Written by Joe Phillips
Everyone is saying “AI Employee” right now. Almost nobody can tell you what it means, and when you ask whether it’s the same thing as an AI agent, the answer usually depends on what the person is selling.
A few months ago I decided to settle it for myself. After more than two decades building software I sat down to write a definition, expecting an afternoon of work and a paragraph at the end of it.
I got a rabbit hole instead. The paragraph needed a test, the test needed evidence, the evidence needed someone accountable for producing it, and that person needed authority they could actually exercise. Every answer opened a door I had not seen. What came out, months later, was the Hybrid Workforce Standard: eight bodies of knowledge, twenty-seven clauses and nine properties, all of it built to answer one question precisely: which AI agents can correctly be called AI Employees.
An agent executes. An employee stewards.
Capability is not the axis. A very sophisticated agent that reasons, calls tools and chains a dozen steps together without help is still an agent, and buying a more powerful one will not move it across the line.
An agent executes a task: “send these twenty follow-ups.” An AI Employee holds a role, which means sustaining a recurring process inside defined limits. Give it an SDR role and it has to find leads, research them, make contact, follow up, record what happened, escalate what it can’t handle, and report how it went. Nobody’s prompting it each morning.
Stewardship stops short of ownership, and that gap is where most deployments get sloppy: the position, its authority and its accountability still belong to a person. What the system carries is the continuing responsibility to sustain the work. What it can never carry is the consequence.
The nine properties
I needed a test a reviewer could run without arguing about definitions, so the test is conjunctive. All nine present in operation, or it is not an AI Employee.
Persistent identity. A defined operational role with results and exclusions. Organizational context. Authorized tools and channels. Autonomy, meaning it can start or continue without a human prompt at every step. Limited authority, with thresholds and prohibited actions written down before it runs. Governed memory, with provenance and retention. Observability: actions, tool calls, costs, decisions and results all traceable. And a named human who answers for configuration, controls, performance and exceptions, however many artificial supervisors sit in between.
Presence is binary. A deployment missing any one of them can still be excellent; calling it an AI Employee is then a commercial metaphor rather than a verifiable administrative category.
The part that surprised me
The test runs in both directions. I did not go looking for that. If a deployment exhibits all nine properties in operation, it is an AI Employee whatever your vendor calls it, and the burden of demonstrating otherwise falls on whoever deployed it.
I found that out on my own system. In August I had an AI assistant build a media-inquiry pipeline for my company: it reads each press request, scores it against my expertise, and drafts a reply for human approval. I built it as a tool. Months later I ran my own test against it and it qualified on all nine. That was not a compliment to my engineering. It meant the thing had been carrying obligations I had never written down, and the only reason nothing went wrong is that I am the sole person who can send anything it drafts.
Most organizations are in that position and don’t know it. The category arrived after the systems did, which is nobody’s fault and everybody’s problem.
What to do with this
Take the system you already run, the one nobody wants to touch, and walk the nine properties against it. Write yes or no for each, honestly.
Nine yeses means you are governing an AI Employee whether or not anyone approved that decision, and you should name who answers for it before you need to know. Eight means the missing one is usually accountability or limited authority, and both are cheap to fix on paper and expensive to fix after an incident.
The AI executes. The organization answers. A human governs. I spent months arriving at a sentence I could have written on day one, and the months are what made it true.
Author Bio:
Master Joe Phillips is the author of the Hybrid Workforce Standard and of AI Employee, a book on designing and governing AI-held roles inside organizations. He has founded several software companies in Costa Rica and works with executive teams on AI governance and implementation. He is based in San José, Costa Rica.