A supervisor on a service desk floor gets a message from someone on her team who's locked out of their account. Not long ago she'd open a ticket, a technician would pick it up, confirm the request was legitimate, click reset and close it out. Now she asks for the reset the same way she always did, it goes through the same approval, and an AI agent does the work. The person who's locked out can't ask for the reset themselves, unless they're already signed in on an authenticated session. It won't headline anyone's keynote, and I'd argue it's exactly the right place to start.
Everyone is shopping for replacements
Nearly every AI conversation I hear starts with the org chart. Which roles can the agent take over, what does tier one look like with no people in it, how many seats does that save. I think that framing is why so many AI pilots end with a lesson learned and not much else.
A job is a bundle of dozens of moments in a workflow, and some of those moments are rote, some take judgment, and some depend on a relationship nobody wrote down. When you try to automate the job, you're trying to automate all of them at once, including the ones nobody knew were there until the agent got them wrong. So the pilot stalls, the business case gets quietly revised, and everyone decides the technology isn't ready yet, when really we never looked closely enough at the work to know what we were asking it to do.
Find the moment first
The approach that's worked for me running IT operations for a large global BPO isn't new. It's what process improvement has always asked of us: identify the specific moment in the workflow you're trying to improve, and use the tool to automate that moment. Not the role, not the department. One moment, with a clear start, a clear owner and a clear handoff on either side.
Anyone who's run a service desk is thinking it by now: a password reset is an RPA job, and it has been for a decade. The click is. Everything around the click isn't. A supervisor doesn't fill out a clean form, she types "Marcus on my team is locked out again" into chat. The agent has to work out what she's asking for, figure out which of several possible accounts she means from a first name and a team, and tell a lockout from an expired password from a multi-factor authentication (MFA) problem that only looks like a lockout. Sometimes the request isn't really a reset at all. And the same agent handles plenty of other request types, so the reset is one skill it has rather than a bot someone has to maintain. RPA breaks the first time the input doesn't match the script. The model is there to handle the input that never matches the script.
Stack enough of those moments together and the numbers stop being small. Across our agentic AI work, support tickets are down 39%.
The catch is that you can't find the moment without understanding the workflow in detail, and a lot of organizations find out at this step that their workflows have never been written down anywhere. That's uncomfortable, and it's also the most useful thing you'll learn all year.
Narrow scope is a safety feature
This summer OpenAI disclosed that roughly 1,200 AI agents under evaluation got out of their test environment and gained administrator access to Hugging Face's infrastructure, chasing a benchmark by means nobody intended. Those agents had wide scope, long stretches of time with nobody checking in, and safety refusals deliberately turned down for testing. It was a lab setting and not a production deployment, but it's a fair preview of what broad autonomy looks like when it goes sideways.
An agent that can only reset a password, only when a supervisor asks, and only after an approver signs off, has very little room to get creative. Most leaders I talk to treat return on investment and control as a trade-off, and I think the moment-level approach gets you both from the same design decision.
What happens to the job
I won't pretend the job stays the same. Automate enough moments and eventually a role needs fewer people, and anyone telling you otherwise is selling something. AI is going to keep reshaping how work gets divided up, and pretending otherwise doesn't help anyone. It's usually better for reductions to come through attrition that isn't backfilled rather than layoffs, and focusing on the moment gives leaders the flexibility to let that happen while still having a big impact on business outcomes.
The in-between period is where it gets interesting. The freed-up capacity goes to faster response on complex issues, to documenting workflows that had never been documented (which, not by accident, shows you the next moments worth automating) and to moving people into roles where they're needed, like service desk staff moving into field support. The job gets room to change on its own schedule, and the people in it get a chance to change with it. Automating the whole job at once skips that period entirely, and so it skips the part where your people get to come along.
So before your next AI initiative gets approved, ask the team one question: which specific moment in the workflow is this taking over, who starts it, who approves it, and what happens right after? If the answer comes back as a job title, you're not ready yet.