PMX internships in the age of AI and agentic coding

Working specification

GenAI
Pharmacometrics
What a pharmacometrics internship should look like when the intern works with AI assistants and coding agents: the question, what has changed since 2020, where mentor time goes, expectations for mentor and intern, and how to tell whether it worked.
Last edited

October 21, 2026

1. Goal

Two deliverables:

  1. A one-page expectations document that the mentor and the intern both read in the first week, saying what each owns, how AI tools are used, and when they meet.
  2. A plan for the next PMX internship built on it, with a record kept during the internship of mentor hours and what the intern delivered, so the plan can be judged at the end.
  3. A validation plan for the internship project, written in week 1 with the specification, saying how the intern’s work is checked and by whom (Section 6).

Out of scope: recruiting and selecting interns, and company policy on which AI tools may see which data. The policy is a constraint the plan works inside, stated in Section 5, not something this project sets.

2. The Question

Can an intern working with AI tools need less mentorship than a 2020 intern and still be a multiplier for the team? A multiplier here means the team gets more done with the intern than without, after counting the mentor’s hours.

Three answers are possible, and the plan should be able to tell them apart.

Three answers to the question
# Hypothesis What it predicts
H1 Substitution. The assistant does what the intern would have done, so the intern adds little the mentor could not get by prompting directly. Mentor hours stay the same or rise; output is no more than the mentor plus an assistant would produce.
H2 Multiplication. The intern with an assistant does more than a 2020 intern, and needs less of the mentor’s time to do it. Mentor hours fall; output rises.
H3 Shift. Mentor hours stay about the same but move: less on how to write the code, more on choosing the question and judging the answer. Same hours, different meetings; output rises.

H3 is the working guess. Nothing here tests it yet.

3. What Has Changed Since 2020

Each row is a task a 2020 intern did, and who now does the first draft of it. The 2026 column is a hypothesis to check against the next internship.

Who does the first draft of each internship task
Task 2020 2026
Write model code (NONMEM, nlmixr2, Stan) Intern, slowly, with mentor help on syntax Assistant drafts; intern runs and checks
Debug a failed run Intern and mentor together Assistant suggests; intern decides
Plots and diagnostics Intern, from templates Assistant
Literature search and summary Intern Assistant drafts; intern checks against the papers
Report draft Intern; mentor rewrites Assistant drafts; intern edits; mentor reviews
Choose the question to answer Mentor Mentor, with the intern learning to
Judge whether the model is right for the decision Mentor Mentor, with the intern learning to
Know what the team needs and who will use the answer Mentor Mentor

The rows the assistant took over were the rows where most 2020 mentor time went. The rows it has not taken over are the ones that make an analysis useful to a team, and they are the ones an intern most needs to learn.

4. Where the Mentor’s Time Goes

The mentor’s influence concentrates at a few points. Listing them is the first step to deciding which ones can be lighter.

Mentor touchpoints and where their weight moves
# Touchpoint When 2020 weight 2026 guess
1 Choosing and scoping the project Before day one High Higher. A well-scoped project is what lets the intern run with an assistant
2 Kickoff: the written specification Week 1 Medium Higher. The specification is the instruction set for the intern and the agent both
3 First review of the intern’s output Week 2 or 3 Medium High. Where to catch work that runs, looks right, and answers the wrong question
4 Weekly meeting Weekly High, mostly code help Lower, and about decisions rather than code
5 Ad hoc questions Any time High Lower; the assistant answers most
6 Midpoint review against the specification Midway Medium Medium
7 Final presentation to the team End Medium Medium. The intern defends the work without the assistant in the room

The guess, if right, moves mentor time to the front of the internship (touchpoints 1 to 3) and out of the middle (4 and 5). This is the same move as writing the specification before commissioning an agent, described in Scoping a project using Agentic Engineering Framework.

5. Expectations

5.1 The Intern

  1. Uses AI assistants and coding agents for code, plots, searches and drafts, within the company’s rules on which tools may see which data. Clinical data does not go into a tool the rules exclude.
  2. Can explain every line of code and every number in the final report without the assistant. The final presentation checks this.
  3. Checks every claim drawn from a paper against the paper, not against the assistant’s summary of it.
  4. Writes down what the assistant got wrong. These notes are a deliverable; they are what the next intern and the mentor learn from.
  5. Brings decisions to the weekly meeting, not code problems.
  6. Owns the specification after week 1, and proposes changes to it in writing.
  7. Runs the validation plan on their own work, including the parts the assistant wrote, and records what each check caught.

