VR Soft Skills Training with AI: Role-Play That Pushes Back

Most soft skills training still means a video, a multiple-choice test and maybe one awkward role-play with a colleague who goes easy on you. Neocortex lets a VR scenario put a trainee in front of a character who doesn't: an upset guest, a nervous patient or a tough buyer who reacts to every word they choose and gets calmer or angrier as a real person would.

What it looks like

The Albury is an invented city hotel, rebuilt in VR for a hospitality training programme. A new front desk trainee puts on a Meta Quest headset and finds herself behind the reception counter at eleven at night. Through the revolving door comes Harriet Lowe, wheeling a suitcase and already annoyed. Her room was given away an hour ago.

Harriet: I booked this room four months ago. Four. And you're telling me it's gone? Trainee: That's not okay, and I'm sorry. Let me look at what I can do for you right now. Harriet: Fine. But I'm not sleeping in a broom cupboard.

The trainee checks the virtual booking screen, and the app sends an event: one suite is free on the top floor. She offers it, at the original rate, with breakfast. Harriet's next line comes back tagged as relieved, she uncrosses her arms, and the reply carries a TAKE_KEYCARD action that the app plays as her reaching for the card.

A different trainee runs the same scenario and opens with "Well, that's hotel policy." Harriet's emotion turns angry, her reply carries ASK_FOR_MANAGER, and the scene gets harder. Neither path was scripted. The trainer defined who Harriet is, what she wants and which checkpoints a good response passes through.

When the scene ends, the headset shows a debrief: the checkpoints the trainee reached, and four quick questions about what the hotel's complaint guidelines say to offer.

How Neocortex makes it work

• AI characters. Harriet is designed in the Neocortex visual node editor: personality, backstory, what she expects, what will calm her down and what will make things worse. Trainers can rehearse with her in the browser dashboard before anyone builds the lobby.
• Emotions. Every line carries an emotion tag. Your animation system turns them into posture and expression, so the trainee reads the guest the way they would at a real desk.
• Conversation Flows. Start, Checkpoint and End nodes define the scenario: acknowledge, apologise, offer a solution, confirm the next step. Each reply returns a flowState keyword your app logs for assessment. Flows are available on every plan.
• Events. What the trainee does with their hands matters too. Checking the booking screen, calling housekeeping or handing over a voucher goes in as an event, and Harriet responds to it.
• Actions. TAKE_KEYCARD, ASK_FOR_MANAGER or WALK_AWAY come back with replies, and your scene plays them.
• Voices, speech-to-text and lip sync. Trainees speak naturally into the headset, the character replies in a streamed voice, and the Unity SDK drives her facial blendshapes so her mouth moves with the words.
• Quiz engine. Question sets and scoring through the API make the debrief (Team plan).
• Usage controls. Send a playerId per trainee and track usage per player and per character, with analytics in the dashboard, so a programme lead can see who has practised and how often.
• Knowledge base. Upload the company's service guidelines so a coach character can explain in the debrief what the policy actually allows.

Ideas to build

Healthcare patient intake. A nervous patient, Tomasz, arrives for a procedure he has been dreading and gives short, evasive answers. The trainee nurse has to put him at ease and still collect the history. Checkpoints track whether allergies, medication and consent were covered, and his emotion tags show whether he is relaxing or shutting down.

Sales objection practice. A procurement lead called Ruth Ambler likes the product, hates the price and has a cheaper quote from another supplier on her desk. Sales trainees practise handling each objection in turn. The flow records whether they asked about her real constraint before discounting, and a quiz afterwards checks they know the pricing rules.

De-escalation training. A shopper in a busy store is shouting about a refund, and the trainee is a security officer or store manager. Events from the scene, such as a crowd forming or a colleague arriving, raise or lower the tension. A CALM_DOWN or STORM_OUT action tells the app how the scene resolved.

The second voice. For scenes with two people, such as a couple disagreeing at the hotel desk, group chat puts both characters in one conversation (Pro and Team plans). Conversation Flows do not run in group chat, so assess these scenes from events and your own scoring instead.

Building it

VR training runs on Meta Quest through Unity with the XR Interaction Toolkit and the Unity SDK. Our open-source VR demo is a good starting point for the headset side.

1. Write the scenario like a trainer would. Who the character is, what went wrong, what a good outcome looks like and the mistakes you want trainees to experience.
2. Create the character in the visual editor with personality, voice, emotions and a few body-language actions.
3. Turn the assessment criteria into a Conversation Flow, one checkpoint for each behaviour you want to see.
4. Rehearse in the dashboard with good, mediocre and bad responses until the character reacts the way an experienced trainer expects.
5. Build the scene from the VR demo: microphone in, voice and lip sync out, actions mapped to animations, events for what the trainee does.
6. Log the flow states per trainee, add a debrief quiz on the Team plan, and review usage in the dashboard.

Make the first version of each character a little easier than real life. Trainees who succeed once are more willing to replay the harder version where the guest doesn't calm down so quickly.

Soft skills training is one entry in our AI NPC, ambient AI and voice command use cases. For other VR ideas read AI characters for VR escape rooms, and for classroom role-play see AI conversation partners for language practice. The education and training and VR experiences pages cover the rest. Start with one difficult guest and four checkpoints.

Checked October 2026.

Frequently Asked Questions