Signal vs. Noise
Eight of Twenty-Five: AI Takes Tasks, Not Jobs
The company that cuts you to save eight tasks just lost the seventeen nobody was counting.
Microsoft offered a buyout to about 60,000 people this week — roughly a quarter of the company, eligible up to vice president. American Express and Barclays announced cuts the same week and named AI as a reason. The headline writes itself: the AI layoffs are here.
Read the research those headlines are supposedly built on and you get a different sentence. AI does not take jobs. It takes tasks. The difference is not semantic. It is the difference between a buyout check and a promotion case, and almost nobody is counting the tasks.
What the study actually measured
In 2018, Erik Brynjolfsson, Tom Mitchell, and Daniel Rock did something the layoff memos never do. They scored work at the task level.
- They took the federal O*NET database — 964 occupations, 2,069 work activities, 18,156 individual tasks — and ran every task through a 23-question rubric for "suitability for machine learning." (Brynjolfsson, Mitchell & Rock, AEA Papers and Proceedings, 2018)
- Finding one: most occupations in most industries have at least some tasks that are suitable for ML.
- Finding two: few, if any, occupations are fully automatable. Suitability varies widely inside a single job.
- Their conclusion, in their words: full automation will matter less than the reengineering of processes and the reorganization of tasks.
Do the arithmetic on their own dataset and an occupation averages about nineteen listed tasks. Brynjolfsson rounds it to roughly twenty-five when he talks about it, and the exact count is not the point. The point is that a job is a bundle, the bundle is uneven, and the machine only wants part of it.
Call it eight of twenty-five. Eight tasks the model can do faster than you. Seventeen it cannot do at all.
The industry's own data says the same thing
- Microsoft Research published its "Working with AI" dataset in 2025 — 200,000 real Copilot conversations mapped against O*NET work activities, scored for how much of each occupation AI actually touches. It is the dataset behind the "Will AI Take This Job?" box on every TMZ blueprint.
- The number one occupation on that list — interpreters and translators — scores 0.49. Half.
- The median occupation, rank 392 of 785, scores 0.14. Eight seats on the list score zero.
- In other words, the company that built the tool measured its own tool and found that for the typical job, AI reaches about a seventh of the work. Even at the very top of the exposure list it does not reach half.
That is the Brynjolfsson result, seven years later, from the vendor's side of the table.
Two companies, same eight tasks
Here is where the headline and the research split, and it is the only part of this piece that changes what you do tomorrow.
Two companies each have an analyst. The analyst's job is twenty-five tasks. Both companies buy the same AI, and it eats the same eight: the pull, the reconciliation, the first-draft summary, the formatting, the status note.
Company A keeps the analyst. She does the eight in an hour instead of a day and spends the recovered time on the seventeen — the conversation with the vendor who is quietly over-billing, the number that is technically correct and obviously wrong, the judgment call the model cannot make because nobody wrote the rule down. Her output goes up. Her value in the seat goes up with it.
Company B cuts the analyst and books the savings. Then it discovers that the seventeen were never on a task list. They lived in the person. The vendor keeps over-billing. The wrong-but-correct number ships. The judgment call gets made by whoever is left, at a worse price, with less context. The savings are real on the expense line and the harvest never shows up anywhere, because nobody was tracking the seventeen in the first place.
Goldman Sachs put a number on this in October. Job growth in the most AI-exposed industries has run about 10,000 a month below trend, nearly offset by hiring for AI-related construction. Tasks are moving. Headcount, in aggregate, is barely moving. The companies that are cutting are not reading a task list. They are reading a capex budget.
This is the mechanism underneath the amplifier argument. AI multiplies the judgment already in the seat. Empty the seat and there is nothing to multiply.
Which seventeen
The tasks AI does not reach cluster the same way in every dataset:
- Anything physical. The Aging Out Board exists because the seats where most of the twenty-five are physical score near zero and are retiring anyway.
- Anything that requires being in the room. Relationships, exceptions, the thing the customer said that was not what the customer meant.
- Anything nobody documented. The rule that lives in one person's head is, by definition, not training data.
- Anything where the input is messy and the output is a judgment. Brynjolfsson's rubric scores a task high when inputs and outputs can both be measured well enough for a machine to learn the mapping. Most of the seventeen fail that test on purpose.
Scot Free Take
Don't rely on your job description. Map your actual day-to-day workflow. Circle the 8 tasks the machine can do, and build an airtight evidence file around the 17 it can't. Those 17 tasks are your promotion case—and your job security.
The move
List your twenty-five. Not your job description — the actual work, the way a time-and-motion study would see it. Circle the eight the model can already do. Be honest; it is probably more than you want it to be.
The seventeen left over are your next promotion case. They are the tasks your company cannot cut without discovering what they were. Build the evidence file on those — dated, quantified, owned — and the next time a budget meeting happens without you in the room, the people in it will be looking at a list of things only you do.
If you want the structure for that file, it is the Side Door Playbook, and it is free.
Sources
- Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). "What Can Machines Learn, and What Does It Mean for Occupations and the Economy?" AEA Papers and Proceedings, 108, 43–47.
- Tomlinson, K., et al. (2025). "Working with AI: Measuring the Occupational Implications of Generative AI." Microsoft Research. Data: github.com/microsoft/working-with-ai (CC BY 4.0).
- Goldman Sachs Global Investment Research, Briggs, J. (October 2026), as reported by Fortune, Oct. 22, 2026.
- Microsoft voluntary exit program, Oct. 26–27, 2026 (Bloomberg; company statement).
- Brynjolfsson, E., interview, Me, Myself, and AI (MIT Sloan Management Review / BCG), 2026.
Get the MoneyZoo LAB Report every Monday and instant access to The Side Door Playbook — one seat, one receipt, one move. Subscribe here.
