AI has moved quickly from drafting emails to running autonomous agents, and corporate training is following. L&D teams now use AI to let employees practise real conversations, get feedback on each skill, and show leadership what changed, at a scale that used to be impossible.
The data shows what works. Across 47,043 AI practice sessions on VirtualSpeech in 2026, at 465 organisations and universities, people who practised the same scenario repeatedly improved steadily: from 5.5 out of 10 on their first attempt to an average best score of 7.0 by their fifth. But the median learner only ran two sessions in total. AI makes practice possible; programme design makes it happen.
This guide covers why traditional training falls short, seven ways L&D teams use AI, what our data shows, and how to introduce AI into your own training programme.
Why traditional training falls short
Employees expect learning to be relevant, flexible, and connected to their role. Long classroom sessions and generic e-learning often miss on all three. The common problems:
- One-size-fits-all content that ignores individual skill gaps.
- Low engagement and poor retention. Asking people to take an hour out for a workshop or a generic video is a hard sell, and much of it is forgotten.
- Little practice. People watch and read about skills such as feedback or negotiation, but rarely rehearse them before using them for real.
- Limited insight. Completion rates show who attended, not who improved.
- Hard to scale. Personalised coaching across hundreds of employees and several countries is expensive.
AI addresses each of these by making training more responsive, more practical and easier to measure.
7 ways L&D teams use AI in corporate training
1. Practice with AI roleplays
The most effective way to learn a skill is to do it. AI roleplays let employees rehearse realistic conversations, such as giving feedback, handling an angry customer or delivering bad news, with an AI character that reacts to what they say. They can repeat it until they improve, in a safe setting, before the real thing.
2. Feedback on every skill, not just a score
AI can score each part of a conversation separately: listening, clarity, empathy, agreeing next steps. That matters because a single score hides weak spots. In our data, the gap between a person’s best and worst skill in the same conversation averaged 2.3 points out of 10, and in 27% of sessions it was 3 points or more.
3. Personalised coaching
AI coaches can review each attempt with the learner, help them plan a real conversation, and suggest what to practise next. Performance in one session shapes the next without an L&D team having to step in, which used to be prohibitively expensive at scale.
4. Custom scenarios and content, built in minutes
AI makes it practical to build training around your own situations: your products, policies and the disputes your managers actually face. Organisations clearly want this. In 2026, 47% of practice sessions at companies and public bodies on VirtualSpeech were on scenarios they had built themselves, and 410 custom scenarios were used at least 10 times. AI also speeds up writing course material, quizzes and summaries, freeing L&D teams for higher-value work.
5. Training in the learner’s own language
Global organisations no longer need to translate and re-record every course. AI roleplays and feedback can run in many languages. In our data, 36% of practice sessions were in a language other than English.
6. Measuring skills, not attendance
Completion rates show who turned up. AI practice produces data on what people can actually do, by skill, team and cohort. L&D teams can see skill gaps across the organisation, target training where it matters most, and show leadership how skills changed over a programme.
7. Flexible, engaging and accessible
AI practice runs in a browser or a VR headset, in sessions of 5 to 10 minutes that fit between meetings. 92% of sessions in our data took place on weekdays, so people fit practice into the working week. Scoring and leaderboards add healthy competition, and because practice runs on a normal computer, nobody needs special equipment or travel to take part.
What 47,000 practice sessions show
Practice with AI feedback works, but only if people come back. When someone practised the same scenario at least five times, their best score rose with every attempt: from 5.5 out of 10 on the first try to 7.0 by the fifth. Their score on any single attempt rose more slowly, from 5.5 to 6.0, because some attempts dip as people try new approaches.
The catch is that most people stop early. The median learner on an individual account ran two sessions in total, and in 65% of cases people tried a scenario only once. For L&D teams, that is the main lesson: AI provides the practice and the feedback, but programme design (assigned scenarios, a deadline, a required second attempt) is what turns access into improvement.
AI supports people; it doesn’t replace them
AI doesn’t replace trainers, facilitators, coaches or managers. It takes on the repetitive parts (scoring practice, giving first-pass feedback, spotting patterns across teams) so people can spend more time on coaching, strategy and human connection.
The best programmes blend the two. A trainer teaches a model in a short session, employees practise with AI that week, and the trainer uses the skill data to decide who needs one-to-one support. For employees, AI practice is a safe place to make mistakes before a real conversation, not a test that decides their future. Being clear about that from the start builds trust and encourages people to use it.
How to introduce AI into your training programme
- Start with a business problem, not a tool. Pick one skill gap that matters, such as new managers struggling with performance conversations, and define what success looks like.
- Run it as a programme. Give it a cohort, a start date, assigned scenarios and a results report. In our data, nearly 30% of learners on individual accounts only ever ran one session. Structure is what turns access into practice.
- Teach briefly, then practise. Pair each short lesson with practice in the same week.
- Build in a second attempt. Across all scenarios, 60% of people improved on their second try. Most won’t retry unless asked.
- Use your own scenarios. Adapt library scenarios or build custom ones around the conversations your people really have.
- Measure skills before and after, and share the results with managers and leadership.
- Review and refine each quarter, using the skill data to decide what to train next.
Risks to plan for
- Data privacy and security. Check where recordings and transcripts are stored, and the vendor’s security certifications.
- Accuracy. Ask how the AI’s scores have been checked against human assessors.
- Fairness. Check that scoring works fairly across accents and languages.
- Regulation. In the EU, AI used to evaluate employees’ performance or to make hiring and promotion decisions is classed as high-risk under the AI Act. Keep practice developmental, and take legal advice before using scores in formal decisions.
Frequently asked questions
How is AI used in corporate training?
Mainly for practice roleplays, personalised feedback and coaching, building custom training content, delivering training in several languages, and measuring skills rather than completion.
Does AI training actually improve skills?
In our 2026 data, people who practised the same scenario five or more times improved from an average of 5.5 out of 10 on their first attempt to a best score of 7.0 by their fifth attempt. Across all scenarios, 60% of people improved on their second attempt.
Will AI replace L&D teams or trainers?
No. AI handles repetitive scoring and first-pass feedback, so trainers and coaches can focus on the human side of learning.
What skills can employees practise with AI?
Communication-heavy skills work especially well: leadership and management conversations, feedback, coaching, customer service, sales, negotiation, interviews and presentations.
How do you measure the impact of AI training?
Compare skill scores at the start and end of a programme, track who practised and how often, and connect the results to business measures such as engagement, retention or customer satisfaction.
How we did this analysis
We analysed 47,043 completed AI roleplay sessions on VirtualSpeech from 1 January to 13 October 2026.
- Who: 5,321 accounts at 465 organisations and universities, plus individual learners, across 1,380 scenarios.
- Library vs custom: 64% of sessions were on VirtualSpeech library scenarios and 36% on scenarios customers built themselves; at companies and public bodies, custom scenarios made up 47% of sessions.
- Languages: we classed a session as English if its written feedback was in English.
- Improvement: some organisations share one login between several people, so improvement figures use only individual accounts. The “first to fifth attempt” figures cover 727 cases, involving 467 people, where someone did the same scenario at least five times.
- Excluded: anonymous free-trial sessions, internal and test accounts, and sessions with no valid score.
- Limits: scores come from AI, not human assessors. We are running a study comparing our AI scores with certified coaches. People who practised more chose to, so their results may not represent everyone.
All data was anonymised and aggregated.
Related reading:
