Insights

Can You Train an AI to Train Your Agents?

September 18, 2026·5 min read

New hires in most contact centers learn the same way they have for decades: shadow a senior agent for a while, read a script, take some test calls, and figure the rest out on live customers while a supervisor hovers nearby. It works, eventually, but "eventually" is the expensive part. A newer approach is trying to close that gap with AI: instead of a script or a shadowing shift, a new agent practices against an AI persona that talks back, pushes back, and grades the conversation. The pitch is a flight simulator for customer service. The harder question is whether the results being reported actually support that pitch, or whether they show something more modest.

What AI roleplay training actually is

The mechanics are fairly simple to describe. Rather than rehearsing a fixed script with a trainer reading the customer's lines, an agent has a live, dynamic conversation with an AI persona built to behave like a real customer, complaints, tangents, and all. The session gets scored against specific behaviors rather than a supervisor's subjective read, and the agent gets that feedback immediately instead of days later in a coaching session. Multiple vendors are now building products around this idea, and the pitch is consistent across most of them: faster time to proficiency, more consistent skill development, and a practice environment where a new hire can fail without it costing a real customer relationship.

The case study that is more complicated than it looks

IntouchCX has been one of the more visible companies publicizing this approach, under a product called Superpunch Roleplay, and its published case study reports a 7 percent boost in first-month CSAT, a 4 percent increase in QA scores, and new agents hitting handle-time benchmarks by week eleven. Those are specific, concrete numbers, which is part of why they are worth looking at closely.

Here is the complication. A separate awards submission from IntouchCX, for what appears to be the same platform, reports a different set of results entirely: 55 percent faster time-to-proficiency, an 18 percent CSAT increase among AI-trained agents, engagement score gains as high as 27 percent, and a 4.2 percent improvement in first contact resolution. Neither disclosure explains the gap, and the two sets of numbers do not obviously describe the same deployment, the same customer base, or the same measurement window. A 2026 AI Excellence Award adds a layer of external validation to the story, but the award announcement used identical language to describe several unrelated winners that same cycle, which limits how much weight it can carry on its own.

None of this means the underlying technology does not work. It means a single public case study, even a genuinely positive one, is not the same thing as a verified, repeatable result. That distinction matters more here than usual, because the entire pitch for AI roleplay training rests on measurable improvement, and measurement is exactly where this example gets murky.

What is more solidly established

Some of what supports this approach does not depend on any one vendor's numbers holding up. Contact center turnover is commonly cited industry-wide as high, often quoted north of 30 percent annually, and new hires are frequently described as needing around 90 days to reach baseline productivity. Neither figure is unique to any single study, but both point to the same underlying cost: the ramp period between hire date and reliable performance is long, and anything that shortens it has a real, calculable value independent of the CSAT percentage attached to any particular pilot.

It is also uncontroversial that traditional roleplay training has real limitations. Scripted scenarios do not adapt to how a specific agent responds, feedback usually arrives well after the practice session instead of during it, and quality often depends heavily on which trainer happens to be running the exercise that day. AI-driven practice addresses those specific gaps by design, whatever the exact percentage improvement turns out to be in a given deployment.

What this means for evaluating the approach

The honest position sits between two extremes. Dismissing AI roleplay training because one vendor's numbers do not reconcile would be an overreaction. Adopting a specific percentage figure from a single case study as an expected outcome would be a different kind of mistake. The more useful approach is treating any AI training tool the way a controlled pilot deserves to be treated: define the metric that matters before the pilot starts, measure the same cohort before and after under the same conditions, and be skeptical of any number that cannot be reproduced outside the vendor's own case study.

That is also the more defensible way to think about whether an AI-plus-human support partner is using this well. The question worth asking is not whether a vendor can cite an improvement percentage. It is whether they can show the measurement behind it, and whether that measurement would survive being run again.

Frequently asked questions

What is AI roleplay training for contact center agents?

It is a training method where new agents practice live conversations with an AI-driven customer persona instead of a scripted roleplay with a human trainer, receiving immediate, criteria-based feedback rather than delayed, subjective coaching.

Does AI roleplay training actually improve CSAT and reduce ramp time?

Early results are promising but not yet independently verified. At least one vendor's own public disclosures for the same product report inconsistent figures across different case studies, which means specific percentage claims should be treated as illustrative rather than guaranteed.

How should a company evaluate an AI agent-training tool before adopting it?

By running a controlled pilot with a metric defined in advance, measuring the same cohort under consistent conditions before and after, and asking the vendor to show the measurement methodology behind any improvement figure rather than accepting a single case study as proof.

Turn this into your numbers.

See how we onboard agents, or book a call and we will map AI-first support to your actual contacts.

Book a call
← Back to all posts