You can call the OpenAI API from Power Automate, but you should not plan a production flow around a free OpenAI account. ChatGPT plans and OpenAI API billing are separate, while Power Automate connectors may also require premium licensing.
The original article's no-cost promise is no longer a safe basis for a build. Check the current commercial terms in both services before you start.
Verified on 24 August 2026 against Microsoft Learn and OpenAI's platform documentation.
Pick the connection route
Three common routes are:
- an approved Power Platform connector that exposes the required operation
- a custom connector with a narrow OpenAPI contract
- an HTTP call through a controlled service such as API Management or an Azure Function
The exact choice depends on your tenant's data policies, licences and security design. Do not add an external AI connector to the default environment without an admin review.
Set up the provider project
Create an OpenAI API project, configure billing and create a project-scoped key. Restrict who can view or rotate it.
Never paste the key into a Compose action, app formula, SharePoint list, email or screenshot. Use the connector connection or a server-side secret store. A flow exported in a solution must be able to move environments without exposing the provider credential.
Build the smallest useful flow
Start with a manual trigger and synthetic text.
Send a fixed task plus the input to the currently supported provider operation. Parse only the fields your flow needs. Add explicit branches for authentication failure, rate limiting, timeout, invalid response and provider refusal.
Return the output to a draft, approval or test record. Do not let model text directly send a customer response, change a permission, approve money or overwrite an authoritative record.
Check the data boundary
OpenAI says API inputs and outputs are not used to train models by default. Its data-controls page also documents abuse-monitoring retention and endpoint-specific application state. Review the endpoint you use, because not used for training does not mean never retained.
Power Platform data policies can prevent certain connector combinations and classify custom connector endpoints. They are useful guardrails, not proof that the data, provider or output is compliant.
Test before calling it complete
Test a valid request, blank input, oversized input, unsafe content, rate limit, bad key, timeout and unavailable provider. Record the flow run, status code and safe error message.
Then confirm the flow owner, connection owner, premium licence route, budget alerts, retry policy and key-rotation procedure.
A successful test proves that one request returned a response. It does not prove output accuracy, predictable cost or suitability for sensitive business data.
For help designing API flows that survive real permissions and failures, join the Power Automate Builders Space.