Use your own OpenAI API key to provide replies in PD AI Assistant. This guide covers a direct OpenAI connection; a custom gateway needs its own compatibility checks.

Prepare the connection

Have PD AI Assistant installed and accessible to your administrator or staff group. Open the PD Modules plugin in your staff area, choose PD AI Assistant, then open Settings → Provider & connection. Note the currently active provider before changing it: saving a new connection changes the engine used across the assistant, including enabled customer chat and automated work.

Use an API key for the intended project. Its permissions and project role must permit model requests; listing models alone is insufficient. See OpenAI’s permissions guide. Keep the secret out of screenshots and support messages.

Save OpenAI as the active provider

1. In Choose a provider, select OpenAI (your own key). This previews its form; the selection alone does not switch the active provider.
2. Enter the key in API key. If a key is already saved, enter a replacement only when you intend to change it; leaving the field blank keeps that saved key.
3. Leave Custom AI proxy URL (optional) blank for a direct connection. If it already contains a gateway address, confirm whether you intend to replace that connection before clearing it.
4. Choose an available Model from the list that your provider account can use.
5. Click Connect OpenAI, then check the active provider and model shown at the top of the page.

Check a reply and knowledge separately

Open Assistant and ask a simple, non-sensitive question that requests no customer lookup or action. Check that a useful reply arrives. The Connected label reflects saved configuration; it does not prove a successful model request.

If you use custom Knowledge, check indexing separately with a short, harmless reference. Replies and indexing use different model operations, so success with one does not prove the other. When moving from another provider, use the knowledge replacement guide for existing references.

If it fails

Check that the key belongs to the selected provider, the model is available to the account, and the provider reports usable quota and access. Review the redacted error before repeating a failed request. Do not select a random model or broaden unrelated permissions to hide the error. Retain the provider, model, original attempt time and redacted message while your hosting team checks the failed operation.

See the PD AI Assistant Blesta guide for related tasks, or explore PD AI Assistant.