AI role play training is practice for hard conversations. A learner talks to an AI avatar that plays the patient, the customer or the colleague, as often as they need, and gets feedback against a rubric. It works when the rubric is fixed, the practice is spaced, and the score measures what was said rather than how the learner felt.
In short
Practice is the scarce resource in training, and content is not. AI role play adds practice that would not otherwise happen: the same hard conversation five times in an afternoon, with a counterpart that stays in character and a score that stays stable. The published evidence supports simulation and spaced practice strongly, and AI role play specifically with small, early studies. In the EU, a training avatar may score what the learner said and how the conversation went. It may not infer emotions from their face or voice in a workplace or a school.
Most people get one chance a year to practise a hard conversation, and it is usually the real one. A junior doctor breaks bad news for the first time to a real family. A new sales rep learns objection handling on a real prospect. A contact centre agent meets their first furious caller on day one of the phones. Role play was meant to fix this. In most organisations it never scaled, because it needs a trainer, a room, a willing partner and a calendar slot.
AI role play training removes the scheduling problem. This article sets out how it works, what the evidence shows, how to measure it, and what the EU AI Act allows a training avatar to score. It ends with six questions to put to any supplier.
This article describes our reading of published research, law and guidance. It is not legal advice. Figures are quoted from the sources listed at the end, with their limits.
What is AI role play training?
AI role play training is practice for hard conversations with an AI counterpart that plays a patient, a customer or a colleague and scores the learner against a fixed rubric. The learner talks, the AI follows a scenario, reacts to what the learner says, and gives feedback afterwards. The counterpart can be a text chatbot, a voice, or a full-body AI avatar that speaks, listens and moves in real time. If you want to see one in action, our product page on AI roleplay training with interactive AI avatars shows how a scenario is set up.
It differs from a training video or an e-learning module in one way that matters. The learner has to produce the words. Watching someone else handle a complaint is recognition. Handling it yourself, out loud, with a counterpart who pushes back, is retrieval and performance. That difference is where most of the learning sits.
It also differs from a quiz. A quiz checks what you know. A role play checks what you do when a person in front of you is upset, evasive or in pain.
Why practice is the scarce resource
Forgetting starts within hours of learning. A 2015 replication of Ebbinghaus’s classic curve found the same steep early drop he published in 1885.[1] Skills decay when they are not used. A meta-analysis of 53 studies found that performance drops the longer a skill sits idle, and that cognitive tasks decay faster than physical ones.[2] Conversation skills are cognitive, and most staff use the hard ones rarely. A nurse may break bad news a few times a year. A rep may face a pricing objection a few times a week. Neither gets a second take.
Most training is never measured. ATD’s 2026 State of the Industry report, covering 340 organisations, found that fewer than a quarter measure whether their training programmes achieve the goals set for them, and fewer still measure return on investment.[3] The same report put formal learning at 16.7 hours per employee in 2025, with direct spend down to $846 from $1,254. Very few of those hours are practice. Most are content.
Leaders know. In LinkedIn’s 2026 Talent Velocity Report, 89 percent of talent leaders said they are concerned about skills agility.[4] In sales, the Bridge Group’s 2025 survey of 351 B2B companies put the average ramp for a new sales development rep at 3.0 months, with a median annual attrition of 40 percent.[5] Every month of ramp is a month of practising on live prospects.
In healthcare the cost of weak conversations is measurable. A meta-analysis of 127 studies found that patients whose physician communicates poorly have a 19 percent higher risk of not following their treatment. Training physicians in communication raised adherence by about 12 percent.[7] A study of 1,265 recorded consultations found that primary care physicians who had never been sued spent about three minutes longer per visit, explained what would happen next, and used more humour than those who had been sued.[8] None of this is taught by a slide deck.
Human role play works, and it does not scale. Five things get in the way, and none of them is about the content. The result is that the people who most need practice get the least of it.
Why Role Play Rarely Scales
Five constraints that limit how much practice a person gets. None of them is about the content.
