AI use case

AI roleplay training for customer conversations

AI conversation roleplay training is a simulator where generative AI plays a realistic customer, by voice or text, so staff can rehearse difficult conversations. Service, sales and crisis staff can practice as often as they need before handling live conversations, and receive structured feedback against the organization's own standards.

By Len Debets · Last verified 26 September 2026 · 9 public deployments

20%
Median time to proficiency reduction, vendor reported
4 deployments.
21%
Reported conversion uplift
GoHealth, vendor claim.
USD 32,400 to USD 450,000
Indicative value per year
A contact centre that hires 200 new agents a year. Worked example, see how it is calculated.

What problem does it solve?

New contact centre, sales and crisis line staff learn the hardest conversations on real customers: the angry caller, the fraud victim, the customer in financial hardship, the person in crisis. Classroom roleplay with colleagues or trainers is limited by trainer time, feels artificial and rarely covers the full range of situations, so new hires often reach the floor with little practice. The risk is long ramp up times, inconsistent handling of disclosures and vulnerability, and avoidable harm to the first customers each new hire serves.

Regulated firms have an extra reason to care. Conduct rules expect staff to recognise vulnerability, give required disclosures and treat customers fairly; in the UK, the FCA's guidance on vulnerable customers asks firms to ensure frontline staff have the skills and capability to recognise and respond to vulnerability. Trainer led roleplay leaves little evidence of what was practised and how well.

How does it work?

  1. Build scenarios from real work. Training and quality teams write scenarios from real, anonymized contact reasons: a disputed charge, a lost card abroad, a hardship request, a complaint, a sales conversation with required disclosures. Each scenario has a persona, a goal, facts the trainee must find out and behaviours to test.
  2. The AI plays the customer. A model plays the persona by voice or text, reacts to what the trainee says, becomes calmer or more upset depending on how the conversation goes, and raises the objections or cues the scenario calls for.
  3. Score against the rubric. After the conversation the AI scores the transcript against the organization's rubric (verification steps, required disclosures, empathy, accuracy of information, next steps) and quotes the moments behind each score.
  4. Give targeted feedback and repeat. The trainee gets specific feedback and can retry the same scenario or a harder variant immediately.
  5. Report to trainers, not to discipline. Trainers see progress per skill and per cohort and spend their time coaching where the simulator shows gaps.
Audience
Employee facing
Autonomy
Assist
Adoption
Early adopters
Channels
Internal tools, Phone and voice, Web chat

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI roleplay training for customer conversations
KPIMedianReported rangeData pointsClaimed by
Time to proficiency reduction20% vendor reported
15% to 55%
44 vendor
Conversion upliftToo few to pool
21%
11 vendor
Conversion upliftToo few to pool
about 3x
11 vendor
Interactions handledNot pooled
at least 1 million
11 organization
Users servedNot pooled
about 5000
11 vendor

Value drivers: Employee productivity, Customer experience, Compliance quality, Speed and cycle time.

Indicative value

A contact centre that hires 200 new agents a year

USD 32,400 to USD 450,000

Value of productive time gained by faster ramp up per year

How this is calculated

Formula: hires * rampWeeks * rampReduction * productivityGap * weeklyCost. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
New agents trained per year hires, agents per year200200The reference organization. Replace with your own hiring volume.
Weeks from start to full proficiency today rampWeeks, weeks610Editorial assumption. Replace with your own ramp time.
Share of ramp time removed by simulated practice rampReduction, fraction of ramp time0.10.3Set against the vendor reported evidence on this page, where BillGO, OneTouch Direct and Oracle NetSuite cut ramp time by 15% to 20% and GoHealth's onboarding fell from nine weeks to four (a 55% saving, treated as an outlier).
Productivity shortfall of a new agent during ramp up productivityGap, fraction of a fully proficient agent0.30.5Editorial assumption.
Fully loaded weekly cost of an agent weeklyCost, USD per week9001,500Editorial assumption. Replace with your own cost.

What it leaves out: Ramp up value only. It leaves out trainer time saved, lower early attrition, fewer complaints and conduct breaches from new hires, and the licence and scenario authoring costs of the simulator.

Who already uses it?

9 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Bank of America

United States · Banking · 2024

ScaledGrade B

The Academy, Bank of America's onboarding, education and professional development organization, uses AI conversation simulators in which employees practise different types of client interactions and receive real time feedback. The bank reports more than one million simulations completed in 2024 and says many employees note that practising client conversations helps them deliver better and more consistent service. No proficiency or client outcome figures are published.

