# If AI does the work, what am I paying for?

> For the decisions and the checking. AI drafts research, first passes and boilerplate code. I decide what to build and what to leave out, make every design decision, read every line before it ships, and answer for the result. The price is fixed for that result, so how fast the AI was is my risk.

By Dion Zeneli · 13 October 2026 · updated 11 November 2026 · 5 min read

Source: https://www.theweb0.com/blog/if-ai-does-the-work-what-am-i-paying-for
Filed under: Process, Numbers

---

Yes, I use AI, and you should know where it sits before you hire me. It is in my research and in the first draft of routine code. It is not in the decisions: what to build, how it looks, and whether a line is good enough to ship. Below: who does what, stage by stage, what the research says about speed and trust, and what you get at the end.

## Does AI make the work faster?

Sometimes, and you can't tell by how it feels. The best test I know of is by METR, a research group that studies AI. In early 2025 it gave 16 experienced open-source developers 246 real tasks in projects they had worked on for years, and randomly allowed or banned AI on each one. Before starting, the developers expected AI to cut their time by 24%. Afterwards they believed it had cut it by 20%. Measured, the tasks with AI took 19% longer.[^metr25]

That was early-2025 tools, and METR now marks the result as out of date. Its follow-up, published in February 2026, points the other way: the raw numbers show a speed-up. But METR calls the new data "only very weak evidence" of how big, because many developers no longer wanted to work without AI and stayed out of the study.[^metr26]

*Figure: How much longer or shorter a task took when AI was allowed, from METR's 2025 study and its follow-up. Dots are estimates. Bars are the ranges METR gives for each one. METR treats the follow-up as weak evidence. Source: [METR](https://metr.org/blog/2026-02-24-uplift-update/). See it at https://www.theweb0.com/blog/if-ai-does-the-work-what-am-i-paying-for*

So I don't read either study as "AI is slow" or "AI is fast". The part that held up is the gap between what people felt and what the clock said. Developers who had just been slowed down were sure they had been sped up. The 2026 update repeats the warning: self-reported speed-ups "can be quite unreliable".

That is one reason I don't sell hours. If AI saves me a day, good. If it costs me a day chasing an answer that was almost right, that day is mine too. The price on the page doesn't move either way.

## Can you trust what it writes?

Not without checking it, and most developers agree. In Stack Overflow's 2025 survey, more of them distrusted the accuracy of AI tools than trusted it.

**46%** distrust the accuracy of AI tools. (Source: [Stack Overflow Developer Survey, 2025](https://survey.stackoverflow.co/2025/ai))

**33%** trust it. Only 3% trust it highly. (Source: [Stack Overflow Developer Survey, 2025](https://survey.stackoverflow.co/2025/ai))

The biggest frustration, named by 66%, was "AI solutions that are almost right, but not quite". The second, at 45%, was that debugging AI-written code takes longer.[^so25]

The 2026 survey, with 30,903 responses, draws the line more clearly. 48% trust AI when they can easily check the answer. Only 16% trust it with most tasks that aren't important decisions. More than 70% found it useful for writing and debugging code. For design, 49% did, and for communication, 35%. Stack Overflow's own summary is that developers "trust AI with machines but not with people".[^so26]

That is where I draw my own line. AI is useful where its answer can be checked, and someone has to do the checking. On your project, that someone is me.

## Who does what on a project?

| Stage | What AI does | What I do |
| --- | --- | --- |
| The brief | Not used | Decide what the site or product is for, and what to leave out |
| Research | Finds sources and drafts summaries | Open every source, check each number and quote |
| Design | Not used | Every design decision |
| Code | First passes, boilerplate, the same change across many files | The structure and the hard parts, then read every line before it ships |
| After launch | Not used | Answer your email |

You never get AI's first draft of anything. The middle column is fast and cheap now. The right-hand column is the part that goes wrong when nobody owns it.

## Why does it suit one person?

DORA, Google Cloud's research programme on how software gets built, put it plainly in its 2025 report. AI's "primary role is as an amplifier, magnifying an organization's existing strengths and weaknesses."[^dora]

The report is about organisations. My reading of it for a one-person practice is this. In a team, the person who asks the AI for something, the person who reviews it and the person who ships it can be three people, and each can assume one of the others looked closely. In my process they are one person. Whoever decides what a button should do is the same person who reads the code that draws it.

That is the good case for an amplifier: a short loop with one owner. It cuts the other way too. If I stopped reading the output, AI would amplify that just as faithfully. So the rule on the About page is the whole method: every line that ships goes through me.

## Why isn't it cheaper, then?

It is cheaper and faster than it used to be. That is how one person can ship a website in 3–⁠5 weeks. What AI doesn't remove is the judgment: what to build, what to cut, and catching the output that looks right and isn't. That is what the fixed price pays for.

If you only need to test an idea, an AI builder on your own is the cheapest route, and I'll say so on the call.

## What do you get at the end?

- The code and the files in your own accounts from week two: your repository, your hosting, your domain, your design files. Nothing has to be handed over at the end. 
- No AI output that I haven't read.
- Your confidential material kept out of AI tools unless you agree otherwise in writing. I can sign an NDA that says so.

Who you'd be working with, and how, is on the [About](https://www.theweb0.com/about) page.

[^metr25]: The tools were mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet, on tasks of about two hours each. Economists and machine-learning experts had predicted 38% to 39% less time. [METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), 10 July 2025, and the [paper on arXiv](https://arxiv.org/abs/2507.09089), both checked 11 November 2026.

[^metr26]: 57 developers, 800+ tasks, from August 2025. Returning developers: an estimated 18% less time, range 38% less to 9% more. New developers: 4% less, range 15% less to 9% more. [METR, update on its developer productivity study](https://metr.org/blog/2026-02-24-uplift-update/), 24 February 2026, checked 11 November 2026.

[^so25]: [Stack Overflow Developer Survey 2025, AI](https://survey.stackoverflow.co/2025/ai), checked 11 November 2026.

[^so26]: [Stack Overflow Developer Survey 2026, AI](https://survey.stackoverflow.co/2026/ai), checked 11 November 2026.

[^dora]: [DORA, State of AI-assisted Software Development 2025](https://dora.dev/research/2025/dora-report/), checked 11 November 2026.