How AI role play training works
The mechanics are simple. A scenario defines who the avatar is, what they want, what they know and how they behave. A rubric defines what good looks like. The learner starts the conversation. The avatar listens, answers in character and reacts to what the learner does. When the session ends, the system scores the conversation against the rubric and shows the learner what to try next time.
The loop is the point. A learner can run the same scene five times in an afternoon, change one thing each time, and see the score move. That is deliberate practice, and it is the part human role play almost never provides.
The Practice Loop
Five steps, and the fifth is the one that produces the learning.
Three design choices decide whether the loop works.
The avatar has to stay in character. A patient who suddenly knows the diagnosis, or a prospect who agrees to everything, teaches nothing. The scenario needs constraints on what the avatar knows and how it reacts, and the system needs to hold them.
The feedback has to be specific. “Good empathy, 7 out of 10” tells the learner nothing. “You gave the diagnosis before checking what she already knew. Try asking first” tells them what to change. The best feedback quotes the transcript and names the rubric step.
The score has to be stable. If the same conversation scores 6 one day and 8 the next, nobody trusts it. A rubric with anchored levels, where each score has a written description, is what makes repetition meaningful.
AI avatar or chatbot for role play: why the format matters
Most AI role play tools are text chat. Some add voice. A few put a full-body avatar in front of the learner. The format decides what gets practised.
Text practises what to say. Voice adds how to say it: pace, pauses, tone, interruptions. An avatar adds the rest: holding eye contact with someone who is upset, reading a face, managing silence, and the simple pressure of a person in front of you. For a cold email, text is enough. For breaking bad news, it is not. The avatar itself also differs: a 3D AI avatar vs 2D video avatar comparison shows why a real-time 3D body reacts more naturally than a looped video face.
What Each Format Lets You Practise
Eight skills a hard conversation needs. Each format adds to the one before it.
- Word choice and structure
- Questions asked, in order
- Pace, tone and pauses
- Interruptions and filler words
- Eye contact
- Reading a face
- Managing silence
- The presence of a person
Best forScripts, knowledge checks, first drafts of a scenario
- Word choice and structure
- Questions asked, in order
- Pace, tone and pauses
- Interruptions and filler words
- Eye contact
- Reading a face
- Managing silence
- The presence of a person
Best forPhone sales, contact centres, telehealth
- Word choice and structure
- Questions asked, in order
- Pace, tone and pauses
- Interruptions and filler words
- Eye contact
- Reading a face
- Managing silence
- The presence of a person
Best forPatient conversations, feedback, de-escalation, any face-to-face scene
No published study yet compares text, voice and avatar formats head to head. The nearest evidence is indirect. In a 2026 Swedish study, 115 medical students all practised with a talking robot patient. Those who also received AI-written feedback after each case passed their OSCE at 96.7 percent, against 79.6 percent for those who did not.[14] The authors of a 2024 Japanese study that used a mostly text-based AI patient noted that text was weak for nonverbal skills, and said the AI should supplement human patients rather than replace them.[13] Both studies are small, and we return to their limits below.
A real-time avatar also changes how people behave. Learners speak to a face differently from how they type to a box. They slow down, and they notice when the other person looks lost. Those are the habits the training is meant to build.
What the evidence says, and what it does not
Many suppliers in this market quote large numbers with no method behind them. Here is what the published research supports, in three layers.
Simulation works. A 2011 meta-analysis in JAMA pooled 609 studies and 35,226 trainees in health professions education. Compared with no intervention, technology-enhanced simulation had large effects on knowledge, skills and behaviours, and a moderate effect on patient outcomes.[9] Most of those studies were procedural rather than conversational. A 2024 meta-analysis of 10 studies looked at conversation directly and found that virtual reality simulation improved nursing students’ communication skills, with an effect size of 0.44 against usual teaching.[24] Read both as evidence for the method, not for any product.
Spacing and retrieval work. A review of 317 experiments found that spreading practice over time beats cramming, and that the best gap grows with how long you need to remember.[10] A review of ten learning techniques rated practice testing and distributed practice as the two with high utility. Rereading and highlighting rated low.[11] Role play with an avatar is practice testing, and it can be spaced without a trainer.
