What does your WordPress AI plugin send to its provider?
A client asks you to add an AI writing tool to WordPress. The demo looks useful. Before you connect it, write down what leaves the site when someone presses Generate.
That answer is often more useful than a long feature comparison. A headline suggestion might need a paragraph. A support assistant might need a conversation containing names and order details. Those jobs deserve different settings and a different review.
AI is moving into familiar editing screens. For example, WordPress.com’s AI Assistant announcement describes tools inside its editor and Media Library, with content used as context. That is a WordPress.com product, not a feature you should assume exists on every self-hosted installation. It is a good reason to ask what an assistant can read before enabling it.
Start with one action
Pick the smallest task the client actually needs: “Suggest three headlines for this public article.” Follow that action from the editor to the provider and back.
Read the plugin’s service documentation, privacy policy and settings. Record the plugin version and the date you checked. Look for a specific description of the information sent, rather than treating “AI-powered” as an explanation.
For plugins listed on WordPress.org, the Plugin Directory guidelines require documentation of external services and address consent for external connections. They also describe service-related exceptions. Directory inclusion does not answer every question about your client’s particular use.
If the documentation leaves a gap, send the vendor a narrow question: “When I generate a headline, do you receive only the selected text, the entire post, or additional site content?” Save the answer with the project notes.
Fill in this worksheet
Use one copy for each feature you intend to enable. A chatbot and an image generator in the same plugin may need separate entries.
- Action: What does the person click, and can the task also run automatically?
- Input: Which text, images, attachments or site records are included? Does the tool add conversation history or page context?
- Destination: Does the request go directly to the model provider, through the plugin vendor, or through another service?
- Stored copies: Where are prompts, responses and diagnostic logs kept? What retention and deletion options are documented?
- Provider settings: Which account and service terms apply? What does the provider say about model training for that exact service?
- Access: Which WordPress roles can use the feature or read its saved history?
- Exit: How do you disable future requests, remove saved history and revoke the connection if needed?
Write “unknown” wherever you lack evidence. That gives the next person a real question to resolve instead of a reassuring guess.
The WordPress Plugin Handbook’s privacy guidance recommends limiting collection and access, documenting sharing, and considering retention beyond the site itself. Apply those principles to the actual workflow: send only the material the task needs, and identify who can retrieve it later.
A worked example: headlines for a client article
Suppose an agency wants headline ideas for a public guide about garden lighting. The editor also has a private draft containing an upcoming campaign budget. This is an illustrative review, not a test of any particular plugin.
The agency’s desired boundary could be:
Send the selected garden-lighting paragraph and my headline instructions. Do not include other drafts, customer records or private editorial notes. Save the chosen headline in WordPress. Tell us separately whether the service retains the request or response.
That statement becomes a checklist for the vendor’s documentation and a test on a disposable site with synthetic content. Use an invented private draft, not real client material, to investigate what the feature includes. If the documented behavior requires the whole post, decide whether that is acceptable before enabling it on private drafts.
A prompt asking the model to ignore private data is not a substitute for controlling what the integration sends. Check the plugin’s configuration and implementation where possible. If you cannot establish the boundary, keep the feature limited to material already approved for sharing.
Separate permission to read from permission to change
Reviewing data transfer does not tell you whether the tool can publish an article or edit a template. Record those abilities separately, and decide which actions require a person’s review.
For team access, start with our guide to giving WordPress clients the permissions they need. If the assistant produces executable code, also use the AI-generated snippet review checklist before installing it.
Keep the completed worksheet beside the site’s handover notes. Revisit it when you change providers, enable another AI feature or install an update that changes data access. The next person managing the site should be able to see what was approved, what was verified and what is still unknown.
Have a correction or a workflow we should cover? Tell the IrmaWP team.