Drupal has spent the past two years building AI into the platform itself, through a coordinated initiative with dozens of partner organizations, dedicated funding and a public roadmap. The result is a large and fast-moving ecosystem: hundreds of modules, integrations with every major model provider, and capabilities that range from proven to experimental. For a business running a Drupal site, the useful question is not "what can AI do?" but "which of these are ready for my site, and how would I know?" This article is our answer, based on how we evaluate these tools before they reach a client.
What the ecosystem actually contains
It helps to sort the capabilities by what they do for the people using the site.
Editorial assistance. Tools inside the editor that suggest a summary, propose alternative text for an image, check tone against a style guide, translate a draft, or generate tags from the body. These act on content an editor is already working on, with the editor deciding what to keep.
Automation. Background tasks that process content according to rules: turning an uploaded document into an accessible page, re-tagging an archive, generating structured data, drafting a translation for review. These run without an editor present and their output is reviewed afterward.
Search and discovery. Semantic search that understands meaning rather than matching words, so site search finds the right page even when the visitor uses different vocabulary. Related-content suggestions built on the same understanding.
Assistants. Conversational interfaces for visitors, which we covered in detail when we wrote about adding an AI assistant to a website.
Agents. Systems that take multi-step actions on the site itself: building a view, configuring a content type, restructuring a page. These are the newest and least mature category.
Readiness tiers
The initiative itself is clear about maturity, which we appreciate. The foundation layer, including the core AI framework, the editor integrations, semantic search, and the guardrail and observability tooling, is production-ready and used on many sites. Multi-step automation and the Context Control Center, a shared place for brand voice, audiences and editorial rules that every AI feature draws from, are described as pilot-ready: usable with supervision, improving quickly. Fully autonomous agents that change site structure are experimental.
Our recommendation follows those tiers exactly. We deploy foundation-tier capabilities on client sites where they solve a real editorial need. We pilot the middle tier with clients who have the appetite and a clear use case, behind review. We watch the experimental tier and do not put it on a production site. The tiers move, and we revisit the line twice a year.
How we evaluate a capability
Every AI feature goes through the same questions any module does before it reaches a site, plus a few of its own.
- What is the failure mode? A tool that suggests alternative text and gets it wrong wastes an editor's second. A tool that publishes content unsupervised and gets it wrong damages trust. We match the level of autonomy to the cost of a mistake.
- Where does the data go? Which provider processes the content, under what terms, in which jurisdiction, and is it excluded from training. We choose providers and configurations that keep client content private and, where required, inside the European Union.
- Is the output reviewable? Everything an AI feature produces should be visible, attributable and reversible. Drupal's revision system helps here: a suggested change is a revision an editor can accept, edit or discard.
- What does it cost per use? Model calls are metered. A feature that runs on every save of every node has a different cost profile from one an editor invokes deliberately. We estimate and cap.
- Does it degrade gracefully? If the provider is slow or unavailable, the editor should still be able to work. AI features are additions to the editorial experience, never dependencies of it.
Where it genuinely helps today
The features clients value most are the unglamorous ones. Alternative text suggestions that make accessibility compliance faster. First-draft translations that a human translator refines, cutting the cost of multilingual content. Semantic search that answers the question the visitor meant. Summaries and metadata that editors adjust rather than write from scratch. Each of these saves minutes many times a day, keeps a person in the loop, and fails harmlessly. That is the profile of a feature worth having.
Governance is the real feature
What makes Drupal's approach notable is not any single capability but the governance around them. The Context Control Center means brand voice and editorial rules are defined once and applied by every feature. Guardrails filter inputs and outputs. Observability logs every call so you can see what the AI did, on which content, at what cost. Permissions decide which roles may use which features. This is the same discipline the platform applies to everything else, and it is why we are comfortable putting AI features in front of editors: they operate inside the system's rules rather than beside them. It is also how we apply AI in our own work: useful, supervised and reviewable.
If you are wondering which of these capabilities would help your team, the best start is a look at where your editors spend repetitive time. Tell us what that is, and we will tell you honestly which tier the solution sits in and whether we would put it on your site today.