Most conversations about "AI in healthcare operations" start with a single, oversized question: what can we get the AI to do? A more useful question, borrowed from how well-built AI systems are actually structured, is: how should the work be divided, and where does a person need to step in?
A recent open-source project called arkan-blog-agents — a teaching example built for an AI course, not a product — makes this concrete. It's a system that writes and publishes content, but instead of one AI model doing everything from one giant prompt, the work is split across specialists: one role generates ideas, another turns the strongest idea into a brief, another researches supporting facts, another writes the draft, another edits it against a fixed set of criteria, another checks technical details a program can verify without judgment at all, and a final role decides whether the result is ready to publish or needs a person to look at it first.
None of this is really about content writing. It's about workflow design, and the same three lessons show up in any clinic that is trying to use AI responsibly.
1. One job per step, not one model for everything
When a single prompt is asked to brainstorm, research, write, and judge its own work all at once, the quality of each part suffers — the system is switching between incompatible modes of thinking without any clear boundary between them. Splitting "generate" from "critique" is a small structural change with an outsized effect on reliability, because the reviewing role is never the same one that produced the draft.
The clinic parallel is direct. A workflow where the same untracked step is expected to both "flag the issue" and "fix the issue" tends to quietly drop things — a missed follow-up, a form that never gets checked, a task that everyone assumes someone else is doing. Naming who (or what) does the first pass and who does the check is often more valuable than any new tool layered on top.
2. Decide in advance what a person must verify
In the Arkan system, an editor scores each draft against a rubric. Above a set threshold, the piece moves forward automatically. Below it, it's held for a person to review before anything goes out. The line isn't drawn by hoping a busy person remembers to double-check things — it's drawn once, as a rule, before the system ever runs.
That's the version of "AI-supported, not AI-driven" we build toward with clinics: a workflow can use AI for the first draft of a follow-up message, a summary, or a scheduling suggestion, but the point where a human confirms it before it reaches a patient is decided in advance, not left to whoever happens to be at the desk that day.
3. Information needs to carry where it came from
The most instructive part of the project's own build notes is a mistake it documents honestly: in an early version, a step generated a plausible-sounding statistic as part of drafting social content, and a later step treated that number as if it were a real, human-reported observation — because nothing in the data distinguished "a person told us this really happened" from "the model produced this as a plausible example." The fix wasn't a smarter model. It was a flag traveling alongside the data itself, marking whether it was verified or generated, so every downstream step could treat it accordingly.
Clinics run on the same risk in smaller form. A patient-reported symptom, a staff observation, and an AI-generated draft summary are three different kinds of information, and a workflow that lets them blur together — in a chart note, a follow-up message, or a report — creates exactly the kind of quiet error that's hard to catch after the fact. Keeping provenance attached to information, not just the information itself, is a habit worth building before AI is involved, and it becomes essential once it is.
The takeaway isn't "build a multi-agent system"
Most clinics don't need eight specialized AI roles handing work to each other. What's worth borrowing is the underlying discipline: break the work into steps with a clear owner, decide ahead of time exactly where a human needs to confirm something before it moves forward, and keep track of where each piece of information actually came from. That's less about the technology and more about how the workflow around it is designed — which is the part that usually gets skipped when a clinic adopts a new tool in a hurry.
It's also the starting point for most of the work we do together: mapping where a clinic's operations already look like this, and where a small structural change — not a bigger tool — closes the gap.