A client came to me last month, a high-volume WooCommerce shop, convinced they needed a custom “agentic AI” layer. They wanted something that could process refunds, check inventory, and talk to their fulfillment API. So I started sketching a Python middleware to bridge OpenAI and their WordPress database, because that looked like the only safe place to put the logic. In hindsight it was a nightmare waiting to happen.
I was halfway through the draft when the State of the Word 2025 keynote went out. The bridge I was about to build was a brittle answer to a problem the WordPress AI Roadmap is standardizing into core. That is thousands of dollars of technical debt I nearly signed my client up for, purely because I had not checked where core was heading first. Over-engineering around a gap that closes in six months is an expensive way to look thorough.
What the WordPress AI roadmap standardizes
WordPress 6.9 launched live on stage in San Francisco, which was a fun bit of theatre, but the part that mattered was the set of “building blocks” for AI. Plugin developers have been rolling their own AI implementations for years. Some call OpenAI directly, others sit on LangChain, and plenty of them are held together with hope. Nothing is consistent from one plugin to the next.
Matt Mullenweg and the team framed all of this as WordPress moving toward an “agentic web.” Two pieces have shipped: the Abilities API and the WP AI Client. The Abilities API lets you describe what a plugin can do in a format a machine can read, so instead of writing a custom prompt for every specific model, your plugin tells any agent: “Hey, I can manage inventory, here’s the endpoint.”
The WP AI Client is the abstraction layer sitting on top of that. You write your prompts once and it stops mattering whether the site owner runs Anthropic, Gemini, or a local Llama instance. That is the point where I killed the Python build. Registering an ‘Ability’ and letting core route the request means far less custom code to maintain. The WordPress.com recap of the keynote goes deeper on the technical shifts: https://wordpress.com/blog/2025/12/04/state-of-the-word-2025-recap/
Registering an ability
If you write custom functionality, your code now needs to be legible to AI agents. This is roughly how we prefix and register a capability against the conceptual structure of the new APIs.
/**
* Conceptual registration of a custom ability for an AI agent.
* Prefixing with bbioon to ensure clean namespace.
*/
function bbioon_register_inventory_ability() {
if ( ! function_exists( 'register_block_ability' ) ) {
return;
}
register_block_ability( 'bbioon/inventory-manager', array(
'description' => 'Checks stock levels and updates inventory counts.',
'parameters' => array(
'sku' => array( 'type' => 'string', 'required' => true ),
'quantity' => array( 'type' => 'integer' ),
),
'callback' => 'bbioon_handle_inventory_request',
) );
}
add_action( 'init', 'bbioon_register_inventory_ability' );
function bbioon_handle_inventory_request( $params ) {
// Logic to talk to your fulfillment API safely
return array( 'status' => 'success', 'message' => 'Inventory synced.' );
}What this changes for a business site
WordPress is being pushed from “content management system” toward “operating system for the web.” Today 56% of WordPress sites run in a language other than English, and adoption keeps climbing among the top 1,000 sites, so the infrastructure underneath has to hold. Hence the 24-hour safety window on auto-updates and the wider WordPress Playground.
If you own the site, ask any developer pitching a “custom AI solution” how it lines up with the WordPress AI Roadmap. A system that cannot talk to other plugins or to future core updates leaves you stuck. Standard APIs are duller to sell, and they are the ones that still work after the next big update.
This gets complicated fast. If you are tired of debugging someone else’s mess and just want the site to keep working, drop my team a line. We have probably seen it before.
Are you wiring AI into your workflow this year, or waiting for 7.0 to put it in core?
