Go Back

Verbal De-Escalation Training for Officers using AI in Virtual Reality

I led research and design for an AI VR product that lets officers practice unscripted de-escalation. I owned setup, dispatch, and after-action review, and partnered with engineering on how those actually work in the headset.

I cut mis-clicks from 20% to 2% by making every choice take two intentional confirms.

I cut scenario start from about 5 minutes to under 2, enough for 3x more drills per headset.

92% of officers found the after-action review useful and actionable.

CSAT moved from 3.4/5 to 4.4/5.

New recruits struggle internalizing communication without repetition.

Those opening phrases largely decide whether a situation escalates or resolves. That is the moment agencies cannot practice enough.

First 45 words spoken.

The problem was scale. Live role-plays and workshops are expensive and slow. Agencies cannot run one-on-one drills every month for every officer.

The alpha trained the wrong thing, the hard way

By late 2024 I reviewed an alpha that had run with five agencies. The idea worked. The product did not. I found:

  1. Subject-type controls buried, so officers stayed on one character and missed the learning objective.

  2. Gaze-only UI produced constant mis-clicks.

  3. There was no dispatch channel, so trainers filled in by voice.

  4. Feedback quality changed with whoever was in the room.

Pilot was heading toward non-adoption at the moment this product was meant to differentiate the suite.

Create a frictionless AI-powered Verbal training experience—that maximizes officer learning and throughputs.

I used three tenets to hold user need, engineering limits, and throughput in the same frame.

Frictionless Entry. Get in, train, move on.

I found officers never hit the hidden subject tab. I led a stepper with Next off gaze center, so each choice takes two intentional clicks. I added a What to Expect screen, hooked to AI, and moved mic checks and dispatch teaching into the flow instead of extra reading.

Tailored Realism. Close to a real encounter.

I tested gestures, hand position, and voice book-ends, then I shipped a dedicated dispatch channel: say the call sign to radio, hear static, talk. I kept in-scenario UI to system alerts. Coaching waits for the after-action review so the drill stays immersive.

UI attached to headset

Testing positions and angles

Testing positions and angles

Actionable Feedback. Coach, do not critique.

I put coaching tips on What to Expect. I led an AI after-action review from real transcripts and user-surveyed categories, then a timeline with icons so an officer can jump to the moment that mattered.

I iterated AAR prompts on real transcripts and trainer language. The model drafted coaching. I decided what was allowed to ship. Officers jump to a timestamp. They do not get a wall of generated notes.

Acknowledgements

  1. Product Management: Anne Hermes

  2. Front End Engineering: Adam Rosyark

  3. AI Research & Backend: Mark Cussick, Wright Esposito

  4. Narrative Design: Shad Miller

  5. Illustrations: Kim Lee