Your CMS Has AI. Your Marketers Have Claude. Nobody Has Connected Them.
DXP vendors built AI agents. Marketers adopted Claude and ChatGPT. Almost nobody has connected the two. Here is the architecture that bridges them without new tools.
Most enterprise marketing organizations I work with are funding two AI programs at once, and they rarely realize it. One lives in the platform budget: the agents, copilots, and orchestration features their CMS vendor shipped over the last eighteen months. They rushed to build those in the last eighteen months to keep up, and some did a genuinely good job while others fell well short. The other lives in their marketers' browser tabs, where ChatGPT and Claude have quietly become part of the daily routine.
Those two investments almost never touch. Closing that gap is now an architecture problem rather than a tooling problem, and it is the work my team at HT Blue has been focused on all year.
Two AI Investments Running in Parallel
On the CMS side, every major vendor has gone all in. Sitecore folded XM Cloud, Content Hub, Search, Personalize, and CDP into SitecoreAI, with Agentic Studio at its center. Optimizely went further at Opticon this year, retiring the DXP label entirely and relaunching Opal as the Optimizely Agent Platform, with its chat assistant now called Mark. Adobe is running Agents in AEM as a beta program, and Sanity has built AI tooling directly into Studio and its Content Lake.
On the people side, adoption has outrun every enterprise rollout plan. Social Media Examiner's 2026 AI Marketing report found that 73% of marketers now use AI daily, with ChatGPT at 81% adoption and Claude at 65%. Ahrefs' survey of marketing and SEO professionals, published at the end of September, found that 93% use AI for content and put Claude ahead of ChatGPT among those users, 83% to 76%.
The problem is the workflow is still clunky. Those marketers draft a campaign brief in Claude, pull numbers from an analytics export, rework landing page copy in ChatGPT, and then open the CMS and paste everything in by hand. It requires copy and paste and a set of disconnected tools, when it could all be brought together into one comprehensive workflow that is genuinely easier to use.
To be clear, I am not suggesting the CMS platforms have closed this gap. Most have not. Some fall short outright, and others force you into the agents built inside their own tool. A few of those built-in agents are little more than wrappers on an AI, where you still end up copying and pasting anyway, so they have not completed the loop either. That is the real issue. Nobody has completed the loop.
The CMS's own agents sit in a separate interface marketers rarely open, because learning another tool is the last thing a marketing team has time for.
In distributed systems terms, the marketer has become the integration layer. Every handoff between the AI they use and the platform they publish to runs through a person doing copy and paste. That is the slowest, least reliable, and least auditable message bus you can build.
The Bridge Already Exists. It Is Called MCP.
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic, that lets an AI assistant connect to external systems through a defined set of tools. Think of it as an API contract written for AI clients. The platform describes which actions are available, the assistant calls them on the user's behalf, and the platform enforces who is allowed to do what.
I argued earlier this year that open protocols beat walled-garden AI features. What changed since then is that every CMS we implement now ships an MCP server, and they all target the chat tools marketers already use:
- Sanity runs a hosted MCP server and ships official connectors for both Claude and ChatGPT, so editors can query, create, and stage content from the chat window they already have open.
- SitecoreAI offers the Marketer MCP server, which connects from Claude Desktop, Microsoft Copilot Studio, Cursor, and other MCP clients and can create and save pages in SitecoreAI.
- Optimizely publishes MCP servers for Analytics and Experimentation that authenticate through the Agent Platform and work in Claude.ai and ChatGPT, so marketers can question dashboards and experiment results in plain English.
- Adobe Experience Manager as a Cloud Service offers Content MCP servers in beta, with separate read-only and read/write servers covering pages, Content Fragments, and assets for Claude, ChatGPT, Cursor, and Copilot Studio.
So the plumbing exists. That is exactly why so many organizations assume the problem is solved. It isn't.
Turning On a Connector Is Not an Integration
After two decades of building enterprise systems, I have learned that the hard part of any integration is never the protocol. It is everything the protocol assumes you already got right. MCP is no different. These are the places where we see connections fail in practice.
The content model decides how smart the agent can be
An agent can only reason about the structure your platform exposes. If your pages are built from large rich text blobs, the assistant sees a wall of markup with no idea which part is the headline, the offer, or the legal disclaimer.
