We Cancelled the Farm System

I read a hiring manager say they were "pausing" entry-level openings because the agents could already do the starter work.

They wrote it like a win. Cycle time. Cost. Headcount they did not have to open.

I run a shop with a lot of agents. I want the boring substrate gone as much as anyone. Pulling the ticket. Cleaning the data. Drafting the first version of the status note. That is tax. Hand it off.

What I read was not a pause. It was a farm-system cancellation with better branding.

The ladder was never the job title

The white-collar career model had a quiet engine. You hired someone early. They did the codified work next to people who did the messy work. They watched how a senior person framed a bad option. They sat in the exception queue. They learned why the "obvious" call from last quarter was not obvious. Then, slowly, they stopped needing the script.

Most of that learning was never written down. It was tacit. How to smell a bad option. When to ignore the clean answer. Which customer gets the exception and which one does not. You get that by reps next to people who already have it. You do not get it from a wiki. You can write the procedure. You cannot write the context that made the expert break the procedure this time.

That engine was never elegant. Status chasing. Rework. Handoffs nobody mapped. Plenty of that deserved to die.

The mistake is treating the developmental part of the starter work as if it were the same waste.

A job is a bundle of tasks. AI does not eat jobs. It eats tasks inside jobs. The tasks it eats first are exactly the ones we used to give people so they could get good. When those disappear and nothing replaces them, you do not have a leaner org chart. You have a missing decade.

Teaching is the keep item organizations drop first, because it never shows up in the efficiency metric. This is that point with a budget attached.

The curse of knowledge

Once you know how to do the hard call, you forget what it felt like not to know. That is the curse of knowledge. You get worse at guessing where a beginner actually starts.

You look at the starter work and see only tax. You do not see the reps. You assume a new person can jump to review and judgment because you can. You cannot remember the months it took you to get there.

So we automate the "easy" layer. Seniors say they will teach by reviewing the agent. It sounds efficient. It is the curse talking. Review is not apprenticeship if the reviewer already has the tacit map and the junior does not. Watching someone reject a clean recommendation is not the same as knowing which signal mattered, what history changed the call, and which alternative almost won.

We undervalue tacit knowledge because it does not show up in a dashboard. You cannot put "knows when the clean answer is the wrong move" in a ticket. So we protect the explicit work, automate the rest, and act surprised when the bench is empty.

Faster, and then hollow

The pattern looks efficient for about two quarters.

Senior people get faster. Agents handle the substrate. The review queue still has names on it. The board deck still gets written. Then someone asks who is ready to own the next hard call, and the room gets quiet.

Experienced people are fine. The tools complement what they already know. The early-career people are the ones who lost the reps.

Stanford's Digital Economy Lab looked at high-frequency ADP payroll data covering millions of U.S. workers. No broad AI employment collapse. Workers ages 22 to 25 in AI-exposed occupations sat about 19% below where they would have been if they had kept pace with peers in less-exposed work. Experienced workers in those same occupations showed no comparable gap. Mostly fewer junior hires, not a spike in layoffs. Strongest where AI replaced tasks instead of helping people do more valuable work. Early signal. Still enough to take seriously.

The technology rewards the people who already finished the apprenticeship.

Klarna ran the customer version of this in public. In February 2024 their AI assistant handled 2.3 million conversations in a month, two-thirds of service chats, workload equivalent to 700 full-time agents. Faster resolution. Fewer repeat inquiries. That "700" was workload math, not a layoff announcement. In May 2025 the CEO said cost had become too dominant and quality dropped. They started hiring humans for customer service again while keeping AI in the mix.

Throughput is not the whole system.

The internal version is quieter. You do not get a public reversal. You get a bench with nobody on it in 2032.

Practice is not a mentoring program

When I say replace the ladder, I do not mean a lunch-and-learn series and a Slack channel called #growth.

Apprenticeship is a step in the work. If you cannot point to the step, you do not have one. That step has to transfer tacit knowledge on purpose, because the curse will keep telling you the junior should already know.

Exceptions before fluency. Sit the new person in the messy cases with someone good, not in the happy path the agent already owns. The agent can pull the account history. The human still has to decide when the right answer is the wrong move with this customer.

Problem statement before the draft. Make them write what the work is actually about before they see the model's version. First rep stays human. Second pass can be assisted. Reverse that and you get fluent output from people who never formed a view.

Explain the call out loud. After a decision, have them say why this option and not the other three. Not a performance review. A practice rep. If they cannot explain it, they did not make it. They approved it.

Protect the recovered time. If the hours come back from automation and you fill them with more tickets, you sped up volume. Put some of that time into judgment, teaching, the design review that kept slipping. Otherwise the organization got cheaper and the people got worse.

Give responsibility while expertise is still nearby. Bounded decisions with someone close enough to question the reasoning. That is how tacit knowledge turns into judgment.

I will take a slower workflow if someone still learns why the exception exists.

The readiness test

Before we automate another starter task, one sentence on the whiteboard:

What is the new apprenticeship, and which step in this workflow is it?

If the answer is "they will learn by reviewing the agent," that is not an apprenticeship. That is botsitting with a career path painted on it. Review that is theater looks busy. Review that is the work looks like framing, judgment, taste, a name on the line, and teaching someone else what just happened.

If you cannot describe the new apprenticeship, you are not ready to automate the old one. You are probably under the curse of knowledge, and you are about to undervalue the tacit part again.

I still want the agents. I still want the tax gone. I am not interested in an operating model that is brilliant at this year's throughput and empty when we need next decade's judgment.

We did not cut the junior roles.

We cancelled the farm system.

Where in your workflows does a new person still get the reps that turn them into someone you would trust with the hard call?