All Academy resources

AI as Clinical Labor · Part 2 · Free reference

The Clinical AI Autonomy Ladder

How much work can AI do on its own? Six levels of delegation help teams describe what AI may do, what people must still do and where approval or escalation belongs.

A stronger clinical team for a healthier tomorrow.

Same purpose. Greater possibilities.

Download Reference PDF Download Image Download Reading edition PDF Download Phone-readable companion PDFRead the six rungs

An original delegation framework, not a validated maturity, clinical-risk or regulatory scale. Examples are illustrative; classify the deployed task, not the product name. Higher autonomy is not the goal.

Version 1.1 · Editorial review 2026-09-30. Educational framework; not clinical guidance.

The Clinical AI Autonomy Ladder: Observe, Suggest, Draft, Act with sign-off, Act and report, Bounded autonomy. Higher autonomy is not the goal. Full text, conditions, human duties and sources follow.
Framework: AI as Clinical Labor · Jivesh Sharma, M.D. · HealthIT.com. Shared here as an Academy learning companion. Open the full-size reference or use the phone-readable companion.

Classify the deployed task

Six rungs of delegation

  1. 1. Observe

    Flags a signal for human assessment

    AI: Detects signals and surfaces risk.

    Human: Interprets the signal and decides the response.

    Examples: Sepsis risk scores, imaging triage flags, wearable alerts.

    Boundary: Fits this rung when the system flags a signal and a person decides the clinical response. High clinical risk can occur at any rung.

  2. 2. Suggest

    Offers options for human judgment

    AI: Presents options, evidence and uncertainty.

    Human: Selects, modifies or rejects.

    Examples: Treatment options, differential diagnosis, candidate trial matching.

    Boundary: Recommendations support human assessment. Candidate trial matching does not establish eligibility or enrollment; generated rationale is not proof of faithful reasoning.

  3. 3. Draft

    Prepares a work product for review

    AI: Drafts notes, messages or other text.

    Human: Reviews and finalizes before use or send.

    Examples: Ambient note drafts, draft inbox replies, prior-auth drafts.

    Boundary: Draft text remains under human review. A system that prepares an executable submission awaiting approval belongs on rung 4 for that step.

  4. 4. Act with sign-off

    Stages an action for human approval

    AI: Prepares an action; releases it after approval.

    Human: Approves or rejects before release.

    Examples: Pended refill orders, referrals awaiting approval, pended order sets.

    Boundary: The executable action requires individual authorization before release. Automatic protocol-based refill or referral actions belong on rung 5 under this framework.

  5. 5. Act and report

    Executes specified, preauthorized actions

    AI: Follows an approved protocol; logs actions.

    Human: Responds to exceptions and reviews outcomes.

    Examples: Protocol scheduling, approved result notices, care-gap reminders.

    Boundary: Use explicit rules and timely escalation. Result notices mean approved nonurgent communications, not unsupervised interpretation, treatment changes or critical-result management.

  6. 6. Bounded autonomy

    Makes a defined, validated decision

    AI: Decides within a specified task boundary.

    Human: Sets limits, monitors performance, handles escalation.

    Examples: Autonomous DR interpretation within labeled use (LumineticsCore).

    Boundary: The autonomous decision here is interpretation of retinal images within the device's intended use. Staff acquire images; clinicians handle results, referrals and treatment. The AI does not perform the whole eye-care pathway.

Before delegating a task

  1. Define the task. Name the function or functions, inputs and decision boundary.
  2. Name the human duties. Record who approves, responds to exceptions and can stop the task.
  3. Match the controls. Use clinical risk, autonomy and reversibility to choose safeguards.
  4. Measure the whole burden. Count time saved after review, correction and exception handling.
  • Choose the appropriate rung. Advancement is not the goal.
  • One product can occupy different rungs for different tasks.
  • Risk, evidence and permitted use constrain delegation.
  • Clinical and organizational accountability remain with people.

The retinal-screening example

LumineticsCore, formerly IDx-DR, autonomously detects more than mild diabetic retinopathy in adults aged 22 or older with diagnosed diabetes and no previous diabetic-retinopathy diagnosis, using the supported Topcon NW400 camera. This is bounded diagnostic interpretation, not a complete eye examination or autonomous treatment. Trained staff capture images. The responsible clinical team reviews outputs, manages inadequate-image results and arranges referral or follow-up. Refer to current labeling for the complete indication, exclusions and warnings.

Connect delegation to your learning

Use oncology workflows to examine the task, supervision and evidence. Start with the free first lesson; modules 5 and 7 extend the exercise within the course.

Free lesson

The Oncology AI Landscape

Recognize the task behind an AI claim.

Course module

Oncology Workflow Fit

Map approval, exception handling and the whole burden of supervision.

Course module

Applied Use-Case Memo

Describe a bounded use of AI with evidence and accountable human duties.

Part 1: The Healthcare AI Labor Stack

Sources and labeling

Explore on HealthIT.com

HealthIT.com provides the open reference library. The Academy connects these ideas to structured study, practical exercises and assessments. The course offers a Certificate of Completion; it is not CME or a professional credential.

Explore HealthIT's free resource library

HealthIT.com is run by HealthIT, the company that operates Oncology AI Academy. Its resources are free and open.