Nobody can quote an AI product from a one-line description, but the shape of the budget is predictable enough to plan around.
Discovery is small and decides everything
One to two weeks to map the workflow, look at the data, agree what success means and write the evaluation criteria. It is the cheapest part of the project and the one that determines whether the rest is spent well.
Data work is the line people forget
Access, cleaning, permissions and a pipeline that survives the source changing shape. On projects with messy internal systems this is often larger than the model work, and it is the main reason two similar quotes differ.
Build is the predictable part
Interface, backend, integrations and the model layer. For a focused first version this lands in the four to six week range with a small team, which is why an MVP quote clusters around a familiar number rather than scaling with ambition.
Evaluation and hardening
The gap between a working demo and something you can hand to customers is guardrails, refusals, retries, monitoring and the evaluation set. Budget for it explicitly or it gets skipped, and you find out about the failures from users.
Running costs
Inference, vector storage, hosting and observability. For most business tools these are smaller than teams fear, and the bill is driven by how much context you send rather than by request volume.
Caching repeated prompts, trimming retrieved passages and routing easy requests to a smaller model routinely cut this by more than half without a user noticing.
What pushes the number up
- Data that lives in systems without an API
- Compliance requirements that keep data on your own infrastructure
- Several user roles in the first release instead of one
- A workflow where a wrong answer has real cost, which raises the evaluation bar
