Does AI roleplay actually reduce ramp time? What the evidence shows, and what to measure instead

Does AI roleplay reduce ramp time? A sales leader who asks gets a percentage, usually from the company selling the roleplay. The percentage sits on a case-study page next to a customer logo, and the page often does not say what was counted, from which start date to which end date, or compared with whom.

That is not a reason to dismiss the tools. It is a reason to separate three claims that get blended into one: that reps get better at practice conversations, that they reach their first live call sooner, and that they reach full productivity sooner. The first has independent evidence behind it. The second has vendor evidence, and some of it is reasonable. The third, which is what "ramp time" means in a capacity plan, has not been shown by anyone I could find.

I run a company that sells an AI roleplay trainer, so read what follows with that in mind.

The best-documented ramp figure measures a shorter schedule, not a faster rep

The most detailed public ramp claim I found is a 55% cut in onboarding time at a health insurance marketplace, and the case study says plainly what it counted: the program went from nine weeks to four. That is a change in how long training lasts. It says nothing about when the new hires started performing like experienced ones.

The case is GoHealth on Second Nature's case-study page, accessed October 21, 2026. GoHealth trains between 1,000 and 6,000 new benefits consultants a year. Before the tool, onboarding ran nine weeks, six of training and three of practice calls. Afterwards the practice calls were folded into training and onboarding took "four weeks and counting", which the page calls a 55% time saving that lets new hires start making effective calls five weeks earlier. On the vendor's index of case studies, the summary card for the same customer describes it as "cutting ramp time by 55%". Between the case page and its summary card, the length of a training program turned into ramp time.

Read carefully, that is a real gain, and it is the gain the evidence supports best: people reach their first live call sooner because practice no longer needs a scarce trainer. The page also states that new hires reach the productivity of established agents, but gives no figure and no timeframe for it. The one measured sales effect comes from a pilot: sales rates compared ten days before and ten days after roughly 34 minutes of training, showing an average 21% increase. The pilot ran in the fourth quarter of 2022, and the same page says GoHealth earns half its annual revenue in the fourth quarter.

None of this means GoHealth is wrong about its own business. It means the 55% answers the question "how long is training", and a buyer who copies it into a ramp plan is answering a different question with it.

Independent trials measure practice skill, which is narrower than ramp

The independent research on AI roleplay is recent and small, most of it comes from medical education, and it supports one modest claim: practicing against an AI counterpart improves how people perform in a practice conversation, compared with reading about it. No trial I found measures time to productivity or follows learners to real customers or real patients.

The strongest design I found is SOPHIE, an AI standardized patient tested at the University of Rochester Medical Center, preprint dated May 5, 2025. Between June and December 2024, 51 students and practitioners were randomized: 26 practiced with the AI patient, 25 read a module with the same content. Each person held a conversation with a human actor before and after, and each conversation was scored by the actor and four reviewers blinded to the group. The AI group improved by 0.15 on a normalized scale, the reading group by 0.06, with an effect size of 0.92. Every one of those conversations was with an actor, though, and none with a patient. Whether a gain like that shows up in a new rep's first month is something their live-call recordings can answer.

Two details matter if you want to apply it to a sales floor. The comparison was reading, the easiest condition to beat. And the AI group started lower, 0.52 against 0.59, then finished about level, 0.67 against 0.65. The authors checked that with a narrower comparison on participants whose starting scores overlapped, and the gap held, but a gain from a low start is still a gain from a low start.

A 2025 randomized crossover study across two UK medical schools compared an AI voice patient with a session run by human actors. Of 396 students, 378 completed at least one survey. Both formats raised self-rated communication skill: by 1.14 points on a 10-point scale for the AI session and by 1.50 for the actors, so the AI session came out slightly worse. The authors name the limits themselves: the outcomes were self-reported, each format was a single short session, and objective assessment would have strengthened the findings.

The third row is the meta-analysis we went through at length in the 14-day onboarding plan. Its short version agrees with the two trials: practice beats a lecture and loses to an active alternative such as hands-on practice or group discussion.

