Risk-Tiered Delegation
Decide what to delegate to AI by weighing how much the output influences a decision against the consequence of that decision being wrong, and separate GxP from non-GxP use.
Not “Can AI Do This?” but “What Happens If It’s Wrong?”
AI can draft almost anything. The useful question is how much a wrong output would matter, and how likely you are to catch it before it does.
Regulators think about AI risk in a similar way. In its January 2025 draft guidance on using AI to support regulatory decision-making for drugs and biologics, FDA describes model risk as a combination of two factors:
- Model influence: how much the AI output drives the decision, compared with other evidence.
- Decision consequence: how serious the impact would be if the decision were wrong.
That guidance was written for AI used to produce evidence for regulatory decisions, but the same two questions work for everyday tasks too.
The Delegation Grid
| Low consequence | High consequence | |
|---|---|---|
| AI output is one input among many | Delegate freely. Brainstorming, first-draft emails, meeting summaries, learning a new topic. | Assist, then verify. Literature scans feeding a target decision, draft protocol sections, code for an exploratory analysis. |
| AI output largely drives the decision | Delegate with spot checks. Formatting tables, reference-style conversion, internal slide drafts. | Keep AI out, or validate formally. Batch release, causality assessment of an adverse event, final statistical results, labeling text, anything a patient receives without expert review. |
Two things move a task to the right of the grid:
- Proximity to a patient. Anything that affects dosing, safety assessment, eligibility, or what a patient is told.
- Proximity to a regulator. Anything in a submission, a GxP record, or an inspection response.
GxP vs. Non-GxP Use
A simple question sorts most tasks: does this output become part of a GxP record, or support a GxP decision?
- Non-GxP (most discovery, internal communication, general learning): your organization’s acceptable-use policy and approved-tool list apply. Discernment is your main control.
- GxP (manufacturing records, clinical trial data, pharmacovigilance cases, quality investigations, submission content): data-integrity rules, validation, audit trails, and SOPs apply. Using an unvalidated chatbot to create or change GxP content is generally not acceptable without a defined, approved process. Module 5 covers this in depth.
Grey areas are common. A deviation-investigation narrative drafted with AI and then edited is still a GxP record. When unsure, ask your quality unit before you use AI, not after.
Common mistake: using AI for a task simply because the assistant is already open in your browser, without asking where the output will end up.
Before you paste or prompt, take five seconds to ask: where will this output go, and who relies on it? If the answer is “a regulator,” “a patient,” or “a GxP record,” stop and check your SOP.
A Five-Question Delegation Check
- 1 Destination
Where will the output go?
- 2 Influence
How much does it drive the decision?
- 3 Consequence
What if it's wrong?
- 4 Data
Can this tool see these inputs?
- 5 Verification
How will I check it, and can I?
The last question matters most. Only delegate what you can verify. If you couldn’t spot an error in the output, because the topic is outside your expertise or the source data isn’t available to you, the task isn’t a good candidate for AI no matter how low the stakes look.
Worked examples
- Summarize 40 abstracts on a target for a team meeting. Low-to-moderate consequence, AI is one input, you can spot-check. Assist, then verify the claims that drive the discussion.
- Draft the inclusion/exclusion criteria for a Phase 2 protocol. High consequence, but reviewed by medical, statistics, and regulatory. Assist with structure and wording; experts own the content.
- Decide whether a spontaneous adverse-event report is serious. High consequence, a regulatory clock is running, and it drives the decision directly. Keep general-purpose AI out; use only validated tools under your PV SOPs.
Take five tasks from the inventory you built in Lesson 1.1 (or five tasks from your week). For each:
- Place it in one of the four quadrants.
- Mark it GxP or non-GxP.
- Write one sentence on how you would verify the AI’s output.
If you couldn’t write the verification sentence for a task, that’s a sign it shouldn’t be delegated yet.
1. In FDA's 2025 draft guidance, model risk is described as a combination of which two factors?
Risk depends on how much the AI drives a decision and how serious a wrong decision would be.
2. You use a general chatbot to draft a deviation-investigation narrative, then edit it heavily. Is the final text a GxP record?
The record's purpose, not its drafting method, makes it GxP. Check your SOPs before using AI for any part of it.
3. What's the most important question before delegating a task?
If you can't spot an error in the output, you can't safely rely on it.
Up Next
The fourth delegation question was “Can this tool see these inputs?” The next lesson covers confidential, proprietary, and patient data.