Before I send anything to someone who is paying for it, I ask one question. If they had pasted their own brief into a chatbot, how close would they have got to what I am about to deliver? If the honest answer is “very close”, I am not selling work. I am selling the fact that they didn’t try.
That is the whole distinction in this post. Using AI in paid work is normal now; I build websites for small businesses and use it every working day. Selling what comes out of it untouched is something else, and I know what that looks like at scale because I did it to my own site. Between June and August 2026 this blog published a hundred posts that were, give or take a light edit, raw output, and an ad network turned it down four times. The counts are in my post-mortem of those hundred posts. Nobody paid for those posts. If anyone had, they would have had every reason to ask for their money back.
Four signs a deliverable is still raw
I check my own work for these before it goes out. None of them needs a detector, only a second read.
- It contains nothing the buyer told you. No service they actually offer, no town, no question a customer really asked. If the text would fit any business in the same trade, it came straight out of the model.
- It asserts things nobody supplied. Models fill gaps with plausible material. In September, going through my own work, I found demo pages I had built for prospects that carried customer reviews nobody had ever written. The layout had a testimonials block, so the model filled it. They are gone now, and a review section only goes on a page if the business hands me real ones.
- Its facts haven’t been checked against a source. A wrong price, a renamed plan, a law paraphrased from memory. The model is exactly as confident about those as it is about the things it gets right, which is why confidence tells you nothing.
- There is nobody behind it who can say why a line is there. Ask “why does this page say that?” If the only answer is “the model wrote it”, the work isn’t finished.
What a person still has to put in
The model does more of the typing than I do. What it can’t do is the list below, and the list below is what the buyer is paying for.
The facts, from the business. The website packages I sell include the base texts. The model drafts them from what the owner tells me; I rewrite them; and every fact in them, from services to hours to service area, is confirmed by the owner before the site goes live. If the owner can’t confirm it, it comes out.
Something only that business knows. One of the maintenance options on my price list includes two articles a month. That is exactly the kind of recurring job where raw output is most tempting, so the rule for it is simple: every article has to contain something the business knows and the model doesn’t, whether that’s a question a customer really asked, a local detail or one of their own photos. If it doesn’t, it doesn’t get published that month. I set out how I write anything with AI help without it turning into filler in my workflow for SEO articles.
Verification. Every number, plan name and legal statement read off the source on the day, with the date written down.
Cutting. A model almost never removes anything on its own. Deciding that a section shouldn’t exist is one of the most valuable things a person does to a draft, and one of the least visible.
A name. Someone who will answer for it when it’s wrong. That, more than the prose, is what separates work from output.
Price the result, not the words
Per-word pricing was always a proxy for effort, and a model has made words nearly free. Charge by the word today and either the price collapses or it quietly stops describing what you do. Per-hour pricing has the opposite problem: it punishes you for using the tool, or invites you to pretend you didn’t.
So I price the deliverable. A site with a defined set of named pages, the features listed, a stated number of rounds of changes, and a fixed price. The buyer knows what they are getting and I know what I have promised. How I scope and package that in practice is in my post on scoping and quoting as a one-person studio.
The trap in pricing the result is underestimating the checking. With AI in the loop, the draft takes minutes and the review takes the time, so the review is what the price has to cover. If a quote only has room for the draft, it is a quote for raw output.
Saying it to the client
This sentence goes in every proposal I send: “I build with AI tools. Nothing reaches you that I haven’t read, checked against what you’ve told me, and wouldn’t put my name to.” It is in the proposal and not the small print, because a buyer who finds out later feels misled, and a buyer who is told at the start gets to ask the better question: what are they paying for, if not the typing?
There is a legal footnote worth knowing, with the caveat that I am not a lawyer. Article 50(4) of the EU AI Act, which applies from 2 August 2026 in the text as published, requires deployers to disclose AI-generated text published “with the purpose of informing the public on matters of public interest”, and exempts text that “has undergone a process of human review or editorial control” where someone “holds editorial responsibility”. A plumber’s services page is not a matter of public interest, so most of my work is outside it. But I find it telling that the one legal line drawn so far sits exactly at human editorial responsibility.
Paid work I don’t do, and what I know about it
Five posts that used to be on this site covered other kinds of AI-assisted freelancing, and they now point here. I don’t offer any of the services below, so this is what I know from published rules, not from doing the work. There are no rates in it, because I have none I could stand behind.
- Translation and post-editing. There is an international standard for it, ISO 18587:2017, “Translation services — Post-editing of machine translation output — Requirements”, which distinguishes full post-editing, meant to be comparable to human translation, from lighter work. A market that has written a standard about the gap between machine output and a finished translation already knows the difference.
- Transcription. Recordings of real people are personal data. Transcribing them for a client in the EU generally makes you a processor under the GDPR, and Article 28 requires a contract setting out what you may do with that data. Automatic transcription is the cheap part. Accuracy on names and jargon, and the data handling, are what gets paid for.
- Voice work. ElevenLabs’ prohibited use policy, last updated 17 August 2026, forbids using its output “to intentionally replicate the voice of another person” without “consent or legal right”. Any job involving a real person’s voice starts with written consent, whatever tool you use.
- Teaching people to use AI. Article 4 of the AI Act, which has applied since 2 February 2025, requires providers and deployers of AI systems to take measures to ensure “a sufficient level of AI literacy of their staff”. That is a real obligation for European employers. I don’t run training, and I would want a record of doing it well before selling it.
- “Prompt engineering” as a service. My opinion only: small buyers pay for a result, not for the prompt that produced it. The skill is real, but it is part of doing a job well, and I wouldn’t know how to sell it on its own to the kind of business I work for.
- Marketplaces such as Fiverr and Upwork. Both take a service fee from the freelancer, and each publishes its own rate. Check the current figure on their own pages before you price anything. Fiverr’s pages refused my requests on the day I wrote this, so I’m not quoting a number I couldn’t read.
If the tools themselves are the question, what I actually ask a model to write each week, and what comes back, is in my notes on AI writing tools, which also says plainly which tools I have never used.
What stays mine
If you strip away the tools, what I sell is short. Deciding what the thing should be. Getting the facts from the one person who has them. Checking everything a model said against where it should have come from. Cutting what shouldn’t be there. And being the person who answers when it’s wrong.
None of that got faster when the models got better. It got more important, because the other half of the job got cheap for everyone at once, buyers included. If you are wondering whether AI work pays at all, and what it costs to find out without betting much on it, that is the question my ChatGPT post tries to answer without a single income figure.