So the honest reading is this. AI roleplay replaces the parts of onboarding where a rep would otherwise read or watch, and it does that better. Whether the skill shown in a practice conversation shows up in the first month of live calls is the step none of these studies took, and it is the step a ramp claim depends on.

A shorter average ramp can come from who left, not from what anyone learned

If ramp is averaged over the reps who made it to productivity, the ones who left before getting there quietly drop out of the number. Any change that makes weak hires leave sooner will shorten that average, even if no one learned faster. AI roleplay with a pass condition is exactly that kind of change.

Picture two cohorts of ten. In the old one, eight reps ramp in six months and two slower ones get there in ten, so the average is 6.8 months. In the new one, the same two slower hires fail the practice gate in week three and leave. The eight who stay ramp in the same six months, and the average is now 6.0. Ramp is down by most of a month and nobody learned anything faster. Losing those two may well be the right call, but that is a hiring decision, and it should not be reported as a training effect. One past cohort's start and exit dates are enough to work out both numbers together.

Season does the same thing from another direction. The GoHealth pilot is a fair example: a ten-day comparison inside a quarter that brings in half the year's revenue would show a rise for almost any change made that month. A team that rolls out roleplay in January and compares the cohort with the one hired the previous summer is comparing two markets, and the training is the smallest difference between them.

The hire date lives in your HR system and the first closed deal lives in your CRM. Ramp gets measured by joining those two dates, and a vendor figure that does not say where its dates came from deserves a follow-up question.

Write two things down before the first rep starts, and keep them fixed. The date that counts as day one, which should be the contract start rather than the first login. And the condition that counts as ramped, in a form someone could check from the CRM without asking you, such as the first month at 80% of quota or a set number of closed deals. If either changes halfway through, the comparison is over.

Ramp is a lagging number, so watch three leading ones and run one fair comparison

Full ramp takes months to observe, which is too slow to steer by. Three numbers arrive within weeks and tell you whether practice is carrying into real work: days to the first supervised live call, first-attempt pass on the gate, and live-call scores in weeks three to eight.

The third one only means something if the new rep's calls are scored on the same form your experienced reps are scored on, and if every call is scored rather than the ones a manager chose.

The third row is the one most teams skip, and it is the closest thing to the evidence the medical trials never collected. If a rep's practice scores climb and their live-call scores in the second month sit where the previous cohort's did, the reps are learning the trainer rather than the job. If both climb, you have the beginning of a claim. Scoring every live call rather than a sample is what makes this cheap; we wrote up how that works in reviewing every sales call without listening to all of them.

A default worth starting from, and it is our recommendation rather than something we have measured: run the comparison against the last two cohorts hired in the same quarter of the year, and do not judge ramp before the new cohort has had as long as the median of those two. Before that point, report the three leading numbers and say out loud that ramp is still open.

If you can split one cohort, do. Half train with roleplay in place of reading and watching, half keep the old plan, everyone keeps their live-call shadowing. It is the only design on this page that answers the question in the headline, and a vendor is poorly placed to run it, because it only sees the half that uses its product.

What to do next

Pull the start dates and exit dates for your last two cohorts from HR, and the first-deal dates from the CRM.

For each cohort, write down two numbers side by side: average ramp for the people who reached your condition, and the share of the cohort who left before reaching it. If ramp got shorter while the share of leavers went up, you have found a filter, not a faster program, and that is worth knowing before you credit any tool with it.

Then write the end condition and the day-one rule on one page before the next cohort starts, and pick which of the three leading numbers you will report at week six. If you are evaluating a trainer, ask the vendor which of their customers measured ramp against a comparison group, and from which system the dates came. A shorter program is still a gain worth having. Just call it that, and keep it out of the capacity plan as a ramp number.

The live-call scores from weeks three to eight do a second job once the cohort is out: they are the starting material for a weekly coaching plan, one behavior per rep per week.

A vendor's ramp figure should come with its denominator. If you are weighing roleplay for your next cohort, bring the question of who was counted and who left. Put it to us on a call.