The new AI support jobs need a decision attached

Intercom’s 2026 Customer Service Transformation Report found that 45 percent of support teams have already rewritten job descriptions to include AI responsibilities, and 40 percent say their people now spend more of the week training AI systems. Microsoft’s Work Trend Index, published in May, found 28 percent of managers weighing whether to hire an “AI workforce manager”, someone to run a team of humans and agents together. Another 32 percent expect to hire AI agent specialists inside 18 months.

The job is not hypothetical either. Search “customer service AI trainer” on the job boards this week and the listings are there, averaging around $52,000 a year in the US.

So the title is arriving. That is genuinely good news for a function that has spent two years reading grim headlines about itself. What has not arrived alongside it is authority, and without that this new rung will not hold anyone’s weight.

What most AI supervisors can actually do today

Look at the levers the tooling hands the person in that seat.

They can read dashboards. Containment rate, CSAT, deflection, escalation counts. They can pull transcripts and read them. They can edit a prompt or add a knowledge base article and hope the behavior moves in the direction they wanted.

Now notice what is absent. None of that is a decision about what the AI is permitted to do. That got settled earlier, by whoever configured the deployment, and in most products it is a single setting: the agent is answering customers, or it is not.

So the supervisor’s actual daily job becomes noticing and reporting. Watch the numbers, flag the bad ones, pass them to someone with deeper access. That is a monitoring job. Monitoring is precisely the sort of work that gets automated in the next cycle, which means a role built this way is the next role on the list. We would be teaching a lot of capable support people a skill with a short shelf life.

Microsoft’s own framing points somewhere better. The report describes human agency shifting toward intent-setting, judgement, orchestration and accountability. Accountability is the word carrying the load in that sentence. Nobody can be held accountable for a system they have no power to change.

The decision that belongs to that seat

Here is the call I think should sit with that person: which topics the AI handles on its own, and when each one earns the right.

That is real work. It takes judgment about the business, a read on where the risk actually lives, and knowledge of which customers get upset about what. A script cannot do it, because it is a judgment about whether the evidence is good enough yet rather than a threshold to compare against. And the outcome lands on the person who made the call, which is what makes it a job rather than a task.

Two things have to be true of the software before that decision can exist at all.

The AI has to be able to work without acting. Celeste learns from a company’s own support history, then works the live inbox writing replies it has no ability to send. The agent builds a record of what it would have said, sitting next to what the team actually said, on real customer mail. That record is the evidence. Without it, anyone approving autonomy is guessing and calling it judgment.

Autonomy also has to be grantable in pieces. Shipping questions can go live while billing stays in practice. Refunds can stay in practice for as long as the owner wants. The person in the seat reads the practice record for one topic, decides it has earned the right, and graduates that topic. Everything else carries on exactly as before.

Build it that way and the role has a spine. Something to decide, evidence to decide it on, and a visible consequence. That is a job somebody can be promoted out of, which is the honest test of whether a rung is real.

Past the org chart

We are loud about where we think the gains from AI should go. There is a nearer-term version of that argument, though, and it is not about fairness at all.

The people answering a company’s mail today know its customers better than anyone else in the building. They are the ones who can tell whether a draft reply is actually correct rather than merely plausible. Giving them the graduation decision is not a gesture. It is the most accurate way to run the system, because they hold the information the decision requires.

The job titles are being written this year. Whether they become careers depends on whether anyone bothers to attach a decision to them.

If you want the mechanics, how Celeste earns autonomy is the whole design.