Well-modeled content with typed fields, reusable components, and clear references gives the model something to work with. In our experience, content model quality is the single biggest predictor of whether an AI connection is useful or merely impressive in a demo.
Permissions have to be designed, not inherited
Most of these servers authenticate with the user's own credentials through OAuth, which is the right default. It also means an over-permissioned author account becomes an over-permissioned agent.
Adobe's decision to ship a separate read-only server signals how vendors expect you to roll this out: start with retrieval, prove the value, then grant write access deliberately and by role.
Nothing should reach production without a checkpoint
In platform engineering, we put circuit breakers in front of anything that can fail loudly. The equivalent here is your publishing workflow. Agent changes should land as drafts, releases, or workflow states that a person approves, never as live edits.
Every platform above supports this. It still has to be configured intentionally, and the defaults are not always conservative.
The assistant needs your context, not just your data
A connector gives Claude or ChatGPT access to your content. It does not tell them your brand voice, your component rules, your SEO standards, or which regions require legal review. That context lives in project instructions, skills, and well-documented schemas, and it is what separates on-brand output from generic output.
Each platform has its own sharp edges
Some servers are generally available and some are still in beta. Some support Claude and ChatGPT on the web, while others work best in desktop clients. Optimizely's servers require the Agent Platform to be enabled on your account. These details determine what a marketer can actually do on Monday morning, and they change almost monthly.
What This Looks Like for a Marketer
When the architecture is right, the experience is almost boring, which is the point. A campaign manager working in Claude can do something like this without leaving the conversation:
- Ask which landing pages underperformed last quarter, with the answer pulled live from analytics.
- Draft a revised version of the weakest page and save it as a draft in the CMS, built from approved components.
- Review experiment results on the current variant before deciding what to test next.
- Stage the new page in a scheduled release and route it to the web team for approval.
No new interface to learn. No exported spreadsheets. No pasting between tabs. The marketer stays in the tool they already chose, and the CMS stays the governed system of record. We have seen how far this can go on a lean stack in how Claude is replacing the CMS for some marketing teams.
None of this makes a CMS's native agents irrelevant. Optimizely's Agent Platform is strong for closed-loop experimentation inside Optimizely One, and SitecoreAI's Agentic Studio has a place for repeatable, admin-defined workflows. We have written about where Agentic Studio and the Marketer MCP server each fit. The point is that marketers should reach those capabilities from where they already work, rather than being asked to move.
Why It Takes Both Halves of the Expertise
Most firms are strong on one side of this problem. CMS integrators understand content models, publishing workflows, and platform quirks, but often treat AI as a feature checkbox. AI consultancies understand prompts, agents, and orchestration, but have rarely untangled a fifteen-year-old content tree or an enterprise permission model.
Connecting the two requires both at once. At HT Blue, our engineers build on and migrate to Sanity, Optimizely, Adobe Experience Manager, and SitecoreAI every day. Our team is also certified through Anthropic's Claude certification program, which launched alongside the Claude Partner Network in March.
That combination lets us design the bridge from both ends: schema and governance work on the CMS side, and connectors, skills, and agent behavior on the Claude and ChatGPT side. It is the foundation of our AI content operations practice.
It is also why we are direct about when this is the wrong investment. If your content model needs a rebuild, or your organization has not settled who owns publishing approvals, connecting AI will amplify those problems rather than solve them. In those cases we fix the foundation first.
Where to Start
- Map what your marketers already use. Find out which assistants your team relies on today and for which tasks. That is your target interface, not a new one.
- Check your CMS's MCP status. Confirm which servers exist for your version, whether they are generally available or in beta, and which clients they support.
- Start read-only. Connect analytics and content retrieval first, and let marketers question live data before they can change anything.
- Audit the content model. Identify the page types and components agents will touch, and restructure the ones that are unstructured.
- Define approval gates before granting write access. Decide which changes land as drafts, who approves them, and how agent actions are logged.
- Measure the handoffs you removed. Track time from brief to published page and the number of manual copy and paste steps you eliminated.
Your CMS vendor spent the last two years building AI. Your marketers spent the same two years learning to use it. The organizations that pull ahead in 2027 will be the ones that stop treating those as separate projects. If you want help connecting them, talk to our team.
