Agencies should price AI builds at a fixed project price backed by a written scope and accuracy target, use paid discovery when scope is unclear, sell a monthly retainer for maintenance, and pass model and hosting costs through at cost rather than marking them up. On white-label work, a client price of roughly twice the partner cost leaves a 50-60% gross margin. For document search at $18,000 against a $7,250 partner cost, that is a 60% gross margin, or about 47% once 25 hours of the agency's own project management are counted.
Why is pricing AI services different from pricing web or marketing work?
Three things make AI work awkward to price with the habits most agencies already have. The effort is uneven - the first working version is fast, and getting from 80% right to 95% right is where the weeks go. Part of the cost keeps running after launch, because every conversation calls a model. And the client usually cannot judge the work by looking at it, so a demo that impresses in a meeting can still fail in month two.
The pricing that works treats those three facts directly: price the outcome rather than the hours, keep running costs out of your margin, and charge separately for the work of proving the system is right.
What pricing models work for agencies selling AI services?
Fixed project price
The default for a defined build: a support agent, document search, an automation. The client gets certainty and you keep the upside of working efficiently. It only works with a written scope - what the system does, which systems it touches and what accuracy counts as done - so pair it with a paid discovery phase when the scope is unclear.
Paid discovery
A short fixed-price phase, typically one to two weeks, that ends in a scope, an accuracy target, a feasibility check on the client's real data and a fixed quote for the build. It protects you from quoting blind and filters out clients who want AI because it was on a slide. Our document search feasibility report resells at $4,000-6,000.
Monthly retainer
For the work after launch: adding documents, re-testing after model updates, tuning on the questions users actually ask, and reporting. It is the cleanest margin an agency has on AI, because the hours are predictable and the value to the client is visible.
Usage costs passed through, never marked up
Model API and hosting costs should be billed to the client's own accounts. Reselling tokens with a markup makes a busy month your problem, ties your margin to a price you do not control, and gives you an incentive your client would not like if they understood it.
Hourly billing is the model to avoid. AI-assisted engineering keeps getting faster, and hourly pricing hands every efficiency gain to the client while leaving you every overrun.
How do you calculate the price and margin on a white-label AI project?
Start from what the client would otherwise pay and check it against your costs, not the other way round. Then do the arithmetic with your own time included - the most common mistake is quoting a margin that forgets the agency's project management.
A worked example: document search over one client corpus, with citations and an admin screen. Resale and partner cost are the mid-points of the ranges on our white-label RAG page; the agency time is an assumption of 25 hours of account and project management at a loaded cost of $90 an hour.
| Line | Amount | Share of price |
|---|---|---|
| Client price | $18,000 | 100% |
| Partner cost (build) | $7,250 | 40% |
| Gross margin | $10,750 | 60% |
| Agency time (25 h at $90) | $2,250 | 13% |
| Net margin | $8,500 | 47% |
A 60% gross margin becomes 47% once your own time is counted. That is still a healthy project, but it is a different number to plan around, and on a smaller build the gap is proportionally wider. If your account management on AI projects runs well past 25 hours, price that in rather than absorbing it.
What should agencies charge clients for common AI projects?
These are the resale ranges and partner costs we publish for agencies, with the gross margin each leaves before your own time.
| Project | Typical client price | Partner cost | Gross margin |
|---|---|---|---|
| Support agent on the client's documentation | $8,000-12,000 | $4,500-5,500 | ~50% |
| Triage and routing agent | $10,000-15,000 | $4,500-7,000 | ~54% |
| Document search over one corpus | $14,000-22,000 | $5,500-9,000 | ~58% |
| Internal copilot across 3-5 systems | $18,000-30,000 | $8,000-14,000 | ~54% |
| Corpus audit and feasibility report | $4,000-6,000 | $2,000-3,000 | ~52% |
The full tables, with timelines, are on white-label AI agents and white-label RAG.
How do you check an AI price is one the client will accept?
Hold the quote up against what the client would do instead. If it beats all three alternatives on their own terms, the price is defensible.
- Hiring. A fully loaded US AI engineer costs $230,000-260,000 in year one. A $20,000 build is a small fraction of one hire, delivered in weeks.
- A SaaS product. If they would otherwise pay per resolution or per seat, show the three-year comparison - the arithmetic is in custom AI agent vs SaaS AI agent.
- Doing nothing. Put a number on the hours the system saves or the tickets it deflects. A build that pays back in under a year rarely loses on price.
What are the most common mistakes agencies make pricing AI work?
- Quoting a fixed price without a written scope and accuracy target, then absorbing the gap between 80% and 95%.
- Leaving evaluation out of the quote because the client cannot see it. It is the part that stops the system failing in month two.
- Reselling model API spend with a markup, so the agency's margin moves with the client's traffic.
- Selling the build without the retainer, then doing the maintenance for free because the client assumes it is included.
- Forgetting the agency's own hours when calculating margin.
- Pricing hourly, which makes every efficiency gain the client's and every overrun yours.
If you are deciding whether to deliver this work through a partner at all, start with how agencies deliver AI without hiring.
How should agencies price AI services?
Price builds at a fixed project price backed by a written scope and accuracy target, use a paid discovery phase when the scope is unclear, sell a monthly retainer for maintenance, and pass model and hosting costs through at cost. On white-label work, a client price of roughly twice the partner cost leaves a gross margin of 50-60% - closer to 47% on a typical document search project once your own time is counted.
For partner costs and how an engagement runs, see white-label AI development for agencies.