The number nobody puts on a slide
There is a pricing page in my head that I have never published.
Every solo operator has one. It is the number you say out loud when someone asks what you charge, and it is almost never the number your work actually costs you.
The gap between those two is the whole subject of this piece.
The market has opinions about the first number. Trawl recent AI-agency content and a fairly consistent picture emerges: retainers in the low thousands a month for small business, a bit more for mid-market, project work banded by complexity from a few hundred dollars for a single workflow up into the tens of thousands for something multi-agent. Those ranges come from marketing rather than anyone’s books, so treat them as the shape of a conversation rather than data. But the shape is real enough: the industry has converged on hybrid pricing — a fixed build fee, then a retainer to keep the thing alive.
What that picture leaves out is the second number.
The cost that does not appear in the stack
The seductive arithmetic goes like this: a few hundred dollars a month of AI tools replaces a salary. Margin becomes enormous. One person does the work of a team.
I run that stack. The arithmetic is not wrong, exactly. It is just answering a different question than the one that matters.
Because the tools are not the cost. The cost is the correction load — the hours that go into noticing that something confidently produced the wrong answer, working out which layer produced it, and repairing whatever downstream artefact already inherited it.
That load does not show up in a subscription line. It shows up as a Tuesday.
In the last fortnight alone, on work I can describe without naming a client: a regulator’s own web page printed a licensing threshold backwards, and that error had propagated into five of our own pages — including one where a previous pass had written a confident paragraph explaining why the wrong figure made sense. A transcription of a two-and-a-half-hour meeting came back with fifteen minutes of fabricated content spliced into the quiet passages. A published research pipeline ran daily for three weeks while nothing downstream turned its output into anything a reader could see.
None of those were tool failures. Every one of them was a verification cost.
Why the margin is real and the ceiling is real too
Both things are true at once, and the market content tends to hold only the first.
The margin is real: one person with the right stack genuinely does produce what used to need several. I would not have shipped what I shipped this year any other way.
The ceiling is real too, and it is not a capacity ceiling — it is a trust ceiling. The work scales until the amount of output exceeds the amount of output you can personally stand behind. Past that line you are not running an agency; you are running a publishing operation for claims you have not checked.
That is why the interesting engineering in a one-person shop is not the generation layer. It is the layer that tells you when the generation layer lied.
What I actually charge for
So the pricing question resolves into something less glamorous than the playbooks suggest.
You are not selling the output. The output is nearly free and getting cheaper every quarter. Anyone can generate a compliance document, a market summary, a landing page.
You are selling the fact that someone checked it — and, more to the point, that someone is accountable when it turns out to be wrong. That is the thing a subscription cannot supply and the thing a client cannot get from the tool directly, however cheap the tool becomes.
Which means the honest structure of a solo AI engagement is closer to professional indemnity than to software licensing. The build fee buys the thing. The retainer buys the continued assurance that the thing still does what it claimed to, after the vendor shipped an update, after the regulator changed a page, after the model changed underneath you.
The uncomfortable implication
If that is what you are selling, then outcome-based pricing is not the clever upgrade the market says it is — at least not for one person. Outcome pricing transfers verification risk onto you at exactly the scale where you have least capacity to absorb it. It reads as sophisticated. It is, for a solo operator, mostly a way of being paid last for being wrong earliest.
I would rather charge less and be able to check everything I ship.
What this means if you are one person
Three things I would tell someone starting where I started.
Price the correction load, not the tool cost. Whatever your stack costs, the real line item is the hours you spend finding out that something plausible was false. Estimate it honestly and put it in the number.
Build the checking layer before you build the volume. It is not the exciting part and it is the only part that scales trust. Every hour spent on a verification step buys back several hours of the repair it prevents.
Treat a confident output as an unverified one. The failure mode of this entire category is not gibberish — gibberish is easy to catch. It is fluent, specific, well-structured and wrong, and it arrives looking exactly like the good work.
The price of being one person is not the stack. It is that there is nobody else to catch it.
