As a Designer

A core consideration for the design of interfaces that expose LLMs is how unstructured the inputs to AI-assisted features should be.

• Chatbot UIs, which enable users to freely go back and forth with the AI using natural language and expect the AI to remember some previous context. They are more unstructured.

• Button UIs, which only accept specific inputs that might be generated by your SaaS application with a command that embeds a prompt you send to the LLM. They are more structured.

There’s Pros and cons for both, but really it comes down to three main factors: use case, complexity of input / outputs, and abuse potential.

For instance, if the inputs are very complex and you expect the user to want to guide outputs, you might prefer a chatbot-style interface. If the abuse potential is large, however, you might want to restrict how open-ended a task your user can give your LLM.

Note that you don't need a universal policy on this. At LogicLoop, it made sense for us to have both Chatbot and Button UI modes, just in different areas of the product.