The junior work is where seniors come from

Hand a set of accounts to someone in their third year and ask them what looks wrong with it.

The pause is the thing. It is a long, careful pause from someone who is trying hard, who is not lazy and not stupid and who has passed every exam you passed. Eventually they say the depreciation looks a bit high. It isn't. They're guessing, politely, because nothing in the file is shouting at them.

Ten years ago that pause was about four seconds long and the answer was usually right. Not because the staff were better. Because whoever you handed it to had, by then, personally coded some eleven thousand bank transactions and knew in their hands what a plumbing business is meant to look like on a page.

The apprenticeship was a by-product, and by-products don't survive efficiency

Casting a schedule taught you that totals are usually right and the interesting one is the exception. Coding twelve months of a bank account taught you what a business's normal looks like, so that the month someone bought a boat stood out without you going looking for it. Chasing a missing statement taught you that clients aren't withholding documents to spite you. They're busy, and unclear about which account you mean.

None of it was on a training plan. It was the price of admission, paid in hours, and it produced people who could open a file and feel that something was off before they could say why.

Then we got faster. Some of the work went offshore, most of the rest went into software, and the price of admission stopped being charged.

The output is better for it. Bank feeds beat manual coding on accuracy and it isn't close, and a machine-cast schedule beats a tired human one every time. What went wrong is that we removed the training without noticing we had removed anything, because on paper what we removed was data entry.

"They can use the software" is not the same as knowing what right looks like

It doesn't hold, and it's worth being precise about why. Exception handling is the hardest job in the practice, not the easiest. Deciding whether a variance matters requires knowing what a non-variance looks like, and there's no way to acquire that except exposure to a great many boring, correct files.

So what you get is someone quick (quicker than you at the software, certainly) who can move a case end to end, and who cannot yet tell you which of the twelve numbers on the screen is the one to worry about. What they are missing is a dataset.

And it surfaces late. Not in their first year, when everything they touch is checked anyway. It surfaces in year four, when you want to stop checking and discover you can't.

What the boring work was actually teaching

What normal looks like. A distribution you end up carrying around, too loose to write down as a rule. Wages as a rough share of turnover for a trade business. How many bank accounts a family trust usually has. What a café's card takings do over a summer. You can't run a session on this. You can only expose someone to enough of it.

Where things hide. The loan account is the interesting page. The last statement of the year is where the odd transaction sits. A prior-year comparative that's suspiciously round has usually been typed by somebody. These are patterns a preparer learns by being burned by them once, in March, at length.

That there's a person on the other end. Ringing a client for a document is where a junior first works out that the client isn't a queue item: they're at a job site, the request was confusing, and they'd send it today if they understood which account you meant. Chasing is the first thing anyone automates, so you can end up with a senior who has never had the conversation.

Teach it deliberately, because it won't happen by accident any more

The work is gone. The exposure doesn't have to be. Four things that seem to work, none of which cost much.

Review in reverse. Give a junior a finished file (a good one, prepared by a senior or built by software) and ask them for the three things they would query. No client waiting, no stakes. This aims straight at the gap: they've seen a hundred files they prepared and almost none that somebody else did. Use a file you've already signed off and you can then tell them exactly what you looked at and in what order, which is a thing nobody has ever explained to them.

Let them write the query, not just collect the answer. Chasing a document is an errand. Working out what's missing and why it matters is the job. The second one teaches; the first one fills a folder.

One judgement, defended, every week. Repairs or capital. Deductible or not. Something arguable out of a live file, with the requirement that they say which way they'd go and what would change their mind. Five minutes. You already know the answer. What's being built is the habit of holding a position in a domain where a great many things really are arguable.

Show them the ones that went badly. Every practice has them: the amendment, the thing found in year three, the schedule that reconciled beautifully to itself. It's the highest-value teaching material in the building and it's locked in two partners' memories, because nobody enjoys retelling it.

We're part of this, so here's the honest version

None of that is reversible, and it shouldn't be. It would be dishonest to sell it as pure gain without saying that the by-product goes out with it.

What software can do about it, and what we've tried to do, is leave the reasoning visible instead of hiding it inside a result. A file that shows which document each figure came from, and that raises the calls it wasn't sure about rather than settling them without saying so, is at least readable as an example: a junior can see how a case was thought through instead of only what it totalled. Call that what it is: better than a black box, and a long way short of eleven thousand bank transactions.

The practices that will still have seniors in ten years are the ones treating this as a training problem now, while the people it affects are in year two and it all still feels fairly theoretical.