5.2 The Mentor

  1. Scopes the project before the intern arrives, to a question the team needs answered, with a named person who will use the answer.
  2. Writes the first specification with the intern in week 1.
  3. Reviews the first real output closely, in week 2 or 3, for whether it answers the question rather than whether it runs.
  4. Keeps the weekly meeting to decisions, and says so.
  5. Records hours spent on the internship each week, for Section 7.
  6. Does not rewrite the intern’s work. Comments go back to the intern, who may use the assistant to act on them.
  7. Writes the validation plan with the intern in week 1, and repeats a sample of the intern’s checks rather than all of them.

6. Validating the Intern’s Work

The intern’s work is checked with the same validation plan that would check an agent’s, taken from the Validation Tools Taxonomy. An intern and a coding agent are both authors whose work someone else has to be able to trust, and the questions about the work do not change with the author. What changes is where errors are expected: an intern is more likely to miss an implausible parameter value or a data convention, an agent to write code that runs and computes the wrong thing.

The internship puts the intern in two roles. The intern is the first reviewer of everything the assistant drafts, and the mentor is the reviewer of the intern. Learning to write and run a validation plan is therefore part of what the internship teaches, and an intern who goes on to direct agents will keep using it.

The rows below are a first assignment for a typical internship project, using the taxonomy’s thirteen kinds. The plan written in week 1 replaces it.

First assignment of validation checks for an internship project
# Kind Checked by When
1 Question Mentor Before day one, and again at the final presentation
2 Study design Mentor, when choosing the study the project will use Before day one
3 Causal assumptions Intern draws the causal diagram, or says why none is needed; mentor reviews Week 1, with the specification
4 Assay Intern finds out what each assay measures and how values below the limit of quantification are reported; mentor reviews Week 1
5 Analysis dataset Intern, with the Validation Checklist; mentor reviews Touchpoint 3, the first review
6 Sources and prior knowledge Intern, against the paper; mentor spot-checks As each source is used
7 Code Intern, with the R coding checklist; code the assistant wrote counts as agent-written Throughout
8 Model Intern runs the diagnostics; mentor judges plausibility Touchpoint 3 and the midpoint review
9 Predictions Intern, with the uncertainty checklist, and against a simple benchmark model where the main model is complex Before the final presentation
10 Conclusions Mentor Before the final presentation
11 Communication Intern, with the presentation checklist; team gives feedback on the form Final presentation
12 Reproducibility Mentor reruns the analysis from its inputs End of internship
13 Use Mentor Three months after

The checks map onto the mentor touchpoints in Section 4. The first review at week 2 or 3 is where kinds 5 and 8 are checked, which is one reason the 2026 guess gives it a high weight.

7. How to Tell Whether It Worked

At the end of the internship, compare against Section 2’s hypotheses with:

  • Mentor hours, total and by touchpoint, from the weekly record.
  • Output: what the intern delivered, against the specification.
  • Use: whether the named person used the answer, checked three months after the internship.
  • Learning: whether the intern can explain the analysis and defend its choices without the assistant, judged at the final presentation.
  • Validation record: what each check in the plan caught, and whether the mentor’s sampled checks found anything the intern’s had missed. The same record is evidence for the Validation Tools Taxonomy on which checks earn their time.

One internship cannot separate H2 from H3 with confidence. It can rule out H1 if the output clearly exceeds what the mentor would have produced with an assistant in the same hours.

8. Open Questions

  1. Does an intern who leans on the assistant learn pharmacometrics, or learn to prompt? Section 5.1 item 2 is the safeguard; is it enough?
  2. Should an internship project be one the mentor could do alone in a week with an assistant? If so, the intern’s value is in work the mentor would not otherwise get to, and the project choice should reflect that.
  3. What is the right project size now? A 2020 summer project may be a two-week project in 2026, which leaves room for a second project or a larger one.
  4. Is there published experience of internships or graduate supervision with AI tools, in pharmacometrics or adjacent fields? Not yet searched. If this project starts reading sources, it needs a references.qmd.
  5. Should the mentor’s share of the checking fall as the internship goes on, as it would for a junior associate gaining a track record? The validation record from Section 7 would show whether the intern’s checks can be trusted with less sampling.

9. Steps

  1. Draft the one-page expectations document from Section 5.
  2. Scope the next internship project against Section 5.2 item 1.
  3. Draft the validation plan for that project from Section 6.
  4. Set up the weekly record of mentor hours by touchpoint.
  5. Run the internship.
  6. Review against Section 7 and revise Sections 3 to 6.
Back to top