AI role play specifically: promising, but the studies are small and the results are mixed. This is the layer buyers should read closely.
What the Studies Found
Six results worth knowing. Figures are from the papers.
Two findings deserve a closer look. The Nottingham and Lincoln study gave 378 third-year medical students both an AI patient and a session with actors. Students rated the gain from actors slightly higher, 1.50 against 1.14 on a ten-point scale, and were more satisfied with them. The AI session cost about half as much per student, £33.48 against £61.75.[15] The AI arm was also unsupervised and asynchronous, so the comparison is not clean. The honest reading is cheaper, scalable, and not yet as good as a trained actor.
A 2025 feasibility study of 18 learners rated psychological safety with the AI patient at 5 out of 5, and fidelity at 6 out of 10. The authors wrote that the AI does not yet match the authenticity of a human simulated patient.[17] That pairing, safe but not yet fully convincing, is the current state of the art in one sentence.
Two smaller results round out the picture. A Korean pilot compared an AI chatbot with peer role play and found no significant difference, but it had 19 students and one hour of practice, so it could not have found much.[16] PwC’s 2020 study of virtual reality soft-skills training found that managers finished a course up to four times faster than in a classroom and reported far higher confidence. It was VR rather than AI conversation, and most of its figures are given as “up to”.[12]
What follows from this for a buyer: use AI role play to add practice volume that would not otherwise exist, keep human role play for assessment and for the hardest scenes, and measure your own results. The next sections show how.
AI role play for sales, customer service and healthcare: five use cases
The common thread is a conversation that is high stakes, happens rarely per person, and has a right way to do it that can be written down.
Five Conversations Worth Practising
Each has a written standard, happens rarely per person, and goes wrong in predictable ways.
AI virtual patients for healthcare communication training
Healthcare is the clearest case. The outcome evidence is strong, the standards are written, and the current alternative, simulated patients played by actors, is expensive and scarce. It is also where AI avatars in healthcare already have a foothold, at the reception desk and in patient information. The 2026 Swedish study and the 2024 Japanese study both used AI patients to add practice before an OSCE, with human patients kept for the exam itself.[13][14] The same approach fits AI avatars in education beyond medicine. That split, AI for volume and humans for assessment, is the model we expect medical schools to adopt first.
AI role play for sales training
Sales is where the market started, and where the claims are loudest. The real value is narrower than the marketing. Reps practise the three or four objections that lose most deals, until the words come without thinking. Measure ramp time and call quality.
AI role play for customer service and public services
Customer service and public services share a problem. The hardest conversation of the week lands on the newest person. One contact centre industry estimate puts a new agent at about six months to full productivity, with a few weeks of classroom training at the start.[6] Most of those weeks are about the system, not the caller. Practising de-escalation with an avatar that stays angry until it is handled properly is the closest thing to a safe first shift.
AI role play for corporate training: leadership conversations
Leadership conversations are the quiet opportunity. Most managers give feedback a few times a year and dread it. A ten-minute practice with an avatar employee, the night before, can change how the real conversation goes. Avatars also work at the scale of a conference. At PwC’s Tax Leadership Conference 2025, three RAVATAR avatars ran scenario-based tax and compliance training rotations, described in the PwC Tax and Compliance AI Hosts case study.
How to measure it: KPIs that survive a budget review
Most training is measured by completion and a smile sheet. Role play can be measured properly, because every session produces a transcript and a score. The frame most L&D teams already use is Kirkpatrick’s four levels: reaction, learning, behaviour and results.[18] Each level below maps to something a role play programme can actually report.
Four Levels, Four Kinds of Proof
Levels 1 and 2 come from the platform. Levels 3 and 4 come from the business, and they are the ones a budget holder believes.
Did people use it, and did they come back?
Did the skill improve in practice?
Did it show up in real work?
Did the business or the patient benefit?
Three rules make the numbers credible.
Measure the first attempt, not only the last. The gap between attempt one and attempt five is the learning. A programme that reports only final scores hides the baseline.
