What to charge for AI work without lowballing yourself

The fastest way to lose money with AI is to price the way you priced before AI. You used to spend two days on a thing, so you charged for two days. Now the same thing takes you two hours, so you charge for two hours, and you just handed most of your income to the tool that made you faster. That is backwards, and the research says exactly how backwards.

Across knowledge work, AI compresses the time a job takes, and the effect is large and measured. In a randomized trial published in Science, ChatGPT cut the time professionals spent on writing tasks by about 37% and raised the quality of the work at the same time. Developers finished a coding task 55.8% faster with GitHub Copilot. Support agents resolved 14% more issues an hour, and 758 BCG consultants working with GPT-4 were 25.1% faster. None of those studies clocked an AI automation build specifically, but the pattern holds everywhere it has been tested: the clock you used to bill by is shrinking.

The short version

If you are quoting a job this week, the method in five moves:

1. Drop the clock wherever the job has a finish line. AI compressed the hours, so a rate times the new, smaller number is a pay cut you wrote yourself. Keep hourly only for genuinely open-ended work.
2. Default to a fixed price on the outcome. Go value-based when the saving is large and provable, and add a retainer for the care the model will need after launch.
3. Set the ceiling from their side, the floor from yours. What the problem costs the client caps the price; what the job costs you, scoping and revisions and breakage included, sets the number it must clear.
4. Anchor with a range, close on one number. Across 53 studies, anchoring reliably moves what buyers will pay, so open with a range whose low end is your target.
5. Write the boundary into the price. Two revision rounds, a change-order clause, and a deposit up front, because 57% of agencies leak $1,000 to $5,000 a month to work they never billed.

The rest is the evidence, and the walk-away rule.

Why hourly punishes you for being fast

Hourly billing ties your pay to how long something takes. AI cuts how long things take. So the better you get with these tools, the less you earn per project. The mechanism is not subtle: in the Science trial, 68% of people submitted the AI’s first draft with little or no editing, spending about three minutes of their own after they pasted it in. A rate times three minutes is not a business.

There is a second twist. The gains land hardest on the less experienced worker, who improved about 34% while the expert barely moved. AI is quietly commoditizing the speed and polish a client used to pay a premium rate for. What it cannot commoditize is knowing which tool, wiring it up so it holds, and pricing the result. So price the result.

Hourly is fine for open ended work where the scope genuinely is not known yet, a research spike, a “help me figure out what is even possible” session. For anything with a clear finish line, price the finish line.

Three ways to price, and when each one fits

You are in good company dropping the clock. In a 2023 survey of nearly a thousand consultants by Consulting Success, a firm that sells value-pricing coaching, hourly was only the second most common model at 29%, behind project-based fixed fees at 30%, with value and retainers taking most of the rest. Agencies split about the same way. Hourly is one option among several, and often the worst one.

Fixed price per outcome. You agree on a deliverable and a number up front. Good when the scope is clear: “an automation that turns these emails into CRM entries,” “a chatbot trained on your docs that answers the top 20 support questions.” The client knows the cost, you keep the upside of working fast. This is the default for most AI build work.

Value based. You price against what the result is worth to the client instead of what it costs you to make. If an automation saves a team 15 hours a week, that is worth far more than the afternoon it took you to wire up. In the same survey, 51% of consultants who price on value report an average project over $10,000, against 39% of those who bill hourly. Treat that as correlation rather than a magic switch, since the people confident enough to value-price were probably already charging more. But the direction is consistent, and value pricing is where most people undercharge by an order of magnitude.

Retainer. A flat monthly fee to keep things running, tune prompts, handle the model changing under you, add the next small thing. AI work is not “ship it and walk away,” models shift and break, so ongoing care is a real service worth charging for. Retainers turn one project into steady income.

The menu, in one look:

Model	Fits when	The evidence
Hourly	Scope genuinely unknown: research spikes, discovery	29% of consultants; punishes speed everywhere else
Fixed per outcome	The finish line is clear	The default for AI builds; you keep the speed upside
Value based	The saving is large and provable	51% report $10k+ average projects, against 39% hourly
Retainer	The work needs ongoing care	Models shift and break; one project becomes steady income

How to set a number you can defend

Start from the client’s side. Ask what the problem currently costs them: the salary hours going into the manual version, the leads they drop, the customers who churn waiting on slow replies. That number is your ceiling, and it is usually higher than the price you were about to name.

Then sanity check from your side. Whatever a fixed price comes out to, make sure it clears what you would want per hour of your real time, including the unglamorous parts: scoping, revisions, the model breaking the week after launch. If it does not, the scope is too loose or the price is too low.

A few working rules, and the research behind them:

• Name the price as one number for the outcome. “This is $X” lands better than “it is $Y an hour and I think about Z hours.”
• Quote a range before you quote a single number. Anchoring is one of the sturdiest findings in pricing: across 53 studies it reliably moves what buyers are willing to pay. Put your target at the low end of the range and let it run upward, which anchors the buyer high while the range still reads as cooperative.
• Build revisions into the price (two rounds, say), and put a line on what counts as new scope. Scope creep is where fixed price jobs go to die, and it is a majority problem: PMI found 52% of projects hit it, and in a 2025 survey of 273 agencies, 57% lose between $1,000 and $5,000 a month to work they did but never billed, with nearly 80% saying they rarely or only sometimes charge for it. You will not bill your way out of overruns after the fact. Price the boundary in.
• Charge a deposit. Half up front is normal, about half of agencies take one and only 16% wait for full payment, and it filters out the people who were never going to pay.

Those levers, anchoring with a range and charging a deposit, lean on the same psychology that works on your own product.

When to walk

Some clients want the AI discount: they read that the tool is cheap, so they think the work should be too. They are not your clients. The value is in knowing which tool, wiring it up so it does not embarrass them, and being there when it breaks.

Generative AI could in theory automate the activities that fill 60 to 70% of the workday, by McKinsey’s estimate, which is exactly why clients now assume AI is in the mix, and exactly why billing for your keystrokes is a shrinking base. If someone only wants to pay for the API bill, let them go build it themselves, and if you want the numbers behind that argument you can point them at what each tool costs. There are better clients out there, and the harder part is usually how you find your first client.

The real version of pricing is refusing to let a tool that made you faster make you poorer. Price the outcome, show the value, and keep the upside of being good at this.