What McDermott Said and What He Didn't
On March 13, ServiceNow CEO Bill McDermott told CNBC that AI agents could push unemployment for new college graduates into the "mid-30 percent range" within a few years. He said "so much of the work is going to be done by agents" and that "it's coming quicker than people anticipate."
The statement got amplified fast. The framing was doom: AI is coming for your career before it even starts. The data backdrop is real — the Federal Reserve Bank of New York put recent college graduate underemployment at 42.5% at the end of 2025, already the highest since 2020. Adding AI agent automation to that picture produces alarming numbers, and McDermott provided them without much friction.
What McDermott didn't specify — and this is the part that matters if you're trying to understand what's actually happening rather than generate anxiety — is what kind of work the agents are replacing. That's the whole question. Not whether automation is happening, but which functions get automated and which don't.
The Work That Disappears First
McDermott told CNBC that ServiceNow's tools have "already taken out 90% of the use cases that previously relied on humans in customer service." That's a specific, operational statement and it's worth taking seriously. Tier 1 and Tier 2 customer support — the work that follows scripts, accesses standard information, handles defined exception paths — is now largely automatable. That work was, historically, a significant entry point for new graduates in service organizations.
The same pattern holds across most coordination work. Entry-level functions that involve processing defined inputs, following documented procedures, formatting outputs, and routing decisions to senior people — that category of work is exactly what AI agents handle well. It requires process knowledge, not judgment. And process knowledge can be encoded.
The functions that are harder to automate are the ones where the work itself can't be fully defined in advance, where judgment about ambiguous situations is the actual value, where relationship context matters, and where the output is a decision rather than a deliverable. These are not entry-level characteristics. They develop with experience. Which means the career ladder logic has shifted: the rungs at the bottom are being removed, not just made less comfortable.
What 14 Years of Operations Shows You
I've been running productions — commercials, branded films, content operations — for 14 years. In that time I've watched multiple waves of tool-driven disruption touch every function: editing software that eliminated the need for offline editors as gatekeepers, digital cameras that changed the DP-to-PA ratio, cloud-based project management that reduced coordination overhead, and now AI models handling script breakdowns, shot lists, and first-draft briefs in minutes.
Each wave removed the function that was most purely mechanical. Each wave also raised the baseline skill expectation for everyone who stayed in the room. The PAs who survived and advanced weren't the ones doing the mechanical tasks faster than the machines. They were the ones who already knew more than their job title required.
The automation wave McDermott is describing follows the same dynamic at much higher speed and much broader scope. What's different now is that it's not one industry's coordination layer being disrupted — it's every knowledge-work coordination layer simultaneously. The junior analyst, the junior coordinator, the entry-level account manager: all affected in roughly the same way at the same time. That's why McDermott's projection sounds extreme but isn't technically incoherent.
What Survives and Why
Two things are reliably hard to automate, and both develop through doing rather than through training.
The first is contextual judgment: the ability to read a situation that doesn't fit the standard pattern and respond appropriately. A production that goes sideways on day one of a three-day shoot doesn't need a workflow. It needs someone who has been in enough broken situations to know which problems to solve first and which to manage around. AI agents are not good at this yet because contextual judgment requires a model of reality that goes beyond the data available in the immediate context. Experience builds that model. Time builds that model.
The second is relationship-based work — the kind where the value isn't the output but the trust and communication that produced it. Clients who trust you don't trust you because your deliverables are technically correct. They trust you because you've demonstrated judgment, consistency, and honesty across interactions, including the difficult ones. That can't be automated because it's not a function. It's a history.
McDermott's warning is real for people who entered their field as a function. It is less real for people building expertise and relationships, even at the early stages of their career. The question for anyone starting out is whether the work they're doing is teaching them judgment or just teaching them the procedure.
The Operator's Actual Response
I'm not interested in the doom framing because it doesn't change the practical question. The practical question is: what should someone who creates, produces, or operates a content business actually do in response to what McDermott described?
The answer is direct. You automate the coordination in your own operation before someone else's automation makes your coordination skills irrelevant. You identify the parts of your workflow that are process rather than judgment, build systems to handle them, and use the recovered time to go deeper on the parts that require judgment and relationship. You increase the ratio of meaningful work in your day — not as a productivity exercise, but as a strategic positioning move.
The operators who will have leverage in the environment McDermott is describing are the ones who understand deeply what they're automating and why, can direct AI agents with precision because they know the work, and are adding complexity and relationships at a rate that outpaces what automation can handle. That's the same profile as a skilled operator at any point in the last 30 years of production. What's different is the timeline. It has compressed significantly.
McDermott said it's coming quicker than people anticipate. On that specific point, he's right.
Sources: CNBC — AI agents could easily send college grad unemployment over 30%, ServiceNow CEO says, March 13, 2026 | CNBC Video — ServiceNow CEO Bill McDermott on unemployment and AI agents | Seeking Alpha — ServiceNow CEO says AI could push jobless rate into 30% range
