Generative AI development in saas
In a SaaS product the model is a feature, and it is judged like one: does it work, how fast is it, what does each request cost, and does it hold up for every customer.
Where generative AI fits
- In-product assistants that answer questions from your documentation and the user’s own data.
- Natural-language search and filters over product data.
- Drafting and summarization features, such as emails, reports, and ticket replies.
- Agents that carry out actions in the product on the user’s behalf, with a confirmation step.
What shapes the build
Latency and unit cost
Model choice, caching, and prompt length set both. A feature whose cost per request is unknown cannot be priced.
Tenant isolation
Retrieval must return only the requesting customer’s data, every time.
Testing before each release
A model or prompt change can shift behavior without any code change. A fixed evaluation set run before release catches it.
Provider dependence
Design the feature so you can change models without rebuilding it.
What we ask first
- What should users be able to do that they cannot do today?
- What can one request cost before the feature stops making sense?
- Which data is per customer, and how is it separated?
- How will you ship the feature gradually, and roll it back?
Services that apply
This page describes the field. We have not published a case study in saas yet. The case studies page lists the work we have.