Keep the rubric fixed for a cohort. If the rubric changes mid-programme, the trend is meaningless. Version it, and report which version each score used.
Pair every level 2 number with a level 3 or 4 number. A rubric score that rises while call quality stays flat tells you the rubric is measuring the wrong thing. That is useful to know early.
One number to avoid: “confidence”. It is easy to collect and easy to inflate. The 2025 Nottingham study measured self-rated gain and found that it favoured the human actors.[15] Self-reports are a reaction measure. Treat them as one.
How to build a scenario that works
A scenario is a short document. Written well, it takes an afternoon and lasts for years. Written badly, it produces an avatar that agrees with everything and a score nobody trusts. Six parts.
Anatomy of a Scenario
Six parts. The rubric and the approved content are the two that suppliers most often leave to the customer.
- 01One objectiveWhat the learner should be able to do afterwards, in one sentence. “Tell a patient their biopsy result and check what they understood.”
- 02A persona with limitsName, situation, what they know, what they want, what they will not say unless asked. The limits are what make it realistic.
- 03A rubric with anchorsFour to eight criteria, each scored 0 to 5, each level described in words, such as “asked what the patient already knew”.
- 04Approved contentThe facts the avatar may state, signed off by someone accountable. A patient avatar must not invent a drug dose.
- 05Feedback rulesWhat the learner sees after the scene: the score, two quoted moments, one thing to try. Not a wall of text.
- 06A spacing planWhen the learner repeats the scene. Day 1, day 3, day 10, then monthly. The plan is what turns a demo into a skill.
Two practical notes. First, start with three scenarios, not thirty. The three conversations that go wrong most often will carry most of the value. Second, let the people who do the job write the persona. A sales manager knows exactly how a procurement lead stalls. A senior nurse knows exactly what a frightened relative asks first. Their words make the avatar believable.
What the EU AI Act and US laws allow a training avatar to score
This is the section most supplier pages skip, and it decides how a role play tool may be designed in Europe.
Since 2 February 2025, Article 5(1)(f) of the EU AI Act prohibits “the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions”, except where the system is put in place “for medical or safety reasons”.[19] The Act defines an emotion recognition system as one that identifies or infers emotions or intentions from biometric data, such as facial images and, per the guidelines, voice.[19] Recital 18 draws a line between emotions such as anger or embarrassment, which are covered, and readily apparent expressions such as a smile, a frown or a raised voice, which are not, unless they are used to infer an emotion.[19]
The Commission’s guidelines on prohibited practices, approved in draft on 4 February 2025 and adopted on 29 July 2025, read “workplace” broadly, including recruitment, and “education institutions” as all types and levels of education, including vocational and continuous training.[20] The guidelines add one exception of their own. Emotion recognition used only for personal training is allowed if the results are not shared with HR and cannot affect the person’s assessment or promotion.[20][21] That exception is in the guidelines, not in the Act. An employer would have to show the conditions were met, years later, if asked. We would not build a training product on it.
The practical consequence for a training avatar is clear. Score what the learner said and how the conversation was structured. Do not infer how the learner felt from their face or voice.
What a Training Avatar May Score in the EU
The line runs between observing what a person did and inferring what they felt.
From the transcriptAllowed
- What was said
- Questions asked, and in what order
- Rubric steps completed
- Talk time and pace
- Filler words and interruptions
- Facts stated correctly
All of this is observable behaviour and text. None of it infers an emotion from biometric data.
Empathy as behaviourObservable
- Named the concern in words
- Paused after the hard sentence
- Checked what was understood
- Reflected the person back
Empathy can be scored as what the learner did, not as what the avatar guessed they felt.
- Emotion from the face“The learner looked anxious”
- Emotion from the voice“Stress detected in the tone”
- Engagement or sincerity scoresInner states under any label
- Expressions used to infer feelingA smile counted as “warmth”
- Emotion from posture or gestures“Crossed arms read as hostility”
- Sensitive traits from biometricsBeliefs guessed from a face
Banned at work and in education since 2 February 2025, except for medical or safety reasons.
