Price AI work so that speed pays off for you

You quote a price for a defined result, such as a lead-qualification agent with CRM integration. If your agents get it done in half the time, the difference is yours. If the scope grows or the model struggles, so is the overrun.

Fits when the result can be described precisely and you have built something similar before. Productized offers, the same automation sold to several clients, are fixed prices too.

Watch out: AI costs belong in the calculation, and after the project you need to know what it really took. Only then does the next quote fit better. A deposit and payment in parts reduce your risk on larger projects.

The client books a budget per month or quarter, as hours, as an amount or both: operating, monitoring and extending their automations, prompt maintenance, new workflows on request.

Fits when the work keeps coming but cannot be planned item by item. It gives you predictable revenue and the client a fixed point of contact.

Watch out: a retainer only works if both sides see how much of the budget is used. Unused hours lead to discussions about the value, overused ones to unpaid work. Agree upfront what happens when the budget runs out.

You bill your hours at your rate and pass AI costs on separately. The client pays what the work took; overruns are not your risk.

Fits for exploratory projects, prototypes and anything where nobody knows the scope yet.

Watch out: every time your agents get faster, the same result brings in fewer hours. And the client sees two kinds of cost on the invoice and will ask about both, so you need records that hold up.

This only matters where you work at API prices. On a flat AI subscription, the costs barely register against revenue; there, usage shows the effort without being billed.

• At cost: the AI costs as incurred, as an expense on the invoice. Transparent, but no margin on the risk you carry.
• With a markup: AI costs plus a percentage for setup, monitoring and price risk. Common in agencies for third-party costs in general.
• By token volume: your own price per million tokens. Independent of which model you use, so switching to a cheaper model raises your margin.
• By runtime: active agent runtime at an hourly rate. Easy for clients to grasp, but agents that work faster bring in less, as with hours.

The fee follows the client's gain: hours saved in support, more qualified leads, a process that no longer needs a person. Sometimes as a base fee plus a success component.

Fits when the gain can be measured and you can influence it. Most of the time that means a client who already trusts you, not a first project.

Watch out: you need a baseline and a measuring method agreed upfront, and you carry the risk that the result depends on things outside your control. Without knowing your own effort, you cannot tell whether the deal was good.

AI costs fluctuate. They depend on the model, the length of the context and how much a prompt cache can be reused. At Anthropic's list prices, an output token from Opus 5.5 costs 40 times as much as one from Haiku 5.5 (prompts up to 100K tokens), reading from the prompt cache costs a twentieth of the normal input price, and batch processing halves the price (Claude pricing, as of 10 Oct 2026). The same task can cost very different amounts depending on how it is run.

Agents get faster, and the hourly price erodes. An assumed example: an integration took you 20 hours a year ago, at 100 euros per hour that was 2,000 euros. With agents it now takes 6 hours. Billed by the hour, the same result brings in 600 euros. As a fixed price of 2,000 euros, it now earns 333 euros per hour, minus AI costs.

Both risks speak for prices based on results, and for knowing your effort precisely. The less you bill by the hour, the more important it is to know what really went into a project.

Whatever the model, clients of AI agencies ask the same question: what did we actually get? A client report answers it with working time, AI sessions with tool, model and runtime, and the commits that came out of them.

On a fixed price or a retainer, the report shows the effort without billing it. Amounts, rates and model costs can stay hidden. On time and materials, it backs up every item on the invoice.