A slow WordPress site is not an AI problem. It is years of plugins, a page builder that owns the content, a theme that copies itself into child overrides, and a field map nobody wants to print.
WordPress modernisation is the work of making that site fast, editable, and honest again. AI can help the team see the mess. It cannot replace the rebuild, the performance pass, or the editorial training. Treat it as a surveyor, not as the contractor.
What modernisation actually is
The jobs are mundane and specific:
- List every plugin, who needs it, and what breaks if it goes.
- Find the builder shortcodes, widgets, and nested HTML that trap copy.
- Map ACF groups to templates so you know what must survive a theme change.
- Separate “this query is slow” from “this asset is huge”.
- Plan a migration that does not empty the editor on Monday.
Humans already do this. It takes time because the evidence is scattered: wp_options, wp_postmeta, a 6,000-line functions.php, three abandoned mu-plugins, and a README from 2019.
That scatter is where a coding agent earns a day. It can grep, inventory, and draft a map. It cannot decide that Elementor can go, that a membership plugin is load-bearing, or that legal still needs that PDF embed.
Where AI helps on day one
Plugin inventory with a hypothesis. Given a list of slugs and versions, an agent can group “caching”, “SEO”, “forms”, “page builder”, and “unknown”. A developer still opens the site and confirms. The value is a first pass that is complete, not clever.
Field and template mapping. Dump ACF field groups (or load Local JSON) and grep the theme for get_field, the_field, and have_rows. The output is a table: field name, group, files that call it, suspected unused. That table is the input to a later FieldCraft run or a theme rewrite. The agent must not rename fields. It must not write UPDATE statements.
Builder archaeology. Search posts for builder signatures, shortcodes, and leftover <!-- wp: comments mixed with classic content. You want a count of trapped pages, not a generated theme.
Performance suspects. From a crawl, a Query Monitor dump, or a list of enqueued scripts, cluster “we should look at these twelve handles”. Core Web Vitals still get measured on a real device. The agent is not Lighthouse.
Risk notes. “This plugin has not had a release in four years” is a useful flag. “Replace it with this other plugin the model likes” is usually noise.
Keep the agent on a local or staging copy. Production stays read-only. If you connect MCP at all, use a read-heavy, non-admin setup.
Where AI actively hurts
Autogenerating the new theme. A generated theme that “looks like the old one” will not match the ACF map, the redirects, or the editorial habits. You will spend the budget teaching the generated code what the old site already knew.
Bulk rewriting content into Gutenberg. You can convert. You should not let a model paraphrase every page while converting. Modernisation is not a content refresh unless the client bought a content refresh.
Installing a new AI plugin as part of the rescue. Adding a chat widget or an SEO generator to a site you are trying to slim down is how the plugin count goes up during a project whose metric is plugin count down.
Letting the agent apply plugin and option changes on production. That is not modernisation. That is unreviewed change. Staging exists so you can break a copy.
Trusting “unused” from a grep. A field can be loaded dynamically. A plugin can be required by a legal process that never touches the theme. Unused is a hypothesis until an editor and an engineer agree.
A sequence that works
- Copy production down. Database and uploads. Note PHP version and object cache. Do not start on live.
- Inventory. Plugins, theme, mu-plugins, cron, transients that look like caches. Agent drafts; human labels keep/drop/replace.
- Map content. Post types, ACF groups, builder usage, media that is actually referenced. Agent drafts the table.
- Pick the smallest cut. Remove the obvious dead plugins on staging. Measure. This is a person with a backup, not a loop of tool calls.
- Performance pass. Images, queries, scripts, caching. Tools like Optimise belong here if the problem is off-origin weight. AI does not.
- Architecture. Move custom logic out of the builder and out of throwaway snippets. This is engineering. An agent can sketch a plugin; a human owns the merge.
- Editor path. If ACF schema must move, that is a migration with preview and rollback — FieldCraft or an equivalent — not a chat instruction.
- Only then talk about AI features. If the site still needs a draft assistant or moderation, that is a separate production decision, after the site is thin enough to host it.
Clients sometimes want to skip to step 8. The honest answer is that a 40-plugin builder site will not become a good AI platform by installing one more plugin. Make WordPress boring again, then add a constrained assistant if there is a job for one.
What we still do by hand
We still talk to the people who edit the site. We still watch a form submit. We still check the booking or checkout path on a phone. We still decide whether to stay on WordPress or to move a piece of it.
AI does not sit in those meetings. It sits in the repo, turning a day of archaeology into a morning, so the meetings are about decisions instead of missing field names.
That is the only productivity claim worth making. Not “AI rebuilt your WordPress”. Not “the site now writes itself”. A faster map, a cleaner cut list, and an engineer who still ships the modernisation.
What “done” looks like
A modernised WordPress site is not a new theme with the same 40 plugins. It is a site where:
- An editor can change a property page or a landing page without opening a builder canvas that takes 20 seconds to load.
- The plugin list fits on one screen and each entry has an owner.
- ACF groups match the templates, and unused groups are archived rather than left “just in case”.
- Mobile LCP is a number you would show a client, not a number you hide.
- Redirects, forms, and the primary conversion path were tested on staging by a person.
AI can draft the punch-list that gets you there. It cannot sign it off.
The rule
Use AI to see a WordPress estate. Use people to change it.
If a vendor sells modernisation as an AI product, you are buying a generator. Generators do not reduce plugin debt. They add a dependency. Keep the model in the survey. Keep the rebuild in engineering.