Start with the information flow
Consider a sales team preparing for a customer call. Relevant information may be spread across CRM notes, support requests and order records. An integration could assemble a draft briefing from approved sources for the employee to check. This is an illustrative use case: the right design depends on the APIs available, record quality and permissions in your systems.
Connect capabilities to controlled actions
OpenAI supports tools such as function calling and remote MCP connections. These provide ways to connect model interactions with application capabilities. Our integration work starts by defining the actions a workflow actually needs. Reading an order status and changing an order are different permissions, so they should be considered separately in the design.
Plan for incomplete data and failed requests
A useful integration must handle more than the successful path. Discuss how it should behave when a record is missing, a service is unavailable or two sources disagree. We can scope validation, user feedback and a handover to a person around these situations. Important business updates should have clear approval rules and a way to trace what happened.
Prepare for an integration discussion
Bring a list of applications, the process you want to improve and examples of the input and desired output. Include who owns each system and whether an API is available. Together, we can identify the smallest useful integration, the work needed around it and how it fits the broader software roadmap.
Source and further reading: OpenAI: using tools