Your Next CMS Is a Bet on How Your Organisation Will Work

When you sign off a CMS replacement, you are committing budget, moving content and asking people to change how they work. You need that investment to keep paying off after launch.

AI adds another question to that decision: what will this platform let your team do next?

The opportunity is better search, fewer repetitive editorial tasks and tools that work across your systems. The risk is choosing a platform that makes each improvement an expensive integration project, or leaves you waiting for a vendor's roadmap.

That is why I would make this a central procurement question: how strong is the CMS's AI ecosystem, and how easily can our team extend it?

Start with the work people want to improve

On a project we're currently working on, we're seeing benefits at two levels. The first is everyday editorial work: generating image alt text, SEO metadata and similar supporting content. These quality-of-life improvements give editors a useful starting point.

The second is connecting an external assistant through MCP, the Model Context Protocol, to update content through a conversation in ChatGPT or Claude. People can work from the assistant they already use. I've written separately about those connections and the permission checks they need.

These are observations from ongoing work, rather than measured productivity gains. They suggest a practical way to evaluate the investment through three lenses: ecosystem, extensibility and control.

1. Ecosystem: who will keep it useful after launch?

Start with the projects the proposed solution actually depends on. Are they maintained? Who contributes fixes? Does the documentation explain how to extend them? Are supported releases available for your CMS version?

Then ask for examples with enough detail to assess: what ran in production, what required custom development, what people still reviewed and what happened when processing failed.

Drupal has evidence worth examining. Its contributed AI and AI Agents projects have recent stable releases covered by Drupal's security advisory policy. AI Agents documents configurable agents, selectable tools and an explorer for inspecting their behaviour.

In June 2026, the Drupal AI Initiative reported 32 participating organisations and more than 50 contributors. That indicates breadth of participation, although the maintainers of your specific dependencies still matter.

There are concrete applications too. Chicken describes building Southwark Council's AI-powered PDF importer. It is a useful delivery example, although a supplier account cannot establish likely results for your organisation.

2. Extensibility: what will the next improvement cost?

Ask a supplier to explain what your next improvement would involve: configuration, a reusable integration or custom development. That distinction matters to the budget and the team maintaining it.

Drupal's content types, fields and taxonomy give information an explicit structure. Its AI Agents framework supports agents with selectable tools. Those foundations are worth evaluating across several kinds of work.

Help people find the answers you already have

AI Search integrates semantic search with Drupal's Search API, using embeddings and vector databases to retrieve content by similarity. It can also supply retrieved content to an assistant answering questions.

For a membership organisation, that could help someone find guidance using their own terminology. Test real member questions, source attribution, restricted content and missing answers. Measure whether people find useful information. The standalone project currently has alpha releases, so budget for evaluation and integration work.

Give editors a shared starting point

Context Control Center manages context as Drupal content: brand guidance, editorial standards and other instructions, with revisions and moderation. Context can be scoped to particular uses, languages, sections or entities, with agents configured to receive relevant material.

That gives the team a shared place to maintain guidance. The potential reward is less repeated briefing and correction. Test whether the right context reaches each integration and improves outputs; supplying a rule does not guarantee compliance. CCC is currently at release-candidate stage.

Reduce the handoffs in routine work

AI Integration – ECA connects AI operations to Drupal's event-condition-action framework. Site builders can combine events, conditions and AI actions in workflows; the project has a stable release.

Consider an incoming publication: suggest a category, draft a summary, flag missing information and send it for review. Evaluate the whole process, including connections to document stores, failed steps and recovery. A quicker first draft is useful only if it does not create more work downstream.

Migration is another example. Drupal's Migrate API separates extraction, transformation and saving into source, process and destination plugins. An agent could inventory old content, propose mappings and flag exceptions. Approved mappings can become repeatable migration rules, with conventional checks validating imported records. The complete process still needs project-specific engineering.

3. Control: can you change direction without starting again?

Model choice matters because requirements change. Drupal AI offers a common integration framework with separate provider modules. Its provider matrix documents supported operations and shows a mixture of stable and prerelease integrations. Choice exists, but capabilities and maturity vary.

Test alternative models against the same tasks. Compare output quality, tool use, cost and failures. Check where content is sent, what the provider retains and what changing providers would require.

Permissions deserve equally concrete scrutiny. Drupal's JSON:API respects entity and field access and applies validation constraints when modifying data. Those guarantees describe that API path; every custom agent tool needs its own enforcement checked.

Drupal's Content Moderation separates authoring from publishing through review states and permissions. I would use that foundation for reviewing generated content, and explicitly design approval for other consequential actions. Record what was proposed, approved and executed.

Some surrounding infrastructure remains immature. The MCP Server project's current releases are beta releases. External agent connections deserve a dependency-by-dependency assessment, including tests that blocked actions are actually refused.

Leave room for capabilities you cannot specify today

TypeSafe's Jev, introduced in September 2026, returns structured decisions with probabilities rather than generated prose. Classification, routing and scoring are the relevant possibilities here.

Models built for these tasks could make frequent background checks more practical: choosing an editorial queue or flagging a draft for review. That is potential to test. Structured output can still contain the wrong decision; confidence scores need evaluating against your own examples.

This is why extensibility matters. A CMS should let your team add a different kind of model to a useful workflow without rebuilding the whole system around it.

Make the decision with your own work in front of you

Ask shortlisted platforms to demonstrate search, a context-aware editorial task and a workflow using representative content. Include failure cases and blocked actions. Measure review effort and maintenance needs alongside output quality.

Drupal deserves evaluation where structured content and custom integration matter. Its foundations and developing AI ecosystem offer room to extend, with implementation and maintenance costs to weigh against the benefits.

You do not need to predict every AI capability your organisation will want. You need evidence that the platform, the ecosystem and your delivery team can support the next useful change at a cost you can justify.

Cover photograph by Prescott Horn on Unsplash, used under the Unsplash Licence.