Three more rules that apply
Keep practice scores out of HR decisions. Annex III of the Act lists AI used to evaluate workers’ performance, or to assess learning outcomes in education, as high-risk.[19] The obligations for those systems apply from 2 December 2027 after the AI Omnibus moved the date.[22] A score shown only to the learner, for practice, is not a performance evaluation. A score that feeds a promotion decision is a different system with different duties. Decide which one you are buying, and write it into the contract.
Tell the learner it is an AI. Article 50(1) requires people interacting with an AI system to be told, unless that is obvious from the context.[19] A training avatar is usually obvious. Say it anyway, at the start of each scenario. It costs nothing and removes an argument.
Recordings are personal data. A role play session produces audio, video and a transcript of an employee. GDPR Article 88 lets Member States set specific rules for employee data, and consent from an employee is a weak basis because of the power imbalance.[23] Decide who can see a session, for how long it is kept, and whether managers see scores at all. In several countries a works council will ask the same questions, so answer them before rollout.
EU AI Act Dates for a Training Avatar
Three duties already apply. One high-risk date is still ahead, and it only matters if scores feed HR decisions.
- 2 Feb 2025
Article 5(1)(f) appliesNo inferring emotions in the workplace or education, outside medical or safety uses.
- 2 Feb 2025
Article 4, AI literacyProviders and deployers must support the AI literacy of their staff. Reworded by the Omnibus, still in force.
- 2 Aug 2026
Article 50 disclosure appliesPeople must be told they are interacting with an AI, unless it is obvious.
- 2 Dec 2027
High-risk, Annex IIIOnly if role play scores are used to evaluate workers or assess learning outcomes.
What changes for US customers?
The US has no federal law that bans emotion recognition at work. The rules come from states and cities, and they work differently. EU law bans certain inferences outright. US law mostly asks for consent before biometric data is collected, and for notice, audits or explanations when AI influences an employment decision. The trigger is the same as in the EU: the moment practice scores feed hiring, promotion or selection for training, the tool is in scope.
US Rules That Touch a Training Avatar
State and city laws, not one federal rule. Check where your learners are.
Running it on-premise or in a private cloud
A role play platform records your staff saying difficult things badly, on purpose. It holds the personas your best people wrote, the rubrics that encode how you work, and the approved content a clinician or a compliance lead signed off. For a hospital, a bank or a public body, that is not data to send to a supplier’s cloud by default.
Most tools in this market run only in the supplier’s cloud. Ask where the recordings, transcripts and scores are stored, and who at the supplier can access them. Ask whether sessions are used to train the supplier’s models, and whether the platform can run inside your own network. The same questions apply to any patient-facing AI, and we set them out in detail in Patient-Facing AI in EU Hospitals. The deeper question, whether a supplier can be compelled by a government outside the EU to hand over data, is the subject of What Is a Sovereign AI Avatar? The EU scales a buyer can use to grade a supplier are in Sovereign AI Avatars: The EU Criteria That Decide.
On-premise and private cloud deployment also answer the works council question in one move. If nothing leaves the building, the discussion is about who inside the building can see a session, which is the right discussion to have.
How to choose AI role play software: six questions for a supplier
The answers separate a practice tool from a demo. Ask for each one in writing.
- 01
Can I see the rubric behind the score?
Ask for the criteria, the anchored levels and three scored example transcripts. Ask what happens when two people run the same conversation.
- 02
What does it analyse about the learner?
Ask whether it infers emotion from face or voice. In the EU workplace, the answer has to be no, outside medical or safety uses.
- 03
Who controls what the avatar knows?
Ask how approved content is loaded, who can change it, how changes are logged, and how the avatar is stopped from inventing facts.
- 04
Where do recordings and transcripts live?
Ask for the storage location, retention period, who at the supplier can access sessions, and whether it can run on-premise or in your private cloud.
- 05
Who sees a learner’s results?
Ask whether managers see scores by default, whether the learner can keep practice private, and how that setting is enforced.
