Meridian AI

Meridian rolled out four AI assistants: one for patient help, one for HR policy, one for clinical protocols, and one for IT support. Staff across clinical, administrative, and IT roles were using them without shared guidance on which assistant fits which question, how to ask clearly, what information to keep out of a chat, or how to tell whether a confident, cited answer is actually right.

Build safe, accurate AI habits. Staff should choose the right assistant, write a clear prompt, protect sensitive information, check each answer against its source, and know what to do when an answer doesn't hold up.

I designed a story-based simulation set during one shift on a Meridian cardiology unit, inside a rebuilt version of the MeridianAI interface. Colleagues, a unit educator, and a visitor each bring the learner a request across five core missions, and support fades as the shift goes on. Mission 1 highlights each step and coaches the learner through it. Missions 2 through 4 offer hints only when the learner asks. Mission 5 is an independent, scored challenge with no hints. Learners build prompts with a structured builder, so prompt quality is scored against clear criteria instead of keyword matching. Their choices lead to scripted consequences: a vague prompt gets a vague answer, the wrong assistant gets a mismatched reply, and following an unsupported answer has a realistic workplace cost. Every misstep offers a way to recover. One mission presents a confident, well-cited answer that answers a completely different question. Each mission ends with a debrief of the learner's decisions and what they demonstrated. Two optional side missions extend the practice: spotting a citation that doesn't support its claim, and building and testing a custom AI agent with clear boundaries.

Delivered a complete, self-contained simulation with five core missions, two side missions, and a scored challenge. The challenge rates five criteria, requires both safe information handling and source verification to pass, and sends learners who miss either one to targeted remediation. First attempts and retries are tracked separately, so the score reflects what the learner did, not how many times they tried.

AI training is judgment training. The skill that matters most isn't which button to click. It's knowing when to trust an answer, how to check it, and what to do next when it doesn't hold up. Fading support from guided to independent practice lets learners build that judgment before they're scored on it.