Five minutes · A worked example
A junior's week
Anthropic's economists recently published scenarios for how AI might reshape the economy. Their model treats every job as a bundle of tasks. AI leaves some alone, helps with some, takes some over, and creates new ones. In their example of a nurse, the new tasks might be checking how well an AI triages patients, or reviewing an AI-proposed care plan.
That's a useful way to look at a job, and it leaves one question open. Reviewing work is a skill, and people learn it by doing the work first. So it matters which tasks you automate, not only how many.
Here is one week for a first-year associate in an accountancy practice, sorted the way most automation plans sort it, and then the way I would. The tasks are typical of the role. The hours are my own illustration.
The week: 40 hours
Outline: software does it · Faded: deleted · Hours: what stays with the associate
01 / 08
The week
Forty hours, twelve kinds of work. Some of it matters, some of it is habit, and most of it is how a new accountant learns what a good set of numbers looks like.
02 / 08
Sort it by value
This is where most automation plans stop. The top row is worth doing well and the bottom row is worth less. The plan automates the top, cuts the bottom and banks the saving.
03 / 08
Now sort it by what it taught
The same twelve tasks, split again by whether doing them taught the associate something they'll need in five years. Half the top row turns out to be how people learn the job, and so does half the bottom.
04 / 08
Automate it
Matching bank lines, keying data from PDFs, building lead schedules. They need to be right, and after the first month they teach very little. This is the saving to bank, in the same quarter.
05 / 08
Delete it
Reformatting files, re-keying the same numbers, status emails. No AI needed. Fix the template, connect the systems and stop sending the email. It's often the best box on the grid.
06 / 08
A person decides
Drafting working papers, explaining variances, reviewing a peer's file. This is where judgement is built. Let the software draft, keep the associate deciding, and have them defend the result to a senior. It's the same kind of review work Anthropic's model expects AI to create, and this box is where the people who can do it come from.
07 / 08
Keep a share
Sampling invoices, chasing documents, writing up client meeting notes. The output is worth little and the lesson is worth a lot. Automate most of the volume, keep a set share for people, and budget it as training.
08 / 08
What's left
Fifteen of the forty hours stay with the associate, and most of them are the ones that teach. Twenty-five come back, and how they're spent is for the business to decide.
What the second sort changes
Sorting by value alone would have automated all six tasks in the top row, including the three that teach the job. The saving looks bigger in the first year. The cost arrives later, when the associates who should have become reviewers never did the work that makes a reviewer.
Sorting by what the work taught gives up a little of that first-year saving and keeps the route by which people become experienced. It's a small decision in any one week. Across an intake, over five years, it decides who can check the software's work.
The reasoning is in The bottom was the training programme, and the grid is a tool you can run with a team in an afternoon in How I run it.
See the week as a table
| Task | Hours now | Worth | Taught | Box | Hours after |
|---|---|---|---|---|---|
| Drafting working papers | 6 | A lot | A lot | A person decides | 4 |
| Keying data from PDFs | 5 | A lot | Little | Automate it | 1 |
| Sampling invoices | 5 | Less | A lot | Keep a share | 2 |
| Matching bank lines | 4 | A lot | Little | Automate it | 1 |
| Explaining variances | 4 | A lot | A lot | A person decides | 3 |
| Reformatting files | 3 | Less | Little | Delete it | 0 |
| Lead schedules | 3 | A lot | Little | Automate it | 0 |
| Chasing documents | 3 | Less | A lot | Keep a share | 1 |
| Reviewing a peer's file | 2 | A lot | A lot | A person decides | 1 |
| Re-keying numbers | 2 | Less | Little | Delete it | 0 |
| Client meeting notes | 2 | Less | A lot | Keep a share | 2 |
| Status emails | 1 | Less | Little | Delete it | 0 |
| Total | 40 | 15 |
Where the example comes from
The tasks are typical of a first-year associate in an audit or accounts practice, and the hours are illustrative. Anthropic's model is Economic Scenarios for Transformative AI (Korinek et al., 2026), with an interactive explorer. Treating a job as a bundle of tasks goes back to Autor, Levy and Murnane (2003) and the work that followed. The grid is mine.