ai:pricing
How to price AI services at your agency.
By Loan Laux · September 7, 2026 · 8 min read
How to price AI services comes down to one rule: stop billing AI-delivered work by the hour. AI compresses delivery time, and hourly billing hands that gain straight to the client. Price fixed against the value of the outcome (the cost of the workflow you are replacing), split every engagement into a build fee and a run fee, and meter model usage in the contract so the client's growth is revenue instead of a cost surprise.
The logic holds whether you deliver with your own team, with embedded AI engineers working alongside it, or through a partner behind your brand (the delivery models are covered in our white-label playbook). Who builds changes; how you price does not.
The margin trap of hourly billing.
The trap is mechanical. Suppose a deliverable used to take 40 hours at a $150 blended rate: a $6,000 invoice. AI-assisted delivery cuts it to 12 hours, and on hourly billing the same outcome now invoices at $1,800. Your output is identical, your costs went up (tooling, training, the time spent changing the process), and the client captured the entire gain.
It compounds. The better your team gets, the less you can invoice for the same outcomes, so hourly billing punishes exactly the capability you are investing in. Agencies that keep hourly rates on AI-delivered work end up in one of two places: they quietly stop getting faster, or they start padding timesheets. Both are worse than fixing the pricing model.
How to price AI services: the baseline method.
Every workflow worth automating has a cost baseline: what the client spends today to get the same output manually. A coordinator assembling reports ten hours a week, two support agents triaging intake, freelancers producing product copy at a per-piece rate. That number, not your delivery time, is the anchor for the price.
01:
Quantify the baseline
Hours per week times the loaded cost of whoever does the work now, plus a defensible number for errors and delay if you have one. Do the math with the client in the room; a baseline they helped calculate is a price they have already half-accepted.
02:
Scope the workflow tightly
Inputs, outputs, review stages, escalation paths, and a success measure. Fixed pricing goes wrong in exactly one situation: vague scope. The tighter the definition, the safer the fixed fee.
03:
Set a fixed price between cost and value
Your floor is delivery cost plus target margin; the ceiling is a fraction of the first-year saving. If the baseline is $100,000 a year and your cost to deliver is $15,000, anything in between leaves room for both sides. Where you land in that range is a judgment call about the relationship, not a formula.
04:
Split the build from the run
Quote the build as a fixed project and the operation as a monthly retainer. One invoice for getting the system live, a recurring one for keeping it that way.
05:
Meter the usage
Put model usage in the contract as a pass-through or a bounded allowance, never as an unlimited inclusion. The usage section below covers the structures that work.
Where each pricing model still fits.
Value-anchored fixed fees are the default for AI builds, but not everything is a build. A working map:
| Model | Use it for | Avoid it for |
|---|---|---|
| Day rate | Discovery, audits, and exploratory prototypes where scope is genuinely unknown | Any delivery work where AI compresses the hours |
| Fixed fee | Scoped builds with defined inputs, outputs, and a success measure | Vague scope or requirements that shift weekly |
| Retainer | Running systems: monitoring, prompt updates, model swaps, small changes | One-off projects with a real end date |
| Per unit | High-volume outputs with a countable unit: documents processed, tickets resolved, articles produced | Work where the unit is hard to define or verify |
Packaging: sell named outcomes, not hours.
A price is easier to defend when it is attached to a package instead of a person's time. Productizing does three things at once: it makes the sale repeatable, it takes your (shrinking) delivery hours out of the negotiation, and it gives the pipeline a natural sequence.
- An entry-point audit at a fixed price. A scoped assessment of the client's workflows with a ranked roadmap as the deliverable. Low risk for them, and it produces the baselines that price everything that follows.
- Named build packages. 'Intake triage automation', 'reporting pipeline', 'support assistant grounded on your docs': defined deliverables, a fixed fee, a stated timeline. Comparable past builds are what let you quote these with confidence.
- Operations tiers. Two or three retainer levels that differ by response time, change allowance, and what gets monitored. Tiers turn the maintenance conversation from 'why am I paying this' into 'which level do we need'.
- A rate card that feeds proposals. Named packages plus comparable past projects are exactly the inputs that let AI proposal automation propose a price a human then confirms.
The run fee is not optional.
AI systems are living systems. Prompts drift as models get updated, integrations break when tools change, edge cases accumulate, and the workflow the client's team actually runs mutates over time. A build invoiced once and then supported for free is the second margin trap: quieter than the hourly one and just as expensive.
Price the run as a monthly retainer covering monitoring, fixes, prompt and model updates, and a bounded allowance for small changes. If you would rather not carry the operational load yourself, that layer can sit with a partner behind your brand; it is what managed AI operations exists for. Either way, someone is paid to keep the system alive, because someone will be doing the work regardless.
Model usage: meter it or eat it.
Model providers bill API usage by the token, metering what you send and what the model generates separately. That cost is trivial at pilot volume and real at production volume, and it scales with the client's usage, not with your effort. A flat monthly fee that includes unlimited usage quietly moves the cost of the client's growth onto your side of the table.
Three structures work: pass usage through at cost plus a handling margin, include a bounded monthly allowance with a stated overage rate, or absorb it only for low-volume workflows where you have measured the ceiling. Any of the three is fine. Deciding after a surprise invoice is not.
Do it yourself, or bring us in.
Nothing above requires outside help. Pick your two most repeatable AI services, write the packages, set fixed prices against real baselines, and put a run fee and a usage clause in the next contract. The usual sticking point is not the pricing theory; it is that fixed pricing takes estimating confidence, and estimating confidence takes engineers who have built these systems before.
That is the gap we fill. Our AI engineers embed with agency teams to scope and build these systems alongside your people, which is what makes a fixed quote defensible, and the estimating ability stays with your team when we leave. If you want the delivery capability without the hiring, that conversation is free.
Frequently asked questions.
Should we charge less for AI-delivered work because it takes fewer hours?
No. The client is buying the outcome, and the outcome did not get cheaper for them because you got faster. If competition eventually forces the market price down, meet it through packaging and volume, not by voluntarily discounting your own efficiency gains before anyone asked you to.
How do we price AI work when there is no baseline to anchor on?
Charge a day rate for a short paid discovery phase, which is the one place hourly pricing still fits, and make producing the baseline part of its deliverable. A fixed-scope workflow audit does the same job for a whole operation: it ends with the numbers that price everything after it.
What do we do when a client insists on hourly billing?
It usually means procurement needs a comparable unit, so give them one that is not your delivery hours: a day rate for discovery, then fixed packages with named deliverables. If they insist on hourly for the build itself, cap it, and keep the cap close to your fixed quote. Uncapped hourly on AI-delivered work concedes the whole margin argument before the project starts.
Should we tell clients that AI is doing part of the work?
Yes, framed as process rather than confession: what is automated, what a person reviews, and what you measure. The disclosure conversation goes fine when the price was anchored to the outcome. Clients only reach for 'AI did it, so it should cost less' when the invoice was denominated in hours to begin with.
How should we price custom agent projects?
More carefully than workflow automations. Multi-step agents consume many times the tokens of a single-pass workflow and fail in more expensive ways, so bound the usage clause tightly and keep a larger contingency inside the fixed fee. Read custom AI agents for agencies before quoting one.