You have a training budget. It was never where most of the learning came from.
Almost everything your senior people know arrived while they were doing work that had to be done regardless. Nobody expensed it. It came attached to the task, as a byproduct, at no charge. The arrangement is old enough that most organizations have never had occasion to name it, which means they have certainly never funded it.
Agents are remarkably good at removing the middle of a task and leaving both ends where they were. The brief survives: someone still has to say what is wanted, precisely enough to act on. The artifact survives: someone still receives the draft, the plan, the proposed change. The stretch in between is what compresses, and on your books that stretch appeared only as hours. An unbilled input producing an unrecorded asset. Take it out and every ledger you keep reports a straight gain.
That is the awkward property of this particular loss. No instrument registers it. Output holds or rises, cycle time falls, nothing on the dashboard turns amber. The bill lands years later, in the form of a team that can competently approve an answer and cannot readily produce a second one.
One clue about where understanding still forms comes from a field experiment inside a large marketplace's customer-service operation, where researchers followed workers taking over chats an AI system had been running. When the worker stepped in early, on her own initiative, she went on to write about 60 percent of what remained of the conversation. When the takeover waited for the system to flag the customer's frustration, that fell to 43 percent, and the outcomes were worse. Escalation timing wasn't assigned, so this is association rather than proof, and participation is not learning; I am inferring exposure from how much of a conversation a person actually touched. But where you sit in a process governs how much of the process you can see. I have written elsewhere about what a late handoff costs the customer. This is the other half: what it costs the person doing the handing.
Most agents in production run a stretch of work and pass it to a human. Exposure has quietly rearranged itself. It used to be distributed across the work. It now collects at the seams.
So team composition over the next few years becomes a question of placement. Where do you put people so they still meet a range of possible answers rather than one finished candidate?
Part of the answer is cheap:
- Framing a problem well is still entirely human and thick with judgment. Hand it deliberately to people early in their careers instead of guarding it as a senior privilege.
- Leave a few stretches of work un-automated, chosen for what they teach rather than for how hard they are.
- Now and then, ask a reviewer to produce her own answer before she sees the generated one.
The rest has a price. A parallel human attempt at work an agent could have finished alone, paid for twice. Cycle time you could have banked and chose to spend instead. And a junior tier you have to defend in front of a finance committee on a case that returns nothing for three years, measured in units nobody has agreed on. No one is promoted for holding that line, which is a fair description of why the line rarely holds.
The two horizons behave differently. Inside eighteen months you will see intuition about edge cases thinning, approvals that someone with scar tissue would have stopped to question. That is recoverable. At five years it turns compositional. Stop hiring the tier that used to do the middle, let your seniors retire off a bench that no longer refills behind them, and you have made a decision that takes a decade to undo. Tooling reverses in a quarter. Team shape does not.
Nobody circulates a memo when a subsidy lapses. You simply notice, one quarter, that you are paying full price.