- 06
Which metric moves, and from what baseline?
Ask the supplier to name the level 3 or 4 metric they expect to move, and to help measure it for one cohort before rollout.
Where RAVATAR fits
RAVATAR builds real-time interactive AI avatars that listen, speak and move as a full-body person, on a screen, a kiosk or a life-size holographic display. The same platform that lets a patient talk to an avatar at a hospital reception lets a junior doctor practise the conversation that happens next. The product side is described on our AI roleplay training page.
Full-body avatars, in real time
The avatar plays the patient, the customer or the colleague face to face. The learner practises the whole conversation, including the pause after bad news and the moment a customer stands up. Personas, voices and scenarios are set up in Genesis AI Avatar Studio, where the people who do the job can write the persona themselves.
Coaching from the transcript, not the face
Every practice session is saved with its full message history in Conversations, so trainers coach from what was actually said. Feedback rests on the words, the questions and the structure of the conversation, not on the learner’s face or voice. That keeps workplace and education deployments inside Article 5(1)(f). Any capability in this area follows the AI Act requirements.
Four deployment modes
The platform runs on-premise, including on a single edge device, in the institution’s private cloud, on EU sovereign cloud, or on global public cloud. On-premise includes air-gapped installations with no outside connection. In the on-premise and private cloud modes it runs local models, so the language model, speech recognition and speech synthesis stay inside the institution’s network. Recordings and transcripts never have to leave it.
One avatar, every screen
The same persona runs on a laptop for solo practice, on a kiosk in a training room, and on a Holobox display for face-to-face scenes. Holobox is an interface RAVATAR delivers on partner display hardware.
Bring us the conversation that goes wrong most often in your organisation. We will build the first scenario with you and agree the metric before anyone practises.
Start with one scenario
Tell us the conversation your team dreads most. We will show you an avatar that plays the other side, and how your trainers coach from each session.
Start a conversationFrequently asked questions
What is AI role play training?
AI role play training, sometimes written AI roleplay training, is practice for conversations with an AI counterpart that follows a scenario and scores the learner against a rubric. The counterpart plays a patient, a customer or a colleague, reacts to what the learner says, and gives feedback afterwards. It can be a text chatbot, a voice, or a full-body avatar that speaks and listens in real time.
What is AI role play for sales training?
Practice for the conversations that decide deals: discovery calls, pricing objections and renewals. The avatar plays a prospect with a brief, for example sceptical about price and asking for references. The rep practises until the objection no longer surprises them. Measure it with ramp time and call quality scores, not self-rated confidence.
What is an AI virtual patient?
An AI system that plays a patient so that a student or clinician can practise history taking, breaking bad news, consent or intake. It can be text, voice or a full-body avatar. Published studies show it adds practice volume at lower cost than actors, with human simulated patients still used for assessment.
Does AI role play actually work?
The evidence for simulation and spaced practice in general is strong. The evidence for AI role play specifically is early and mixed. A 2026 study of 115 medical students found a higher OSCE pass rate when AI-written feedback followed practice with a talking robot patient. A 2025 study of 378 students found an AI patient cost about half as much as actors but produced slightly lower self-rated gains. Use it to add practice volume, and measure your own results.
Is an AI avatar better than a chatbot for role play?
For conversations that happen face to face, yes. Text practises what to say. Voice adds pace and tone. A full-body avatar adds eye contact, reading a face and managing silence, which are the skills that matter in patient conversations, feedback and de-escalation. For scripts and knowledge checks, text is enough.
Do the same rules apply in the US?
Not the same rules. The US has no federal ban on emotion recognition at work. Illinois requires written consent before collecting face geometry or voiceprints. Illinois, New York City and Colorado require notice, bias audits or explanations when AI influences employment decisions such as hiring, promotion or selection for training. A tool that scores only the transcript and keeps scores private to the learner stays out of most of these rules.
Can an AI role play tool score emotions in the EU?
