AI automation agencies usually charge a one-off build fee plus a monthly retainer for hosting and changes. Model usage is passed through at cost or with a margin. Outcome-based pricing works once you can measure results reliably.
You do not have to pick just one model. The right mix depends on how much the work changes after launch and how well you can measure cost per client. This guide covers each model, how to handle token costs, and the data you need before you quote.
Quick Answer
How should an AI automation agency price its work? Charge a fixed build fee for the first version. Add a monthly retainer for hosting, monitoring and small changes. Pass model and platform usage through, with or without a margin. Move to outcome pricing only when results are easy to measure.
How do AI automation agencies price their work?
AI automation agency pricing usually draws on five models. Each one shifts risk between you and the client. The table compares them side by side.
| Pricing model | What the client pays for | Best for | Main risk to the agency | Data you need |
|---|---|---|---|---|
| Fixed build fee | Scoped delivery of a workflow or agent | Clear, well-defined first builds | Scope creep eats the margin | Estimated hours per build stage |
| Monthly retainer | Hosting, monitoring, fixes and small changes | Live automations that need care | Retainer drifts below real effort | Monthly hours and support load per client |
| Usage pass-through | Model tokens, platform runs and API calls | Variable or high-volume workloads | Surprise bills damage trust | Token and run costs per client |
| Outcome or value pricing | A measured result, such as tickets resolved | Mature, measurable automations | Paying for results you cannot prove | Baseline metrics and attribution |
| Hourly or day rate | Time spent on discovery or change work | Audits and open-ended requests | Efficiency lowers your revenue | Accurate time records |
A sensible default combines a build fee, a retainer and usage. That mix covers the upfront effort, the ongoing care and the variable running cost. For a deeper look at each structure, see our guide to AI agent billing models.
The next question is how to split the work between the build fee and the retainer.
Should you charge a build fee, a retainer or both?
Charge both for most client work. The build fee pays for design and delivery. The retainer pays for keeping the automation running once real data hits it.
What goes into a build fee?
A build fee should cover discovery, workflow design, prompt and tool setup, testing and handover. Price it from estimated hours per stage, not from a gut feel. Add a contingency line for integration surprises, because client systems rarely behave as documented.
Write the scope tightly. List the triggers, systems, outputs and number of revisions included. Anything outside that list becomes a change request at your hourly or day rate.
What should a retainer cover?
A retainer should cover hosting, monitoring, prompt tuning, failure handling and a fixed allowance of change hours. Automations break when a connected app changes its fields or a model update shifts behaviour. Someone has to notice and fix that.
Set the retainer from measured effort, not a round number. Track how many hours each live client really takes in its first three months. Then publish clear tiers using AI agent rate cards so clients know what each level includes.
What are the trade-offs?
| Approach | Upside | Downside |
|---|---|---|
| Build fee only | Simple to sell and invoice | No income for maintenance; support becomes unpaid work |
| Retainer only | Predictable monthly revenue | Large upfront builds strain cash flow |
| Build fee plus retainer | Pays for delivery and ongoing care | Needs a clear line between build scope and changes |
With labour covered, the remaining variable is usage. That is where margins leak if nobody is watching.
How do you pass AI model and token costs to clients?
Pass model and platform costs through as a separate line, either at cost or with a stated margin. Hiding them inside a flat fee works only when usage is small and stable.
Usage costs are real and variable. On Anthropic’s pricing page (checked 5 October 2026), Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens. Larger models cost more. Automation platforms add their own fees. n8n’s pricing page (checked 5 October 2026) lists Starter at €20 a month and Pro at €50 a month on annual billing. Plans are priced by workflow executions.
Here is an illustrative example using that Haiku 4.5 rate. An agent run with 20,000 input tokens and 2,000 output tokens costs about $0.03. At 10,000 runs a month, the client’s model bill is about $300. Run your own numbers with the AI token cost calculator.
| Usage approach | How it works | When it fits |
|---|---|---|
| Pass-through at cost | Client pays the exact provider bill | Clients who want full transparency |
| Cost plus margin | Provider cost plus a stated percentage for handling | You manage keys, limits and billing risk |
| Bundled allowance | Retainer includes a usage cap, with overage above it | Stable workloads with predictable volume |
| Client-owned accounts | Client holds the API and platform accounts | Regulated clients or very high volumes |
Whichever approach you choose, show clients what the AI did and what it cost. Our guide to transparent AI billing covers report formats that prevent invoice disputes.
Once costs are visible, some clients will ask to pay for results instead. That needs a different foundation.
When does outcome-based pricing make sense?
Outcome-based pricing makes sense when the result is easy to count, easy to attribute and stable month to month. Examples include leads qualified, tickets resolved or invoices processed.
Check three conditions before you agree to it:
- A clean baseline. You need the client’s numbers from before the automation went live.
- Clear attribution. Both sides must agree which results came from the automation and which did not.
- Shared data access. You need reliable access to the systems that record the outcome.
Outcome pricing shares risk. If volumes fall for reasons outside your control, your revenue falls too. A safer structure keeps a smaller base retainer and adds an outcome fee on top. If you are making that shift, read our guide to moving from hourly to value pricing.
Every model above depends on knowing your real costs. That is the final piece.
What data do you need to price AI work profitably?
Profitable AI automation agency pricing starts with three numbers per client: human hours, agent runs and usage costs. Without all three, any quote is a guess. Track them alongside revenue:
- Human hours on discovery, building, reviews, support and change requests.
- Agent runs per workflow, including failed runs and retries.
- Usage costs for tokens, platform executions and third-party APIs.
- Revenue per client from build fees, retainers and usage lines.
Combine human and machine effort into one figure using a blended rate for human and AI work. That shows what an hour of delivered work truly costs you.
How do you derive a retainer from cost data?
Use a simple floor formula. The figures below are illustrative, not a benchmark.
| Line | Example monthly figure |
|---|---|
| Support and change hours (8 hours × $75 loaded cost) | $600 |
| Platform share | $60 |
| Expected model usage | $300 |
| Total monthly cost | $960 |
| Retainer at a 50% gross margin (cost ÷ 0.5) | $1,920 |
Swap in your own hours, rates and usage. Review the numbers after three months of live data and reprice if effort has drifted. Tools built as AI agency billing software keep these figures per client, so repricing is a report rather than a spreadsheet exercise.
Key Takeaway
Price AI automation on measured effort and usage: know your human hours and model costs per client before choosing a model.
Ready to see the real cost of every AI project?
Keito tracks human hours, agent runs and token costs per client so you can price with confidence.
Frequently Asked Questions
How much should an AI automation agency charge?
Charge enough to cover delivery hours, ongoing support and usage, plus your target margin. Price the build from estimated hours per stage. Then add a retainer and usage lines based on measured cost per client.
Should AI agencies charge hourly?
Hourly billing suits discovery, audits and open-ended change work, but rarely the build itself. Automation reduces the hours an hourly model rewards. A fixed build fee plus a retainer usually protects margin better.
How do agencies bill clients for AI token usage?
Agencies bill token usage as a pass-through line, at cost or with a stated margin. Some include a usage allowance in the retainer and charge overage above it. Others have clients hold their own API accounts.
What is a typical AI agency retainer?
There is no reliable public benchmark, so derive your retainer from cost data. Add monthly support hours at loaded cost, platform fees and expected model usage. Then divide the total by one minus your target margin.
How do you calculate margin on AI automation projects?
Margin is revenue minus human labour cost, model usage and platform fees, divided by revenue. Track each cost per client and per project. Recalculate monthly, because usage and support effort change after launch.