Let’s talk about your project

What an AI agent costs, in four separate lines

Four separate lines nobody separates: build, model usage, hosting and maintenance — and which one surprises everyone.

Published · Neterius

Most AI quotes give you one number. That is where the surprises come from, because an agent has four costs with completely different shapes: one is a project, one scales with usage, one is fixed monthly, and one is a contract.

1. Build — a project cost

Defining the task, connecting the sources, bounding the actions, wiring the handoff and assembling an evaluation set. It is quoted by scope, once. What moves it is the number of systems the agent has to read from or write to, and whether the source knowledge has to be written before it can be indexed.

2. Model usage — scales with volume

This is the line that surprises everyone. It is charged per conversation and depends on the model, the context length and how much retrieved material each answer carries. It has to be estimated against your real volume before the pilot, not after. A demo with twenty conversations tells you nothing about a month with twenty thousand.

3. Hosting and infrastructure — fixed monthly

The vector store, the application, the logs. Predictable and usually the smallest of the four, but it exists and belongs in the estimate.

4. Maintenance — a contract, not a courtesy

Keeping the knowledge current, re-running the evaluation set when the model changes and fixing what breaks. Without a support contract this does not exist, and an agent on stale documentation repeats errors very confidently.

What no honest quote will contain

  • A guaranteed accuracy percentage. Nobody can commit to that.
  • A promised savings figure. Without a comparable baseline of yours it is invented.
  • A flat monthly fee that covers unlimited model usage without a volume cap.

How to estimate the usage line yourself

Take your expected monthly conversations, multiply by the average number of turns per conversation and by the context each turn carries (the question, the retrieved passages and the answer). Apply the provider’s current price per token, then run the same sum for a quiet month and a busy one. The gap between the two is the range your budget has to absorb.

What makes the build line grow

  • Each additional system the agent reads from or writes to.
  • Knowledge that does not exist in writing yet and has to be documented first.
  • Actions with side effects, which need approval steps and audit trails.
  • Extra channels (web, WhatsApp, voice), each with its own handoff rules.
  • A stricter evaluation set, when mistakes are expensive.

How to compare two quotes

Ask both vendors to split the four lines and to state the assumed monthly conversation volume. If one gives a single number, you are not comparing the same thing.

If you want the four lines for your case, our AI agent development service starts with exactly that estimate.

What a bounded agent looks like

Internal demo: a valid answer with sources, a question it cannot support, and a handoff to a person.

Frequently asked questions

Can we start with a small pilot?
Yes, and it is the sensible way to start: one task, one channel and a fixed evaluation set. The pilot measures quality and real usage before committing to the full build.

Do we need to train our own model?
Rarely. Most business agents use an existing model grounded in your own documents and systems. Training or fine-tuning only makes sense with a clear, measured gap that retrieval cannot close.

Who owns the prompts, the data and the code?
That is settled in the contract. In our projects the repository, documentation and access are delivered as agreed, and your data stays in your systems or in accounts you control.

Related services

Want the four lines for your case?

Tell us the task, the sources and your expected volume. We estimate build, usage, hosting and maintenance separately.

Evaluate your AI use case