How much does an AI agent cost?
The honest answer: it depends on five items, and they do not weigh the same. Before comparing quotes, you need to know which item each offer includes, and which it leaves out.
One published figure, to anchor the usage cost: at OOTI, the agent that answers questions about figures in the ERP data costs under $0.10 per question. That figure is measured, and it is on the case study. It says nothing about the build cost, which is a separate item.
Build it yourself or start from a platform?
This is the first item, and the most variable. An agent that really does the work needs building blocks nobody sees: the graph of steps, knowledge of the business, a map of the software, per-role rights, evaluation, the credit counter.
Building those blocks in-house represented twelve months of R&D for us. That is the work a platform saves you from redoing. What remains for each project is the business configuration: describing the activity, the tools, the rules and the test set.
A frequent trap is costing the prototype. An agent that demos well is quick to put together. An agent you let act on a production database needs evaluation, thresholds and a log, and that is where most of the time goes.
What do model calls cost?
Every action the agent takes consumes model calls, document searches and tool calls. That cost depends on three choices:
- the model chosen for each task: a simple classification does not need the same model as reasoning across several documents;
- the length of the context sent on each call, which a good graph reduces by loading only what the step needs;
- the number of steps in the path, which is set when the workflow is designed.
An agent designed as a graph has an advantage here: each step has its own model and context, instead of one large call that does everything.
What do evaluation and maintenance cost?
Evaluation is the item quotes forget, and the one that decides whether the agent reaches production. You need a set of real cases, built with your business experts, and a measurement at every change: of model, of prompt, of software version.
Maintenance follows the same logic. Your software evolves, models change, use cases widen. An agent without a test set degrades without anyone noticing. An agent with a test set is re-measured in minutes. The page on measuring an AI agent sets out what to measure.
Where to host it, and what does that change?
On your servers, on your cloud, or with us. Hosting on your side reassures your customers about their data and leaves you the choice of model. It also moves model consumption onto your account. Location questions are covered in AI sovereignty.
How does a software company bill an agent to its customers?
Three formulas exist, and they are not equal:
| Formula | What your customer pays | Limit |
|---|---|---|
| Premium tier | One more flat fee on the subscription | The price follows neither usage nor gain |
| Cost-based credits | Each action, at cost plus your margin | You leave value on the table |
| Outcome-based credits | Each successful action, priced on the work saved | You need to measure the gain once |
Our recommendation: prepaid credit packs, a price set on the outcome, and cost as the floor. One simple rule for the price: your customer keeps 90% of the gain, you bill the rest. The details are in the credit model.
To see what that means on your installed base, the My AI ARR calculator makes the estimate from your number of customers and users.
How long before the first revenue?
From workshop to production, six to eight weeks on a known activity with API access. Lemon Learning generated its first AI revenue three months after the first conversation. How we run those weeks is described in how we work.
