Three jobs, one name.
The phrase covers three kinds of work. All three are real. Few products need all of them.
Training or fine-tuning a model
Machine learning engineers, your own data and compute. Needed when no existing model can do the task, or when the data cannot leave your systems.
Building on a model through an API
Prompts, retrieval over your documents, tools the model can call, and tests that show whether a change made it better or worse. This is what most assistants, copilots and agents are made of.
Building the product around it
The interface, the permissions, what a person checks before anything is sent, and what happens when the model is wrong. This decides whether people trust it enough to use it.
We build the third, and the code that connects it to the model. We do not train models.
What it costs in the United States.
Clutch publishes what clients paid for AI projects and what AI companies charge per hour in each country. The Bureau of Labor Statistics publishes what the people cost if you hire them yourself.
An AI project, as reviewed on Clutch
$10,000 to $49,999
The band most AI development projects fall in. The average is $120,594 over about 10 months, far above the typical band, which is what a few large projects do to an average.
Source: Clutch, AI Pricing Guide 2026
A US AI development company, per hour
$50 to $99
The average cost per hour Clutch lists for AI development companies in the United States. Across all countries most companies charge $25 to $49.
Source: Clutch, AI Pricing Guide 2026
A data scientist on your payroll
$120,230 a year
The median wage in the United States in May 2025. For computer and information research scientists it was $140,300, and for software developers $135,980.
Source: US Bureau of Labor Statistics, Occupational Outlook Handbook: Data Scientists
Running it
Model providers bill for usage, by the amount of text sent and returned, every month. That is not part of a build quote, so ask for an estimate at your expected volume.
Figures as published on the pages linked, read on 18 November 2026. Clutch states no sample size, so treat them as direction, not as a quote.
What to ask before you hire.
Six questions tell you whether a team has shipped an AI product or built a demo.
Which model, and why
An existing model should be tried on your real cases first. A team that proposes training its own before that should be able to say what the existing ones failed at.
How will we know it works
Ask for the test set: real cases the product is checked against before every change. Without one, a new prompt that improves the average can quietly break one kind of request.
What does a person check
Where the model is unsure, the product should say so and send the decision to the person who owns it, with the source beside the claim.
Where does our data go
Which provider sees your prompts and documents, what it keeps, and for how long.
Who owns the prompts, the tests and the code
All three should be yours, in your repository, from the first day.
What will it cost to run
A monthly estimate at your expected volume, and what happens to it if usage doubles.
How we design interfaces where a person checks what the AI did