Not in the workplace or in education. Since 2 February 2025, Article 5(1)(f) of the EU AI Act prohibits AI systems that infer emotions in those settings, except for medical or safety reasons. A training avatar should score what was said and how the conversation was structured, not how the learner looked or sounded.
Is a role play score a high-risk AI use under the AI Act?
Not if it stays with the learner for practice. AI used to evaluate workers’ performance or to assess learning outcomes in education is listed as high-risk in Annex III, with obligations from 2 December 2027. If role play scores feed promotion, pay or formal assessment, treat the system as high-risk and plan for those duties.
How do you measure the ROI of AI role play training?
Set a baseline for one cohort before the programme starts: ramp time, call quality, pass rate or complaint volume. Then report rubric scores from first to fifth attempt, and the same business metric 90 days later. Pair every practice score with a real-world number. Avoid self-rated confidence as the main measure.
Can AI role play training run on-premise?
Yes, on the right platform. RAVATAR runs on-premise, including air-gapped, on a single edge device, or in a private cloud, with local models for language, speech recognition and speech synthesis. Recordings, transcripts and scores stay inside the institution’s network.
How many scenarios do you need to start?
Three. Pick the three conversations that go wrong most often, write a rubric with anchored levels for each, and let the people who do the job write the personas. Add more once the first three are being used and measured.
Sources
Every figure in this article is traceable to one of these. Where a source carries a caveat, the caveat is in the text.
- Murre JMJ, Dros J. Replication and analysis of Ebbinghaus’ forgetting curve. PLOS ONE, 6 July 2015. One subject, nonsense syllables. journals.plos.org
- Arthur W, Bennett W, Stanush PL, McNelly TL. Factors that influence skill decay and retention: a quantitative review and analysis. Human Performance 1998, volume 11, issue 1, pages 57-101. 53 studies, 189 data points. doi.org
- ATD, Highlights from the 2026 State of the Industry report, May 2026. 340 organisations, 2025 data. Fewer than a quarter measure whether training achieves its organisational goals, fewer still measure ROI. Formal learning hours 16.7, up from 13.7. Direct spend per employee $846, down from $1,254. td.org
- LinkedIn Learning, 2026 Talent Velocity Advantage Report, formerly the Workplace Learning Report. 1,240 talent professionals and 607 learners, surveyed September 2025. 89 percent of talent leaders concerned about skills agility, 86 percent of companies lack adequate talent velocity. learning.linkedin.com
- The Bridge Group, SDR Metrics, 10th edition, February 2025. 351 B2B companies. Ramp 3.0 months, median annual attrition 40 percent. bridgegroupinc.com
- ICMI, The true cost of contact center agent attrition, 2017. Practitioner estimate of about six months to full productivity. icmi.com
- Haskard Zolnierek KB, DiMatteo MR. Physician communication and patient adherence to treatment: a meta-analysis. Medical Care 2009, volume 47, issue 8, pages 826-834. 127 studies. 19 percent higher adherence with good communication, 12 percent improvement after communication training. pubmed.ncbi.nlm.nih.gov
- Levinson W, Roter DL, Mullooly JP, Dull VT, Frankel RM. Physician-patient communication: the relationship with malpractice claims among primary care physicians and surgeons. JAMA 1997, volume 277, issue 7, pages 553-559. 1,265 recorded visits. pubmed.ncbi.nlm.nih.gov
- Cook DA, Hatala R, Brydges R, et al. Technology-enhanced simulation for health professions education: a systematic review and meta-analysis. JAMA 2011, volume 306, issue 9, pages 978-988. 609 studies, 35,226 trainees. pubmed.ncbi.nlm.nih.gov
- Cepeda NJ, Pashler H, Vul E, Wixted JT, Rohrer D. Distributed practice in verbal recall tasks: a review and quantitative synthesis. Psychological Bulletin 2006, volume 132, issue 3, pages 354-380. 317 experiments. pubmed.ncbi.nlm.nih.gov
