The AI Productivity Paradox: Produce More, Deliver the Same
By Anis Hammouche·June 29, 2026
Producing more with AI does not mean delivering more. The draft appears in minutes, but reviewing, testing and fixing it takes back a good part of the time you saved. If you freelance or run a small agency, what counts is what reaches the client in good shape.
Take Malik, a freelance developer (an invented case, like every figure in his examples in this article). Without AI, he writes and checks an invoicing module in 6 hours. With AI, the first draft comes out in 1 hour, then he spends 4 hours reading it, testing it and fixing what does not hold. Total: 5 hours instead of 6, which is 1 hour saved out of 6, about 17%. The draft arrived very fast, the delivery only a sixth faster.
Where the saved time goes
The AI productivity paradox fits in one line: the time saved writing is lost again checking. Malik's case is nothing exotic. A survey run by The Harris Poll for GitLab between 6 and 21 April 2026, among 1,528 IT and development professionals in six countries including France, describes the same mechanism. Blog du Modérateur summed it up on 29 June 2026 (in French):
- 79% of respondents say individual developer productivity has improved, but that the delivery process as a whole has not kept pace;
- 85% think AI has moved the friction point from writing code to reviewing and validating it;
- respondents say they spend 16% of their time writing code.
These figures come from Blog du Modérateur's summary, I have not read the GitLab report itself. Two caveats. The survey was commissioned by a vendor of development tools, which has an interest in code validation becoming a topic. And its respondents work inside organizations, not alone at a desk. I read these percentages as an order of magnitude, not as a measurement of your work. For that, you have to count your own time.
The bottleneck moves
A chain of work always has a slowest link. When you speed up another step, the queue grows in front of it. Picture a four-person communication agency (invented too). Without AI, it writes 5 draft posts a week for its clients. With AI, it writes 20. But one person reviews and signs off, and she gets through no more than 8 posts a week before they go out.
The math is quick. The agency publishes 8 posts instead of 5, which is 3 more on 5: 60% more, not four times more. The other 12 drafts pile up every week: 48 after a month. The agency feels fast because the start of the chain runs flat out, while the output is capped by approval.
| Job | What you can see | What decides |
|---|---|---|
| Freelance developer | Lines of code generated | Modules delivered without a comeback from the client |
| Communication agency | Draft posts produced | Posts approved and published |
| Online shop | Support replies drafted | Customers whose request is settled without a second message |
| Freelance consultant | Meeting notes generated | Notes sent without corrections |
The left column climbs fast and is effortless to measure. It reassures, and it says nothing about what the client receives. The right column is slower to collect, but it is the only one that pays.
Two debts that volume hides
The first is about code, so about maintenance. In the same survey, 73% of respondents worry about their teams' ability to maintain AI-generated code over time, and 82% think it risks creating a new kind of technical debt. If you freelance, that is a client calling you back in six months to fix what you shipped fast.
The second I cannot put a number on. In a small agency, beginners learn on the small tasks: the first draft, the formatting, the meeting notes. Those are the tasks AI absorbs first. If your first hire no longer does them, how do they learn to spot bad copy or bad code? That is reasoning, not measurement, and I have no source to settle it. Ask yourself the question before you hand the machine everything that trains people.
Measure before you spread it
Before you extend a use of AI to all your work or the whole team, ask three questions. Which precise result should it change for the client: a delay, a number of errors, a number of comebacks? How long does the work take before and after, on the same scope, review and rework included? And where is the slowest link in the chain, the one that caps the output?
If the use speeds up a step that was not the bottleneck, you are producing noise. If it relieves the slowest link and the result moves, you have found a lever. To limit what AI misses from the start, brief it like an intern. The full measurement method is in how to measure what a tool really brings you. Here is a sheet to fill in for one use at a time, over a week or two:
AI use tested: [what the AI does for me]
Scope: [task and period, for example October quotes]
Time before, without AI: [hours per week]
Time after, with AI: [hours per week, review and rework included]
What happens to the client: [delay, errors, comebacks]
Verdict: [I keep it / I adjust it / I stop]
Log the review time on its own line. It is the line nobody counts at the start, and the one that tips the verdict.
Frequently asked questions
So does AI waste time? Not in Malik's example: AI saves him 1 hour out of 6. The gain exists, it is just smaller than the draft makes it look. It depends on what you count: if you only time the generation, you always overestimate.
How do I know if a use creates value or just volume? To know whether a use of AI creates value, tie it to a result the client sees, then compare end-to-end time before and after on the same scope. If only the number of drafts goes up, it is volume.
Should I keep training tasks for a beginner? That is my opinion, not an established fact: yes, some of them. Keep the tasks that teach them to judge, and give AI the ones that teach nobody anything.
Is review really the weak point of AI? On content that shows figures or sources, yes: CNIL reminds readers that generative models can produce inaccurate results that look plausible (in French). I go through the reflex in AI invents figures and sources.
Sources
- Blog du Modérateur, Code généré par IA : les développeurs face au paradoxe de la productivité, 29 June 2026 (in French)
- CNIL, Les questions-réponses de la CNIL sur l'utilisation d'un système d'IA générative (in French)
Malik's and the agency's examples are fictional. The survey percentages come from the first source.
The next step is From solo to small team: what AI replaces, where the same question comes up on the hiring side. I share this kind of test in my newsletter, you can sign up right below.
