Everyone in Medical Affairs is using AI now. Almost no one is using it for real work.
The dominant pattern in pharma teams today looks like this: open ChatGPT, paste in a meeting summary, ask for bullet points, copy the result back into the slide. That’s not AI adoption — that’s a slightly better spell-check.
Real adoption is agents doing the work, not chat producing text. Agents that read the event PDFs, extract the speakers, match them against your KOL list, draft the follow-up emails, populate the tracker, and surface only what actually needs human judgment. Text output is the cheap part of AI. Doing the work is the hard part — and it lives one layer deeper than where most teams are looking.
After two years of building, breaking, and shipping AI workflows inside Medical Affairs, I’ve stopped chasing tools and started thinking in stacks. A stack is what separates “we have ChatGPT licenses” from “we have agents that ship work.” A stack is what holds when conference season hits, when compliance asks the hard question, when the model gets swapped out for a better one.
This is the field guide I wish I’d had when I started.
Layer 1 — Foundation: How Your Information Is Encoded
Before you talk about agents, talk about substrate. The single most underrated decision in any AI workflow is what format your information lives in.
PDFs and PowerPoints are where institutional knowledge goes to die. Markdown — plain, version-controlled, machine-readable — is where it stays alive. It travels between humans and models without friction. It diffs cleanly. It’s the closest thing pharma has to a lingua franca that both compliance and code can sign off on.
I’ve written more about this in Markdown as the Operating Layer for Medical Affairs AI. If you’re trying to introduce AI tools into a team that still emails decks back and forth, that’s the place to start. Get the substrate right and everything above it gets easier. Skip it and every layer above will leak.
Layer 2 — Governance: Making AI Audit-Proof
Pharma doesn’t get to skip compliance. Every AI workflow in Medical Affairs has to survive a question from someone wearing a “Quality” badge. That’s not a brake on innovation — it’s the bar you build to.
Three pieces sit at this layer:
- Structured prompts as SOPs. Treat your prompts like SOPs, because functionally they are. Version them. Review them. Approve them. I unpack this in SOPs for Agentic AI Workflows.
- Auditable execution. Every agent run should leave a trace a human can read months later. Making AI Harnesses Audit-Proof in Pharma describes what that actually looks like in practice — and what regulators are starting to ask for.
- Industry alignment. What I learned listening to the AI governance conversations at ESMO 2025 is documented in ESMO 2025: AI Governance for Medical Affairs. The short version: the rules are coming, and the early movers are the ones writing them.
Skip this layer and your AI initiative becomes a Q3 audit finding instead of a competitive advantage.
Layer 3 — Execution: From Dashboards to Agents
This is the layer most people associate with “doing AI.” It’s not the most important — but it’s where the visible work happens, which is why it gets all the attention.
The progression I keep seeing in real Medical Affairs teams:
- Dashboards that visualize what already happened. Most teams stop here.
- Dashboards with intelligence that flag what actually matters. A small step that pays back fast — described in From Dashboards to Decision Engines.
- Agents that take the next action: draft the email, route the insight, populate the tracker. The case for moving here now is in Why Medical Affairs Needs to Embrace Agentic AI — Now.
Both jumps are technically easier than the foundation work in Layer 1. Both are useless without it.
Layer 4 — Documentation: The Hidden Failure Mode
There’s a problem nobody warns you about: as agents start generating output, your documentation problem doesn’t shrink. It explodes.
Suddenly there are ten times more artifacts — drafts, decisions, summaries, execution traces — and nobody knows which version is canonical. This is what I describe in The Documentation Crisis AI Agents Create. It’s not a minor bookkeeping issue. It’s the thing that quietly kills the second wave of AI adoption inside an organization, after the initial pilots look great in slides.
The teams that solve documentation early move faster than teams with better models.
Layer 5 — Strategy: The Operating Posture
The top of the stack isn’t a tool. It’s how you think about the work.
Two pieces matter here. First: stop treating your medical strategy like a slide deck you update once a year. Treat it as a living system that absorbs new evidence weekly. I make that case in Medical Affairs Strategy as a Living System.
Second: know when to start. Most teams either rush into AI without the substrate, or wait until “the technology matures” and arrive late. The middle path — what I call AI readiness — is laid out in AI Readiness in Pharma: Right Time, Right Tech.
And if you’re a Medical Affairs professional wondering what all of this means for your own career, there’s a separate conversation we need to have. I started it in The AI Layoff Trap in Medical Affairs. The short version: the people who lose are the ones who only touch AI as users, never as operators.
What to Do With This Map
If you’re starting from zero, work bottom-up. Not because it’s the prettiest path, but because it’s the only one that holds:
- Pick one workflow. Just one. The one that wastes the most time on your team this week.
- Move its information into Markdown. That’s your foundation.
- Write the SOP first, then the prompt. That’s your governance.
- Prototype the agent. Make it boring. Make it traceable.
- Document what changes. Especially what doesn’t work.
Don’t skip layers. Don’t optimize Layer 3 before Layer 1 holds. The teams I’ve seen succeed move slowly through the substrate and quickly through the surface. The teams that struggle do the opposite — and they all sound the same when it stops working.
The Bigger Picture
In two years, the question won’t be whether Medical Affairs teams use AI. They will. The question will be which teams built the operating substrate to use it well — and which teams have a graveyard of half-deployed pilots and a Slack channel full of frustrated MSLs.
The stack isn’t optional. It’s the thing.
This article was co-authored with Anthropic’s Claude Opus 4.7 model. The ideas, domain expertise, and editorial direction are mine — the AI helped structure, draft, and refine the text.