- Dunlosky J, Rawson KA, Marsh EJ, Nathan MJ, Willingham DT. Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest 2013, volume 14, issue 1, pages 4-58. journals.sagepub.com
- PwC, The effectiveness of virtual reality soft skills training in the enterprise, 2020. New managers at 12 US locations. Figures are reported as “up to”. pwc.com
- Okayama University, Enhancing medical interview skills through AI-simulated patient interactions: nonrandomized controlled trial. JMIR Medical Education 2024, volume 10, e58753. AI group 35, control 110. OSCE interview score 28.1 vs 27.1, d 0.48. mededu.jmir.org
- JMIR Medical Education 2026, volume 12, e90368. Quasi-experimental study of 115 Swedish medical students. All practised with a social robot patient (Furhat). AI-written feedback vs no feedback. OSCE pass rate 96.7 vs 79.6 percent. One author is a cofounder of the robot maker. mededu.jmir.org
- JMIR Formative Research 2025, volume 9, e71667. Multicentre randomised crossover study, Nottingham and Lincoln, 378 third-year medical students. AI patient vs actors. Self-rated gain 1.14 vs 1.50, cost £33.48 vs £61.75 per student. formative.jmir.org
- BMC Medical Education 2025, volume 25, article 1755. Pilot randomised trial, 19 Korean medicine students, AI chatbot vs peer role play for OSCE preparation. No significant difference after correction. pmc.ncbi.nlm.nih.gov
- JMIR Medical Education 2025, volume 11, e70766. Mixed methods feasibility study, 18 participants. Psychological safety 5 of 5, fidelity 6 of 10. pubmed.ncbi.nlm.nih.gov
- Kirkpatrick Partners, The Kirkpatrick Model: reaction, learning, behavior, results. kirkpatrickpartners.com
- Regulation (EU) 2024/1689, the EU AI Act, Articles 3(34), 3(39), 4, 5(1)(f), 50(1) and Annex III points 3 and 4, and Recital 18, consolidated text as amended by Regulation (EU) 2026/1744. eur-lex.europa.eu
- European Commission, Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689, C(2025) 5052 final, adopted 29 July 2025. Draft content approved 4 February 2025 as C(2025) 884. Paragraph 254 on personal training uses. digital-strategy.ec.europa.eu
- DLA Piper, EU AI Act: spotlight on emotional recognition systems in the workplace, 7 April 2025. Commentary on the guidelines’ training exception. technologyslegaledge.com
- Regulation (EU) 2026/1744, the AI Omnibus, published 24 July 2026, in force 27 July 2026. Annex III high-risk obligations apply from 2 December 2027. eur-lex.europa.eu
- Regulation (EU) 2016/679, General Data Protection Regulation, Article 88, processing in the context of employment. eur-lex.europa.eu
- Cho MK, Kim MY. The effect of virtual reality simulation on nursing students’ communication skills: a systematic review and meta-analysis. Frontiers in Psychiatry 2024, volume 15, 1351123. 10 studies, overall effect size 0.44 for communication skills versus no intervention or general classes. doi.org
- Illinois General Assembly, Public Act 103-0769 (SB 2979), amending the Biometric Information Privacy Act, 740 ILCS 14. Biometric identifiers include a voiceprint or scan of hand or face geometry. Written release includes an electronic signature. Liquidated damages of $1,000 per negligent and $5,000 per intentional or reckless violation, at most one recovery for the same identifier from the same person. ilga.gov
- Illinois General Assembly, HB 3773, Public Act 103-0804, signed 9 August 2024, effective 1 January 2026. Amends the Illinois Human Rights Act: AI with a discriminatory effect in recruitment, hiring, promotion, renewal of employment or selection for training is a civil rights violation, and employers must give notice when they use AI. ilga.gov
- NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools. Bias audit within one year of use, notice to candidates and employees, enforcement from 5 July 2023. nyc.gov
- Colorado General Assembly, SB 26-189, Automated Decision-Making Technology, 2026 Regular Session. Passed both chambers in May 2026 and replaces SB 24-205. Covers automated decisions in employment and other consequential decisions. Duties start 1 January 2027, with Attorney General rules on post-adverse outcome disclosures due by the same date. leg.colorado.gov
