Onboarding · For clinical

Clinical work at Luna with AI alongside

How to use agents for the parts of clinical and regulatory work that aren't your highest-judgment work — and where the boundary lives.

Module 01 · Boundaries

What we're not asking you to do

You're not being asked to put PHI into a general-purpose agent, replace your clinical judgment with a model's draft, or skip the human-review step on anything that goes to a regulator, a site, or a patient. The boundaries here are the same boundaries Luna's clinical work has always had — just written down so the agents respect them too.

Your task

If you ever feel like the system is asking you to skip a step you'd normally take, that's a bug. Tell us — that's a workflow we need to fix.

Module 02

PHI scope and the separate path

Most Luna agents are not cleared for PHI. The ones that are have piiScope: phi in their declaration and live behind the phi-authorized Workspace group. If you're working with patient identifiers, trial data, or anything you'd treat as PHI in a clinical setting, route it through the PHI-scoped agent. There's no shortcut, and no agent will accept the data outside that path.

Your task

Confirm your phi-authorized group membership with John. Bookmark the PHI-scoped agent (when it lands — flagged on the roadmap).

Module 03

Where agents earn their keep in regulatory and clinical work

Drafting protocol synopses, summarizing a long thread of monitor visits, comparing two versions of a study report, generating a plain-English explainer of a 510(k) decision letter, pulling structure from a wall of free-text deviation reports. The pattern: turn a wall of text into a structured starting point you then verify.

Your task

Pick one regulatory or clinical artifact from the last month. Ask Basal (general scope, no PHI) for a structured outline of an analogous one. Compare its structure to yours.

Module 04

Always-human steps

Final review of any document going to FDA, an IRB, a site PI, or a patient. Adverse-event causality assessments. Any judgment that would normally carry your signature. These aren't 'AI-assisted.' They're yours, and the audit log records them as such — the agent's contribution stops at the prior step.

Your task

Add 'human review by <name>' as an explicit row in your project tracker for any agent-assisted regulatory artifact.

Module 05

Surfacing risks back to the platform

If an agent gives you a confidently wrong answer about a clinical or regulatory question, that's the most useful signal we get. Log it (a screenshot is fine), drop it in #ai-platform, and we'll either correct the prompt, narrow the scope, or remove the agent from that path entirely. Hallucinations on regulated work are not 'a known limitation' — they're a fix.

Your task

Find one moment in the last month where an agent (here or elsewhere) was wrong about a clinical-adjacent topic. Write it up; it informs the guardrails we ship next.

Done with this track

Where to go next.

Try the agent for your role, configure it the way you'd like it on My Agents, or read the governance model if you want to know exactly how this works under the hood.

← All tracks Open chat