modeling training (Journal of Applied Psychology, 2005). The abstract states that effects on knowledge decayed over time, while effects on skills and job behavior stayed stable or grew. That setting is closer to a sales job than a hospital simulator is.

Four differences weaken the transfer, in our reading:

• Sales calls are hard to predict. Macnamara and colleagues found practice mattered less for less predictable tasks: 6% of the variance for low-predictability tasks, against 23% for high-predictability ones.

• Most medical outcomes were skills in a test setting. A sales leader wants revenue. Buyer, price and territory also move revenue.

• A surgery checklist is easier to agree on than a good discovery call.

• Medical reviews mostly compare simulation with no training. A sales team’s real alternative is manager-led roleplay. We did not find a study of that comparison.

The evidence is thin for AI conversation practice in any field, and we found none for sales. On October 12, 2026, we searched the open web with a general search engine for controlled studies of AI roleplay or AI conversational agents used in sales or communication training. We did not search PubMed, PsycINFO or Scopus directly. That is why we write “we found no,” not “none exists.”

We found no peer-reviewed controlled study of AI roleplay with salespeople. We found no independent study linking AI roleplay to quota, win rate or revenue.

The closest controlled study we found is a 2025 preprint from the University of Rochester, by Haut, Hasan and colleagues. A preprint has not been through peer review. Fifty-one health students and clinicians were randomly assigned to practice with an AI patient or to read a paper. Blinded raters scored a standardized-patient conversation before and after. The AI group improved more (Cohen’s d from 0.59 to 0.92). Practice lasted up to 30 minutes, retention was not measured, and the topic was cancer communication, not selling. We also found a registered trial plan comparing AI roleplay with peer roleplay for consent training in endoscopy. We found no results for it.

The strongest objection to our view goes like this. Medicine shows large effects across hundreds of studies, and the method is general, so asking for sales trials is too strict. It is a fair point. But those effects compare simulation with no training, mostly in skill tests, in small studies. A sales leader asks a different question. The same bar applies to every sales training method, not only to AI roleplay. We are not saying it fails. We are saying how sure anyone can be.

Practice works best when it has a clear goal, tasks pitched at the learner, and feedback that stays on the task. The sources below state these conditions. Our own reasoning is labelled.

Ericsson (2016) lists individualized tasks, a clear goal and immediate, informative feedback. In McGaghie’s review, some studies had learners practice with feedback until they reached a minimum passing standard. Kluger and DeNisi pooled 607 effect sizes from 23,663 observations (Psychological Bulletin, 1996). Feedback raised performance on average (d = 0.41). Over a third of the feedback interventions lowered it. Feedback worked less well when it pulled attention toward the self and away from the task.

The Taylor team’s abstract says transfer to the job was greatest when practice used trainee-generated scenarios, when trainees set goals, when their managers were also trained, and when the workplace used rewards and sanctions. Cook’s team weakly saw a benefit from spreading training over more than one day and from mastery learning. They also found, against their own expectation, that high-feedback designs often had smaller effects. They warned this could be chance or confounding.

Our reading is that people must actually do the practice. A tool nobody uses cannot help.

It adds three before-and-after pictures. It cannot show cause. Eubrics is an AI sales roleplay platform. It offers AI roleplay bots that simulate customers set up for a team’s product, buyers, pricing and market, plus scoring and coaching plans from real customer calls. The figures below are Eubrics’ own, from our published case studies. All are anonymized.

A US-based insurance organization with 500+ advisors: the average call quality score rose from 62% to 67%. Eight of nine sales skill factors improved, by 3% to 15%. All targeted advisors used the platform and rated it 3.6 out of 5. The rollout also included AI-guided conversations and CRM support, so roleplay alone cannot take the credit.

An IT services and consulting organization with 50,000+ staff: the score rose from 64% to 66%. 94.1% of users logged in, and 51.5% of simulations were completed. Users rated the experience 5 out of 5.

A lifestyle retail brand with 600+ outlets: the score rose from 66% to 81%. 99% of AI bot calls were completed, and 94% of coaching plans were completed (1,265 of 1,355). Associates rated the call experience 4.8 out of 5.

Gains ranged from 2 to 15 points. Completion ranged from 51.5% of simulations (IT services) to 99% of AI bot calls (retail). These are two different measures. Practice only helps if people do it, and completion rates varied a lot.

The pages show before and after scores. They do not describe a control group, random assignment, a measurement period or an outside check on the score. The platform produces the score. The pages do not say whether it comes from practice calls, real calls or both. Treat this as one input, not proof.

Does AI roleplay actually work for sales training?

The research supports the ingredients: practice with feedback on the task, mostly in medicine. We found no controlled study that tests AI roleplay with salespeople. Eubrics’ case studies show score gains in three teams, without a control group. So the honest answer is that it may work when the right conditions hold. That has not yet been shown for sales.

Is AI roleplay just hype?

The evidence does not say hype, and it does not say proven. In medicine, simulation beat no training across 609 studies (Cook and colleagues, 2011), though most studies were small and weak in design. In sales, the direct evidence is thin. Judge any vendor, including us, by the quality of the comparison it shows.

Can a company’s own results show that roleplay works?

They show what happened in one team. They do not show why. A fair test needs a comparison group, a score checked by someone outside the vendor, and a fixed measurement period. Without these, a before-and-after gain could come from new hires, a new product, a new manager, or practice of any kind. That is our reasoning, and it applies to our own case studies too.

The research supports practice with feedback. It does not yet show that AI roleplay training raises sales results. Simulation beat no training across 609 medical studies. In a small review of 14 studies, simulation with deliberate practice beat traditional clinical teaching. The deliberate practice debate says practice matters, but less than its strongest fans claimed. The one professions group in that meta-analysis showed no clear link. Our own data is a before-and-after picture, not a trial.

A controlled study with salespeople would change this view. It should use scores checked outside the vendor, real sales outcomes and a follow-up period. Either result would teach us something.

Until then, ask five questions of any claim that practice works. Compared with what? Measured how, and by whom? How many people, over how long? Did people actually do the practice? Who paid for the study?

We found no controlled study showing that AI roleplay training raises sales results. We did find solid evidence that simulation practice with feedback improves skills in medicine.

If you want to see what practice with scoring looks like for your own team, see what AI sales roleplay involves or ask for a walkthrough with your own scenarios.