Everyone is afraid of AI making jobs disappear. Box founder Aaron Levie has a different take. In a clip this week he said the biggest new job title of the next five years doesn't exist today, and it's going to create a million roles. He calls it the agent operator.
Not the person using an agent. The person who walks into a marketing, legal, operations, or life sciences team and redesigns the workflow so an agent can run most of it. They're the interpreter between how work has been done and how it will be done.
The claim is easy to dismiss as another AI-future platitude. But his framing lands harder if you've actually tried to roll out agents inside a Fortune 1000 pharma, a bank, or a consultancy. Regulated. Data fragmented across fifteen systems. Employees wired to run a process a particular way for a decade. An agent can't inherit that. It has to be rebuilt for how an agent actually reads and acts.
The job is translation, not prompting
Most of the AI conversation is about the model. Which frontier model reasons best, which handles long context, which generates better code. But if you've tried to roll an agent into a function at a regulated enterprise, the model is almost never the bottleneck. The bottleneck is that an agent can't inherit a process written for humans.
A legal analyst flags a contract for reasons that aren't in the playbook. They're in her head from twelve years of reviewing deals. A marketer's attribution model is technically "last touch" but she knows to discount three specific channels because of a known spam problem nobody documented. A finance manager sanity-checks a number against a second source because a bad quarter in 2022 taught her the first source lies when volume spikes.
None of that is in the SOP. It's institutional knowledge. Someone has to translate it into explicit instructions, guardrails, data sources, and checks an agent can run. That translation is the job Levie is describing when he says "requires care and feeding and a real level of technical and business process acumen." It's not prompt engineering. It's business process redesign, with a model in the loop.
Why it's a million-job role
Every function in every large company has to be redone. Marketing needs campaign performance workflows. Legal needs contract triage. Operations needs postmortem drafters. Research needs literature monitors. None of these are generic. Each one requires someone who understands the function, the data, and the agent. Multiply by every function at every large company and you get Levie's million.
79% of enterprises face AI adoption challenges despite massive investment. Two-thirds of enterprise AI projects stall in pilot. The model is good enough. The people who can do the translation work are the bottleneck.
The role is already here
The title isn't on the org chart yet. The work is. People inside large companies are already redesigning workflows for agents, writing instructions, wiring up integrations, and proving to legal that nothing sensitive ever reaches the model. They're doing it with a mix of Python scripts, Zapier, and internal platforms stitched together by IT. In five years they'll have a title and a team. For now they're doing the job without the name.