Entry-level support just became a judgment job

PwC read more than a billion job ads across 27 countries for its 2026 Global AI Jobs Barometer, published in June. One finding is odd enough to stop on. Entry-level jobs exposed to AI are now seven times more likely to ask for traditionally senior skills. Judgment. Leadership. Talking to people face to face. Those roles grew 35% since 2019, while other entry-level roles fell 10%.

Pete Brown, who leads workforce at PwC, named the mechanism directly. AI is removing the routine work that used to serve as apprenticeship, and demand for judgment now arrives much earlier in a career.

Customer support is the clearest case of this anywhere in the economy. The tier-one queue was the apprenticeship. You learned the product by answering a few hundred easy questions about it, and somewhere in there you picked up what the company actually promises customers. That queue is the first thing an AI agent takes.

The industry is not responding by shrinking the job. Gartner surveyed 321 service and support leaders and found 85% adding new tasks and responsibilities to frontline roles, with 75% moving agents into higher-value work. Forrester’s Max Ball named the roles that come next: bot unblockers who supervise AI agents, judges and experts who take the cases that need a person, and consultants who sell.

So the forecast is settled. The support job becomes a judgment job.

Here is the part almost nobody has worked out. A judgment job needs something to judge.

Where is the judgment supposed to go?

Think about what supervising an AI agent means in practice with most support tools. You connect the agent, it starts answering customers, and the human gets a dashboard of what already happened. You can read transcripts after the fact. You can write a rule and hope it holds.

But the decisions that actually matter got made once, during setup, by whoever did the integration. What is this agent allowed to say. Which subjects should it handle at all. When should it stop and get a person. After that the frontline role is watching a screen.

That is not a judgment job. It is a monitoring job with a better title.

If the new role is going to be real, the product has to put real questions in front of a person. Not “was that answer any good,” asked after the customer already read it. Questions with a cost attached, asked before anything ships.

What that looks like when you build for it

This is the design problem we started with, so it is worth saying plainly how Celeste works.

Celeste learns from a business’s own support history and then practices in private. It drafts replies to real incoming messages and nobody sends them. A person reads the drafts and says whether each one is right. No customer sees anything during this.

Topics then go live one at a time, and a person decides which ones. Order status might graduate in the first week because the practice drafts were consistently correct. Billing disputes might wait until the owner has read months of them. That is a real decision, made by someone who knows the business, with consequences either way. That is what judgment means.

Underneath all of it, the limits live in our code rather than in the wording of a prompt. The owner sets what the agent may do. Celeste proposes an action, and the code decides whether it is permitted. An agent cannot talk its way past a boundary that is not sitting in the conversation with it.

Notice who those three things need. They need someone who can look at a draft reply and know whether it matches what this company actually promises. That is the tier-one veteran whose queue just got automated.

We also think reading practice drafts is a better apprenticeship than the queue it replaced. Work through a hundred drafts against a hundred real customer messages and you are looking straight at the business’s policy, one decision at a time, with the reasoning attached. The older version of that education was several hundred password resets and a lot of waiting.

Who the savings belong to

One more thing follows from this, and it is why we set the company up the way we did.

The person supervising the agent is teaching it to handle work they used to do themselves. When that goes well, the business saves money, and the saving traces straight back to their experience. Some of it belongs to them. That is The Dividend Standard, our commitment to direct 70% of profits to the workers whose jobs AI automates.

PwC put the wage premium for AI skills at 62% this year, up from 57%. That premium is real money and it is going somewhere. Support veterans learning to supervise agents have a claim on it that is easy to state out loud and very easy to quietly skip.

If you want to see what the judgment seat actually looks like, how it works walks through practice mode and topic graduation.