  • Interactions handled: at least 1 million, simulations completed by employees in 2024
    "Employees completed over 1 million simulations last year, with many noting that practicing client conversations helps them deliver better and more consistent service."
    Claimed by: organization

U.S. Department of Veterans Affairs

United States · Government and public sector · 2024

ProductionGrade B

The Veterans Crisis Line trains new crisis responders with ReflexAI simulations in which generative AI plays eight Veteran personas, each with its own configured motivation and crisis, so trainees can practise in a low risk setting before live calls. Each simulated call produces a scoring summary of strengths and areas of growth. The tools launched with the first full cohort of new trainees in May 2024, and VA's 2025 AI use case inventory lists the system as deployed and not high impact.

No outcome disclosed.

BillGO

United States · Payments and cards · 2026

ProductionGrade C

BillGO, a bill payment technology company, ran onboarding for its Enrollment Specialists through live, manager led sessions that took five days and scaled to only four new hires at a time. It replaced that with AI role play practice through Second Nature so new hires rehearse customer conversations before taking live calls. The vendor reports new hires now reach proficiency in four days instead of five.

  • Time to proficiency reduction: 20%
    "BillGO Cut Ramp Time by 20% and Scaled Training 3x with Second Nature"
    Claimed by: vendor

Calix

United States · Technology and software · 2026

ProductionGrade C

Calix, a cloud and appliance based platform and managed services company, found its sales training too long for reps to finish and manager feedback on recorded elevator pitches inconsistent, because overloaded managers rarely reviewed them. It replaced those recorded pitches with Second Nature's AI role plays and wove them through its 30, 60 and 90 day onboarding program. The organization reports gains in sales team performance, average deal size and new hire NPS scores, alongside more consistent coaching, but discloses no measured figures for these.

No outcome disclosed.

OneTouch Direct

United States · Professional services · 2026

ProductionGrade C

OneTouch Direct, a business process outsourcing contact centre with more than 2,000 agents across four countries, had agents learning soft skills on live calls, which cost sales opportunities and raised handle times. It introduced Second Nature's AI role play so agents rehearse calls before taking live ones. The vendor reports that one team's conversion rate rose from 12% to 49% within 30 days, and that agents overall reach competency 15% faster.

  • Conversion uplift: about 3x, within 30 days, one team
    "Being able to take that group and within 30 days triple their conversion rates"
    Claimed by: vendor
  • Time to proficiency reduction: 15%
    "Agents reach competency 15% faster, in 550 chats instead of 650 chats"
    Claimed by: vendor

Pacira BioSciences

United States · Pharma and life sciences · 2026

ProductionGrade C

Pacira BioSciences, a pharmaceutical company that makes non opioid pain therapies, needed sales reps to discuss its products with FDA-compliant precision and to redirect healthcare providers back to clinical value in high pressure conversations, while mastering clinical terminology. It introduced Second Nature's AI role play at a national sales meeting, with scenarios that flag unsupported claims and superlatives, then extended it from onboarding to preparation for major sales meetings and to upskilling experienced reps. The organization reports more accurate, compliant and confident customer conversations, though it discloses no numeric result.

No outcome disclosed.

United Rentals

United States · Professional services · 2026

ScaledGrade C

United Rentals, the world's largest heavy equipment rental company, found that sales coaches were too overloaded to give reps enough time practising real customer conversations and that manager feedback was inconsistent. It integrated Second Nature's AI role play into its existing online training, first for onboarding and then for a "Leading Change" discovery program covering 6,000 reps, while keeping human coaches in the loop. The vendor reports an average 30 point rise from a rep's first to last role play score, and a University of Houston study of about 2,000 outside sales reps who used the tool found a revenue gain per extra practice.

  • Users served: about 5000, over 12 months
    "Approximately 5,000 reps completed numerous different role play scenarios over 12 months"
    Claimed by: vendor
  • Revenue uplift: 35.2%
    "The study found that every additional role play practice translates into an extra 1.2% revenue gain, with reps who had been relatively low performing achieving revenue uplifts of 35.2% and those with high performance goals increasing revenue by 29.9%."
    Claimed by: vendor

Oracle NetSuite

United States · Technology and software · 2025

ProductionGrade C

Oracle NetSuite handled onboarding and sales training manually. After a proof of concept with 56 emerging sales reps, it rolled out Second Nature's AI role play platform to its SDR and emerging sales departments, so reps rehearse sales conversations before live calls. The vendor reports that opportunities and sales volume rose, and that SDRs win their first customer and finish onboarding faster than before.

