Cost
How much does production AI cost in SaaS?
Production AI cost in SaaS breaks into four lines: inference (tokens or GPU hours), retrieval and vector storage, surrounding data movement including egress, and the engineering to keep it correct. For most vertical SaaS copilots the dominant variable is not the model price but tokens per action multiplied by actions per user per month. A single retrieval-heavy interaction commonly runs several thousand input tokens, so the difference between a tight context window and a lazy one changes your bill by an order of magnitude, not a few percent.
The four cost lines
| Line | Driver | Biggest lever |
|---|---|---|
| Inference | Tokens in and out, model tier, call count | Context discipline and model routing |
| Retrieval | Vector store size, queries per action | Chunking, filters, cache |
| Data movement | Egress, cross-region and cross-cloud traffic | Keep inference next to the data |
| Engineering | Evaluation, monitoring, prompt and model churn | Regression suites over manual QA |
Indicative bands
The following are indicative planning figures, not a quote, and they move with your model choice and traffic profile.
- A pilot copilot for a few hundred internal users: typically low hundreds of dollars per month in inference.
- A live customer-facing copilot across a mid-size vertical SaaS base: typically low thousands per month once context is controlled, and five figures per month when it is not.
- Vector storage and retrieval infrastructure: usually a small fraction of inference cost until you index everything indiscriminately.
- Egress is billed per GB and is easy to overlook: roughly $0.09/GB on AWS, $0.087/GB on Azure and $0.12/GB on Google Cloud for standard internet data transfer out.
The number to model before you build
Cost per successful action, not cost per token. Take the tokens consumed by one complete user outcome, including retries and retrieval, and multiply by realistic monthly volume per account. Compare that against the price of the plan the feature sits in. If the ratio is uncomfortable, fix it in architecture now, not in a pricing renegotiation later.
How Datics helps
Our product review includes an inference-economics read: cost per action, expected volume and the levers that keep the feature profitable at scale.
Sources
Related answers
Unit economics
Why AI features can damage SaaS gross margins
AI turns a fixed-cost software business into a variable-cost one. Here is how margin erosion happens and how to price and architect against it.
Read the answerCost
How to reduce AI infrastructure cost in production
Concrete levers for cutting production AI cost in a live SaaS product: context discipline, model routing, caching, placement and GPU utilization.
Read the answerYour customers already expect AI from your product.
Let us make it real without rebuilding what already works.