Risk-Tiered Delegation
Use a four-tier model to decide which clinical and administrative tasks to delegate to AI, how much review each needs, and which tasks AI should never own.
Delegation Is a Clinical Judgment
Clinicians delegate constantly: to residents, nurses, MAs, pharmacists, scribes. Good delegation matches the task to the delegate’s competence and sets the right level of supervision. AI is a new kind of delegate with an unusual profile: broad knowledge, excellent language, no accountability, no examination skills, and unpredictable errors.
The question is never just “Can AI do this?” It’s “What happens to the patient if AI does this wrong, and will I catch it?”
Four Tiers of Risk
- 1 Tier 1: Low
Non-clinical or generic content
- 2 Tier 2: Moderate
Clinical drafts you fully review
- 3 Tier 3: High
Input to clinical decisions, as a second look only
- 4 Tier 4: Not delegable
Final decisions and judgment
| Tier | Examples | Required review |
|---|---|---|
| 1. Low risk | Meeting summaries, policy drafts, teaching slides, generic patient-education outlines (no PHI), scheduling templates, appeal-letter structure | Normal proofreading |
| 2. Moderate risk | Ambient scribe notes, discharge instruction drafts, portal message replies, chart summaries, prior-authorization letters with clinical detail | Line-by-line review against what you know and observed, before signing or sending |
| 3. High risk | Differential diagnosis brainstorming, “what am I missing?” checks, literature synthesis for a specific patient, drug-interaction screening | Use only after forming your own judgment. Verify any actionable claim against an authoritative source |
| 4. Not delegable | Final diagnosis, treatment and dosing decisions, triage disposition, goals-of-care conversations, informing a patient of bad news, anything requiring licensure | AI may support preparation; a clinician decides and communicates |
Notes on applying the tiers
- Tier moves with context. A discharge-instruction draft is Tier 2 for a straightforward case, but closer to Tier 3 for a complex anticoagulation plan.
- Validated, purpose-built tools may sit lower than general assistants for the same task. An FDA-cleared ECG algorithm used per its label differs from asking a chatbot to read an ECG.
- Your organization’s policy overrides this model. If a use isn’t approved, it’s not available to you, whatever its tier.
Common mistake: using AI to summarize a complex chart before you’ve read it, then reasoning from the summary.
The summary may be accurate, but you skipped the process that builds clinical understanding, and you can’t detect an omission in material you haven’t seen. For high-stakes cases, read first, then use AI to check what you might have missed, not the other way around.
Augment, Don’t Abdicate
Three questions to ask before delegating any clinical task:
- Severity: If the output is wrong and I miss it, what’s the worst plausible outcome for the patient?
- Detectability: Am I positioned to catch the error? (Did I examine the patient? Do I know this domain well?)
- Reversibility: Can the error be corrected before it reaches the patient?
High severity + low detectability + low reversibility = don’t delegate, or delegate only as a supplementary check.
Watch for deskilling
Repeatedly delegating a skill can erode it. Trainees in particular should build core competencies (note writing, differential construction, literature appraisal) before leaning on AI for them, so they can recognize when it’s wrong.
- List ten tasks from a typical week, both clinical and administrative.
- Assign each a tier (1–4) using the table above.
- For each Tier 2 or 3 task, write the specific review step you’d require, such as “Compare every medication in the draft to the MAR.”
- Identify one Tier 1 task where AI could save you meaningful time this week, and check whether your organization has an approved tool for it.
1. Which task belongs in Tier 4 (not delegable)?
Disposition decisions require clinical judgment and accountability. AI can support preparation, but a clinician decides.
2. What's the safest way to use AI on a complex chart?
Reading first preserves your understanding and lets you catch omissions in the AI output.
3. Which combination makes a task the WORST candidate for delegation?
When errors are dangerous, hard to catch, and hard to undo, keep the task with the clinician.
Up Next
Even a Tier 1 task can become a serious problem if it exposes patient data. Next: protecting PHI when using AI.