  • Time to proficiency reduction: 20%
    "Among Oracle Netsuite's SDRs, Second Nature cut time to first logo by 21% and overall onboarding time by 20%, while boosting the quality of SQL opportunities."
    Claimed by: vendor

GoHealth

United States · Insurance · 2022

ProductionGrade C

GoHealth, a health insurance marketplace focused on Medicare, uses AI role play partners so licensed benefits consultants practise sales and compliance heavy conversations, both in onboarding and in ongoing training. For new hires, practice is built into the training weeks instead of a separate three week block of practice calls. After a pilot in the fourth quarter of 2022 GoHealth signed a long term contract. The vendor reports shorter onboarding, a higher sales conversion rate in the pilot and a higher trainee to trainer ratio.

  • Time to proficiency reduction: 55%
    "That’s a 55% time saving that allows new hires to start making effective calls five weeks earlier than before."
    Claimed by: vendor
  • Conversion uplift: 21%, Q4 2022 pilot, sales rates 10 days before versus 10 days after about 34 minutes of AI practice
    "Average 21% increase in sales conversions after 34 minutes of practice"
    Claimed by: vendor

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • The contact reason report, to choose scenarios by volume and risk
  • The quality assurance rubric and required disclosures per conversation type
  • Anonymized example transcripts or call recordings for realistic personas
  • Vulnerability and complaint handling policies the scenarios must test

Systems to integrate

  • Learning management system for assignments and completion records
  • Single sign on for trainees and trainers
  • Voice or telephony softphone for realistic practice calls

Complexity: Low

No integration with customer systems is needed. The effort is in writing good scenarios and rubrics with the quality and compliance teams, calibrating the scoring against human assessors and making voice latency low enough to feel like a real call.

  1. 1

    Pick scenarios by risk and volume

    Start with five to ten scenarios that new hires find hardest and that carry conduct risk, such as a hardship request, a scam victim or a complaint. Add routine ones later.

  2. 2

    Write rubrics with quality and compliance

    Use the same rubric the quality team uses on live calls, so practice and assessment measure the same things. Mark which items are mandatory, such as identity checks and disclosures.

  3. 3

    Calibrate scoring against humans

    Have experienced assessors score a sample of simulated conversations and compare with the AI scores. Fix rubric items where they disagree before trainees see scores.

  4. 4

    Keep personas realistic, not cruel

    Let the persona escalate and de escalate in response to the trainee, but keep abuse within what staff actually meet, and give trainees a way to stop a session.

  5. 5

    Blend into the programme

    Integrate practice into training weeks rather than bolting it on, as GoHealth did by folding practice calls into training, and let trainers use the reports to target coaching.

  6. 6

    Measure on the floor

    Compare ramp time, quality scores and complaint rates of trained cohorts with earlier cohorts on the same contact mix.

Guardrails

  • Scores are used for practice and coaching, not for promotion, pay or termination decisions without human review
  • No inference of trainees' emotions from voice or face, which the EU AI Act prohibits in the workplace outside medical or safety reasons
  • Scenarios and model answers use only approved policies, products and disclosure wording
  • No real customer data in personas; examples are anonymized before they become scenarios
  • Trainees can see their transcripts and scores and contest a score with a trainer

KPIs to instrument

  • Time to proficiency per cohort, before and after
  • Practice sessions per trainee and scenario coverage
  • Agreement between AI scores and human assessor scores on a sample
  • Live quality scores and complaint rates in the first three months on the floor
  • Trainee rating of realism and usefulness

Human in the loop

Trainers and quality leads own the scenarios and rubrics, review the AI's scoring on a sample every cohort, and make every certification or sign off decision. The AI gives practice and feedback; a human decides whether someone is ready for live customers.

Common failure modes

Scoring that trainees do not trust
Scores that disagree with what trainers say undermine the tool. Calibrate against human assessors and show the transcript evidence behind each score.
Unrealistic customers
Personas that are too easy or cartoonishly hostile teach the wrong lessons. Build personas from real contact reasons and review them with experienced agents.
Training on outdated policy
The simulated conversation rewards a disclosure or process that has changed. Tie scenarios to the policy owner and review them when policy changes.
Practice data used as surveillance
Using practice scores in performance management kills honest practice and can make the system high risk. Keep practice and performance evaluation separate by design.

