The AI employee everyone’s building stops one layer short.

There’s a good tutorial making the rounds on how to turn Claude Code into an “AI employee.” Give it a workspace. Give it memory. Give it a clear ticket, eyes to check its own work, a review standard, a schedule, and permission boundaries. It’s a genuinely useful map, and I’m not knocking it — I run every one of those pieces, every day.

But almost every version of this stops at the same place. It gives the AI employee exactly what you’d give a new hire. It never gives it what you’d give a licensed one.

Here is the whole difference in one line. An AI employee that drafts your landing page needs a workspace and a ticket. An AI employee that drafts a letter of medical necessity, a prior-authorization appeal, or a surgical bill needs something no prompt, no repo, and no schedule can provide: a named, licensed human who signs the output and owns the liability for it.

That is not a nice-to-have you bolt on later. It is the entire line between an AI that makes you faster and an AI that can legally do regulated work. Speed is the easy half and it’s commoditizing by the week — the model does the drafting now, and it does it well. What doesn’t commoditize is the accountable human at the seam, and the owned record that proves they were there.

I’ve been building this for a while, so let me be concrete about both sides. I run agents in production — including ones that keep my own live sites honest against the latest regulations, on a schedule, unattended, fixing drift before I’m awake. That’s the “employee” half, and it’s real. But anything that touches a clinical output does not ship until a licensed human puts their name on it. Not because the model isn’t good enough. Because a name on the work is what makes it real, not just fast.

The regulators are arriving at the same seam

Last week the FDA published its first framework for regulating generative-AI medical devices. Read past the summary and you find the same idea, in the government’s own words: competency evaluated the way clinicians are credentialed, explicit accountability for the people who build and deploy these systems, and for agentic AI, a hard requirement for human checkpoints before any irreversible or high-consequence action. The map the tutorials are drawing ends exactly where the regulation begins.

So if you’re building AI employees for anything that carries real consequences — health, money, law, safety — build the accountability layer first, not last. Give the AI a workspace and a ticket, yes. Then give it the one thing that turns a fast draft into work that’s allowed to exist: a licensed human who owns it, and a record that proves it.

That’s the layer I build. It’s the whole thesis under everything on this site.