Use cases by function4 use cases

AI features for your own product, built like the rest of your software

Buyers increasingly want AI inside the software they already use rather than another chatbot beside it. Upwork's 2026 in-demand skills report showed demand for AI integration work growing 178% over 2025, against 71% for chatbot development. For a product team, the question is how to add it without leaking data between customers, eroding the margin on every plan, or shipping a demo that falls over in month two.

The builds here are for founders and product teams: an in-app assistant that acts on the user's own data, an MCP server so customers can use your product from Claude, ChatGPT and Copilot, a customer portal where B2B clients find their own orders and documents, and product data that AI shopping assistants can read.

I have built AI into products of my own, including Caldra AI and OmniAPI, so the parts that decide whether these features survive contact with real users are familiar: tool design, tenant isolation, metering and tracing. The architecture decisions below are the ones I would make on day one of any of these builds.

The decisions that matter on day one

  • The user's permissions, never a service account

    An assistant that calls your API with an all-powerful key will eventually show one customer another customer's data. It acts with the permissions of the person using it.

  • Metering from the first release

    Model costs scale with usage, not with seats. Usage is recorded per tenant and per feature from day one, with caps per plan, so pricing can follow cost.

  • Confirmation before anything is written

    Reading is cheap to get wrong. Writing is not. Actions that change data show a preview and wait for the user, and can be undone.

  • Evaluations in CI

    A set of real tasks with expected outcomes runs on every prompt or model change. Model upgrades stop being a leap of faith.

  • Tools designed for agents

    Exposing every REST endpoint as a tool confuses models. A small set of task-shaped tools with clear descriptions works better for assistants and for MCP clients alike.

Frequently asked questions

How do I add AI to my existing SaaS product?

Start with one job users already do by hand in your product, give an assistant a few tools over your own API that act with the user's permissions, add confirmation for writes, meter usage per tenant, and test it on real tasks. A narrow feature that works beats a broad assistant that sometimes does.

What is an MCP server and does my product need one?

An MCP server exposes your product's capabilities to AI assistants through the Model Context Protocol, so customers can use your product from Claude, ChatGPT or Copilot. If your customers already work inside those assistants, and many products in project management, CRM and finance now have official servers, it is becoming an expected integration.

How much does it cost to add AI features to an app?

The build depends on scope, and the pages here show which pricing tier each kind of feature usually lands in. The running cost depends on usage: model calls per active user, context size and which model handles which task. That running cost is why metering and model routing belong in the first release.

How do I stop an AI assistant leaking data between customers?

Make isolation structural rather than a prompt instruction: the assistant's tools call your API with the current user's credentials, retrieval indexes are filtered by tenant at query time, and nothing from one tenant's context is cached for another. Then test it deliberately, with cross-tenant requests in the evaluation set.

Also browse: SaaS · Modernization and rescue · E-commerce · all use cases