Replacing a Job With AI: The Hidden Cost That Bites Twice
By Anis Hammouche·July 5, 2026
Replacing a job with AI rarely costs what the spreadsheet says. Once you count the review, the rework when the tool gets it wrong and the hard cases nobody handles anymore, the saving melts. If you are torn between hiring, taking on a freelancer or handing everything to a tool, redo the math before you decide.
Take Léa, who runs a three-person communication agency (an invented case, like all of Léa's amounts and durations in this article). Her clients send her 40 requests a week: touch-ups, questions, small emergencies. She is torn between a part-time assistant at 1,400 € a month, all costs included, and an AI tool at 40 € a month that would answer in her place.
The math that looks obvious
The starting mistake is comfortable because it fits on one line. Léa subtracts the price of the tool from the assistant's pay and concludes she saves 1,360 € a month. That reasoning assumes the tool does the same work with nobody watching it. But an assistant is also paid for the 10 requests a week that fall outside the script: the unhappy client, the file that will not open, the change that contradicts the previous one.
The same math, honestly
Here are Léa's assumptions, all fictional. Of the 40 weekly requests, 30 are simple and 10 are hard. She reads every tool answer to the simple ones, 3 minutes each. The tool gets one in ten wrong: 3 errors a week, at 20 minutes of rework each, because the client is already irritated. Nobody else handles the 10 hard requests: she spends 15 minutes on each. Count 4 weeks a month, and value her time at 50 € an hour, the rate she bills.
- Review: 30 × 3 min = 90 min a week, so 6 h a month.
- Rework: 3 × 20 min = 60 min a week, so 4 h a month.
- Hard cases: 10 × 15 min = 150 min a week, so 10 h a month.
| Line | Part-time assistant | Tool alone |
|---|---|---|
| Direct cost | 1,400 € | 40 € |
| Review (6 h at 50 €) | 0 € | 300 € |
| Rework (4 h at 50 €) | 0 € | 200 € |
| Hard cases (10 h at 50 €) | 0 € | 500 € |
| Total per month | 1,400 € | 1,040 € |
The gap is 360 €, not 1,360 €. And it is fragile. If the tool gets twice as many wrong, 6 errors a week, rework rises to 8 h a month, the tool's total to 1,240 € and the gap to 160 €. The table simplifies one point: it assumes the assistant handles all 40 requests without Léa reading anything. Finally, Léa's 20 hours weigh 1,000 € in the table only if they take the place of work she would have billed. If she has nothing to bill in those hours, the real cost is lower and the gap wider. It is up to you to put a figure on them for your case.
What the spreadsheet still forgets
Three costs escape this table. This is my reasoning, not sourced figures.
Quality drops quietly. The answers stay correct on the surface, but they miss the client's context, and an irritated client does not always say so. By the time it shows in the numbers, the damage is done.
Rework costs more than the first pass. Correcting an answer that has already gone out takes longer than reading it before it leaves: in Léa's assumptions, 20 minutes to repair an error that reached the client against 3 minutes to review an answer. That is the "twice" in the title: you pay for the tool, then for the time to repair what it got wrong.
And knowledge leaves with the job. An assistant who knows each client's habits never wrote them down anywhere. When she goes, you only find out the day you would have needed them.
What documented reversals say
Some cases are documented. Siècle Digital published an article on 2 July 2026 (in French) about companies reversing course after replacing staff with AI. According to the article, which does not name the original surveys here, 32% of US hiring managers say they cut a role mainly because of AI and then rehired for it, and 55% of employers who restructured their workforce for AI regret it. It cites Ford, which rehired more than 300 engineers after quality problems, and the Commonwealth Bank of Australia, which reportedly laid off more than 40 customer service agents, found that its system could not handle real requests, then cancelled those cuts.
These are large organizations and American figures, not three-person agencies, and I have not found these cases anywhere but in this article. I take them as a signal, not as proof for your case. The mechanism is the same as in Léa's table: the hard work does not disappear, it changes hands.
Replace or augment: how to decide
Instead of asking which job AI can cut, ask which low-value part of that job it can absorb. If you freelance and are torn between subcontracting and handing the work to a tool, Léa's math applies to you too: replace the 1,400 € with the subcontractor's cost. The table below helps you decide.
| Criterion | Hand the job to the tool | Keep the person, equip the repetitive part |
|---|---|---|
| Assumption | The tool covers all the useful work | The tool covers the repetitive part, the human keeps the judgment |
| Hard cases | Fall back on you | Handled by the person, with more time |
| Rework | Every error is fixed by hand | A human checks before it goes out |
| Knowledge of the client | Lost with the job | Kept and reassigned |
| Gain | One cost line less, fragile | Time given back on the repetitive part |
The gain in the right column looks more modest on paper, but it does not rest on any bet that human judgment will disappear. If you are moving from solo to a small team, also look at what AI replaces and what it will not.
Redo the math with your figures
Paste this into a spreadsheet or a note, and replace the brackets with your values. To see where an AI tool's other costs come from, also read the real cost of AI.
1. Cost of the job or the subcontractor per month: [amount]
2. Cost of the tool per month: [amount]
3. Review hours per month: [number of simple answers] × [minutes per review] ÷ 60 × [weeks per month]
4. Rework hours per month: [number of errors] × [minutes per fix] ÷ 60 × [weeks per month]
5. Hard-case hours on me per month: [number of cases] × [minutes per case] ÷ 60 × [weeks per month]
6. Value of my hour: [my hourly rate]
7. Real cost of the tool = line 2 + (line 3 + line 4 + line 5) × line 6
8. Gap with the job = line 1 - line 7
Frequently asked questions
Can AI replace a whole job? Rarely. A job is paid for its hard cases as much as for its simple tasks. In Léa's example, the tool absorbs 30 requests out of 40, but the other 10 and the review fall back on her.
Why do companies reverse course after replacing jobs? According to the Siècle Digital article, because the tool is not enough on complex points: Ford had quality problems, and the Commonwealth Bank of Australia reportedly found that its system could not handle real requests. It is the gap between the simple case in the spreadsheet and the hard case in the field.
How do I know which part of a job to automate? Separate the tasks that never have an exception, which the tool can handle under review, from the tasks where an error is costly, which stay human. Then measure review and rework time over one real week before you put a figure on the gain.
Does augmenting a job really pay? Augmenting a job gives back time on the repetitive part, not one cost line less. In my view that gain is steadier than a cut, because it does not depend on how good the tool is on hard cases. That is reasoning, not a measurement.
Sources
- Siècle Digital, Après avoir remplacé des salariés par l'IA, ces entreprises font déjà marche arrière, 2 July 2026 (in French)
The amounts for Léa, the durations and the error rate are fictional assumptions. The American figures, Ford and the Commonwealth Bank of Australia come from the article above.
The next step is ROI of AI: how to measure what a tool really brings you. I share this kind of calculation in my newsletter, you can sign up right below.
