When SeatGeek launched ChatGPT company-wide, IT got 300 requests on day one, per Serval's case study. By day three, 60% of the company had filed a ticket. The tool was the easy part, and the tickets were the job. Every new AI tool arrives with licenses, access requests, approvals and a slightly different workflow for each team. The same case study counts 406 Claude licenses provisioned with zero human touches. My guess is adoption creates operational work faster than most teams can hire for it. Automation Engineer sounds like one more hot AI job title. Prompt engineer paid six figures at the peak of the hype, and by May 2025 Fortune was calling it obsolete. I think the difference is what the knowledge is about. Prompt engineering was knowledge about a model, and the models absorbed it. This role runs on knowledge about a company. Who approves access, which step legal refuses to skip, where the spreadsheet everyone depends on actually lives. A better model doesn't learn that from a release note. Building keeps getting cheaper. Serval describes a new automation as a sentence and an afternoon of work. The expensive part is knowing which process to automate and getting a team with real politics to use it. For PMs, it looks a lot like discovery with coworkers as the users. Every automation starts as a few lines of code and ends as a negotiation over who gets to approve it.