This piece is about the least glamorous question in the AI agent boom: when an agent does the work, who answers for it? Researchers at Accenture and Wharton have both converged this year on the same finding, now circulating as “the supervision gap.” Agents are being granted real authority inside companies, to draft, to file, to spend, to reply, faster than anyone is assigning a named human who owns the output. In most organizations, ownership is described, in the research’s own language, as informal or undefined.
I run an AI staffing agency. We place managed AI agents into client organizations the way a temp agency places people, and the supervision gap is the reason the company exists. So read what follows knowing I have a position. I will label the opinions as opinions and show you the market evidence either way.
The market sells the worker and hands you the management burden
Look at what is actually for sale in mid-2026. Motion raised a $60M Series C last December at a $550M valuation selling “AI Employees That 10x Your Team’s Output,” prebuilt roles across sales, support, and marketing. Lindy sells the “AI Employee” and reported 5,000+ customers in May. Cognition’s Devin, the coding version, was valued at $26 billion the same month. Artisan cut its AI BDR to $250 a month in May and went self-serve. These are real products with real traction, and each ships the same fine print: you supervise it. The buyer gets a worker and inherits a management job nobody scoped, staffed, or priced.
That is the gap in commercial form. The vendors sell capability. Accountability is left as an exercise for the customer, which is how a mid-market company ends up with eleven agents holding send permissions and no org chart line that says who reviews what they send.
Two data points tell me the industry knows this is the weak joint. First, Harvey, the $11 billion legal AI platform doing roughly $300 million in annual revenue by May, spent July shipping multi-step agent workflows wrapped in a human review-and-approve pattern. When the category’s most credible player starts building supervision into the product surface, the market concedes that unsupervised agents do not meet professional standards. Second, the cautionary tale already exists. 11x sold “digital workers” harder than anyone, hit a reported $350M valuation, and then came apart in public: customers it listed had churned or never signed, reported churn hit 70 to 90% at three months, and the founder was out as CEO by May 2025. My read of that collapse, and it is a read, not a court finding, is that they claimed the labor frame without building the thing labor requires. If you call it an employee, someone has to manage it. Nobody did.
What a staffing agency does that a software vendor structurally cannot
A temp agency does not mail you a person and disappear. It screens, trains, places, supervises, and, when the work is wrong, answers for it. That relationship is old, boring, and well understood by every buyer who has ever filled a role. We bet that this relationship, not another builder tool, is what the agent economy is missing.
In practice it means the agent arrives named and trained for the role. A management layer rides along: logging what it did, escalating what it should not decide alone, and reporting to you in the cadence a manager would. And the agency stands behind the output, which is a sentence no tool vendor in the list above can say, because their terms of service exist to say the opposite.
One distinction worth being precise about, since the words collide. Anthropic, whose models we build on, ships a product called Managed Agents: a hosted runtime that manages agent execution, the infrastructure layer. We manage the work: the role definition, the supervision, the accountability to the client. Their layer makes ours more reliable. Different floors of the same building.
The closest thing to our thesis on the market today is Maisa, which sells auditable “digital workers” with full traceability in regulated industries, and earned a Gartner nod this year. Auditability is a real step. It is also, in my view, the halfway house: an audit trail tells you what the agent did. It still does not name the party who answers for it. The last unoccupied position in this market is the one where someone stands behind the work.
Current view, subject to change
My current view is that the supervision gap is the buying trigger for the second phase of agent adoption. Phase one buyers were enthusiasts who enjoyed the management burden. Phase two buyers are operators who want the role filled and the risk owned, the same people who use staffing agencies for human roles and for the same reasons. That is why we describe Coworkers.Global as an AI staffing agency rather than an agent platform, and why supervision is the product, not a feature.
What would change my mind: if the platforms ship supervision credibly enough that buyers stop wanting a counterparty, the way payroll software absorbed a chunk of what bookkeepers once did. Harvey’s review-and-approve pattern is the early version of that move, and I take it seriously. Or if agent reliability improves so fast that supervision stops being worth paying for at all. I think both are years out, and the second may be never, for the same reason companies with excellent employees still have managers.
Final thoughts
Every previous wave of workplace automation eventually produced an accountability structure: the machine got an operator, the ledger got an auditor, the fleet got a dispatcher. The agent wave is producing its structures out of order. Capability arrived first, spectacularly. The operator role is being defined right now, mostly by whoever notices the gap.
The research community has now measured the gap and named it. The incumbents are edging toward it one approval button at a time. Somebody will be the accountability layer of the agent economy. Our answer is the oldest structure in the labor market, updated: an agency that places the worker and stands behind the work.
Regards,
Charles Stack