Designing services for an AI conversation

The task before the channel

A conversational interface does not need to reproduce an entire website. Our starting point is one task that already has value: finding a resource, comparing options or preparing a request. Ask what the person needs to accomplish and which parts should remain in the existing product.

OpenAI introduced apps in ChatGPT and the Apps SDK as a way to bring tools and interfaces into a conversation, using MCP. The original OpenAI announcement explains that starting point. For a new implementation, consult the current tool documentation.

A brief for the first use case

• Need: what the person wants to resolve and how they do it today.
• Input: what they provide and what the system can retrieve.
• Result: what they receive, how they check it and what happens next.
• Boundary: what the tool must not do and when it should return control.
• Owner: who maintains the service, its data and its errors.

Example: preparing a catalogue selection

A first trial could help find products from explicit requirements. The response presents catalogue options and their differences, with links to the products. If an important requirement is missing, the experience asks about it instead of filling the gap with an assumption.

This example does not require placing orders or making purchases in the conversation. Reading information and changing a system are separate journeys; the second needs permissions, review and confirmation of the action.

Test what goes wrong too

Before expanding the trial, review ambiguous requests, unavailable products, incomplete data and connection errors. People should understand what was found, what could not be checked and how to continue on the website or with the team.

The decision to continue should depend on whether one task becomes clearer and more useful, not on the number of connected features. We can help define that trial and connect it to your tools.