I once spent three weeks bolting a permissions layer onto custom REST endpoints so an AI agent could not wreck a client’s product catalog. The Abilities API landing in WordPress 6.9 is the standard that job was missing.
A client’s AI recommendation engine turned out to be raw curl calls dropped into single.php, pinned to a deprecated API version. Core’s PHP AI Client and Abilities API work is the standard that kind of build has been missing.
A client’s AI plugins each shipped their own settings screen and each asked for API keys they had already entered. The Core AI check-in from October 8th, 2025 covers both fixes: categories on the Abilities API, and one provider settings screen.
An off-the-shelf AI plugin left a client with a slow backend and alt text that had nothing to do with the images. The core AI discussions point somewhere better: content summaries, post titles and image descriptions behind one PHP AI Client.
A client wanted AI features on his WordPress site and every plugin used a different API. The new WordPress AI Client SDK gives plugins and themes one consistent, WP-native API for multiple providers.
A client spent six months stuck on API plumbing before any AI features shipped. The lesson: build WordPress AI on a proper AI Client SDK, with a standard way to register providers, before chasing features.
A client’s AI prototype broke because their site ran PHP 7.2 while our code used PHP 7.4. Here is how the WordPress Core AI team handles the same version gap, and how to build for it.
A client inherited an AI setup that was really just API calls held together with duct tape, and it kept breaking. Here is the architecture I use instead: provider interfaces and DTOs that let you swap AI models without rewriting everything.
A real question from WordPress Core AI work: should a shared AI adapter ship as a canonical plugin or a Composer library, and why that architecture choice shapes maintenance later.
Tying a WordPress site to one AI provider’s SDK makes switching painful later. This walks through an abstraction layer with per-provider adapters so you can change models without rewriting your core code.