What are the risks and rules?

EU AI Act

Depends on design

Used only for practice and feedback, the simulator is limited risk. Article 50 requires that people know they are interacting with AI unless that is obvious from the context, as it usually is in a training session, and the provider must mark synthetic voice or text output as AI generated in a machine readable format.

It becomes high risk under Annex III point 4(b) if its scores are used to evaluate the performance of workers or to decide on their promotion or termination, and can fall under point 3(b) when a vocational training institution uses it to evaluate learning outcomes. Inferring trainees' emotions from voice or face in the workplace is prohibited under Article 5(1)(f), except for medical or safety reasons.

Guidance

Controls to put in place

  • Written purpose limitation that keeps practice scores out of performance management
  • Periodic calibration of AI scores against human assessors, with results recorded
  • Scenario and rubric change control owned by training and compliance
  • Trainee notice of how transcripts and scores are stored, who sees them and for how long
  • Inventory entry for the simulator with an accountable owner

Frequently asked questions

Does AI roleplay actually shorten ramp up time?

Public evidence is early and mostly vendor reported. GoHealth's vendor reports onboarding cut from nine weeks to four after folding AI practice into training, and Bank of America reports more than one million simulations completed by employees in 2024. Measure ramp time and live quality on your own cohorts before and after.

Is AI roleplay training high risk under the EU AI Act?

Not when it is used only for practice and feedback; then the Article 50 transparency rules apply. It becomes high risk if its scores are used to evaluate employees' performance or decide on promotion or termination (Annex III point 4(b)), and inferring trainees' emotions in the workplace is prohibited (Article 5(1)(f)).

Can it be used for sensitive conversations such as crisis calls?

Yes, with care. The US Department of Veterans Affairs trains new Veterans Crisis Line responders on AI simulations with eight Veteran personas, and each simulated call produces a scoring summary of strengths and areas of growth. Scenarios for sensitive topics need expert review and a way for trainees to stop a session.

How to cite this page

Blits.ai AI Use Case Library, "AI roleplay training for customer conversations", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/conversation-roleplay-training. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 29 September 2026: Added deployments: United Rentals, Pacira BioSciences, Calix.
  • 29 September 2026: Added deployments: BillGO, OneTouch Direct, Oracle NetSuite.
  • 27 September 2026: First published

Related use cases

Cross industryBanking

AI quality and compliance monitoring of every customer interaction

Call quality and compliance monitoring is automated quality assurance that transcribes and scores every customer interaction, voice and chat, against the organization's own rubric. It checks required disclosures and script adherence, flags conduct and mis selling risk, and surfaces coaching opportunities, instead of the small sample a human QA team can review.

Deployments
9 public, best grade B
Reported quality score uplift
about 10%
British Gas, vendor claim
Cross industryBanking

Real time AI assist for contact centre agents

Live agent assist is AI that helps a contact centre agent during a live call or chat by transcribing it and surfacing the right knowledge and next step. It also drafts responses and writes the after call summary and CRM notes, while the agent stays in control of what is said and done.

Deployments
16 public, best grade B
Median productivity gain
20.5%
4 deployments
Cross industryTelecommunications

AI sales call coaching and CRM update

Sales call coaching and CRM update is AI for sales teams that analyses sales calls and meetings against the team's own sales method to coach sellers and their managers. It also writes the call summary, next steps and opportunity updates into the CRM for the seller to confirm, aiming at winning deals and building selling skill, not the regulated advice record or general meeting notes.

Deployments
6 public, best grade C
Reported time saved per task
3 minutes
Sandvik Coromant, organization claim
Cross industryBanking

AI agent for first line contact centre service

An AI first line contact centre agent answers inbound customer contact on phone, chat and messaging, resolving general and routine questions end to end in the customer's own language. It routes everything complex, sensitive or regulated to the right human team, with the context attached.

Deployments
39 public, best grade A
Median containment rate
50%
11 deployments, 1 "up to" value left out
Cross industryGovernment and public sector

AI assistant for employee onboarding

An employee onboarding assistant is AI that guides each new employee from signed contract through their first months on the job. It answers first week questions in plain language, tracks the personal onboarding checklist, triggers the paperwork, equipment, access and training steps in the systems that own them, and keeps the manager and HR informed of what is still open.

Deployments
5 public, best grade B
Autonomy
Supervised agent