What is an agentic CMS?
Last updated: 17-11-2026
An agentic CMS is a content management system where AI agents carry out multi-step content work through the CMS's own tools and permissions, while people set the goals and approve what goes live. An agent can audit pages, draft content, update metadata, translate and prepare changes for publishing.
If you've clicked an AI button to shorten an intro, you've used an AI CMS. If you've asked an AI agent to find every page that still mentions last year's product name, draft the fixes and leave them for you to approve, you've used an agentic CMS. The word "agentic" describes how the work gets done. In most cases the software is a CMS that AI agents can work with safely.
In short:
- What changes: AI goes from suggesting text to completing tasks
- What stays the same: people decide the goals, approve changes and own the content
- What it depends on: structured content, clear permissions and an approval flow
Overview
- AI CMS vs agentic CMS: what's the difference?
- Agentic CMS dictionary
- How does an agentic CMS work?
- Agentic CMS vs headless CMS vs DXP
- Built-in agents or an agent-ready CMS?
- Working across tools: CMS content plus your other data
- What can AI agents do in a CMS?
- Benefits and limits of an agentic CMS
- What needs to be in place first?
- How to evaluate an agentic CMS
- How to get started
- Agentic workflows with Umbraco
- Frequently asked questions about agentic CMS
AI CMS vs agentic CMS: what's the difference?
The two terms are often used as if they mean the same thing. They describe different levels of AI involvement.
| AI CMS (CMS with AI features) | Agentic CMS | |
|---|---|---|
| What the AI does | Helps with one step: suggests a headline, writes alt text, translates a field | Plans and completes a task of several steps, then hands it over for review |
| Who drives the work | The editor, field by field | A person sets a goal; the agent does the steps |
| Typical request | "Make this intro shorter" | "Find our outdated product pages and draft updates" |
| Where the AI works | Inside the editing interface | Through the CMS's tools and APIs, inside or outside the editing interface |
A CMS with AI features becomes agentic when the AI can work across pages, tools and data instead of one field at a time.
Agentic CMS dictionary
- AI agent: software that uses a language model to plan and carry out steps towards a goal, with the tools it has been given access to.
- Agentic workflow: a sequence of steps an agent completes, with checkpoints where a person reviews the work.
- AI CMS (AI content management system): a CMS with built-in AI help, such as generating text, alt text or a translation for a field.
- Agent-ready CMS: a CMS that lets external AI agents work with its content and actions in a controlled way, for example through an API or an MCP server.
- MCP (Model Context Protocol): an open-source standard for connecting AI applications to external systems, such as a CMS, an analytics tool or a CRM. Within the permissions they're given, compatible AI assistants can read and change what those systems hold.
- Structured content: content stored in defined fields and types (title, summary, author, product) instead of one large block of text. Agents rely on it to change exactly the right thing.
- Human in the loop: a fixed point in the workflow where a person reviews, edits or approves what the agent has done.
How does an agentic CMS work?
Take a typical task: "Improve the meta descriptions on our 20 pages with the lowest click-through rate." An agentic workflow runs it like this:
- Goal: you describe the task and the limits ("drafts only, don't publish").
- Context: the agent gathers what it needs: the pages, search data and your brand and tone guidelines.
- Plan: it breaks the task into steps and shows you the plan.
- Act: it uses the CMS's tools to read the pages and save new descriptions as drafts.
- Review: an editor checks the drafts, fixes what's wrong and approves the rest.
- Publish: approved changes go live through your normal publishing flow.
- Measure: a few weeks later, the agent compares click-through rates before and after and reports back.
What keeps this under control is the CMS's permission model. An agent can only do what its tools and permissions allow. If the agent's user can create drafts but can't publish, the agent can't publish either.
In Umbraco: permissions are set per user group and can be limited to parts of the content tree. When you connect an AI assistant through the hosted CMS Editor MCP, it signs in as your own backoffice user and can only use the tools your permissions allow. People and agents follow the same rules.
Agentic CMS vs headless CMS vs DXP
| Agentic CMS | Headless CMS | DXP | |
|---|---|---|---|
| What it describes | How content work gets done: AI agents carry out tasks | How content is delivered: through APIs to any frontend | A suite covering the whole digital experience |
| The question it answers | "Who does the work?" | "Where does the content go?" | "Which tools do we run as one platform?" |
| Typical example | "Find outdated pages and draft updates" | Content shown on a website, an app and a screen from one source | CMS, personalisation, analytics and commerce in one suite |
The three can overlap. A headless CMS can be agentic, a DXP can include agents, and a traditional CMS can be agent-ready. You don't need to go headless before you can use AI agents. Structured content matters more. Read more in What is a headless CMS? and CMS vs DXP.
Built-in agents or an agent-ready CMS?
There are two ways to make a CMS agentic, and the difference is who brings the agents.
With built-in agents, the vendor ships agents inside the platform, and you configure and use them in the CMS's own interface. This is common in large suites and DXPs.
With an agent-ready CMS, agents you choose work with the CMS's content. That might be the AI assistant your team already uses, agents your developers build or agents a partner sets up. The connection usually runs through an API or an MCP server.
| Built-in agents | Agent-ready CMS | |
|---|---|---|
| Getting started | Usually faster: agents come preconfigured | You connect an AI tool and set permissions |
| Choice of AI model and tool | Set by the vendor | Your choice, and you can change it later |
| Where governance lives | In the vendor's platform | In the CMS's permissions plus your AI tool's settings |
| Who maintains it | The vendor | You or your partner, with official tools from the vendor where available |
| Fits best when | You want a prescribed setup and work mostly inside one platform | You already use AI tools, combine data from several systems or want to avoid depending on one AI vendor |
Many platforms offer some of both. So ask the vendor: if we change our AI tool or model next year, what do we have to rebuild? With an agent-ready CMS, the answer is usually the connection. Your content and workflows stay where they are.
Working across tools: CMS content plus your other data
MCP is an open standard, so one AI assistant can connect to several tools at once. That's when agentic work pays off most, because the agent can put your content next to the data that shows how it performs and what customers ask.
Sources worth combining with CMS content:
- Analytics and search data: which pages get traffic, which convert and which searches bring people in
- CRM data: which industries and roles you sell to, and where deals stall
- Sales call notes and recordings: the questions and objections prospects raise
- Reviews and support tickets: how customers describe their problems in their own words
- Planning tools: the campaign plan, owners and deadlines, for example in Airtable or on a project board
Many of these tools now have their own MCP servers, including Google Analytics, HubSpot, Salesforce and Airtable. Where there isn't one, an API or an export can often fill the gap.
Example: plan a campaign from real customer questions
- The agent reads the campaign brief and plan in your planning tool.
- It pulls recurring questions and objections from CRM notes and sales calls, summarised without personal details.
- It checks analytics for the pages you already have on the topic and how they perform.
- It drafts a landing page and supporting content in the CMS, saved as drafts.
- A marketer reviews, edits and approves, and the plan is updated with what was done.
Each connection keeps its own permissions, so the agent only reaches the data that its user or account can reach in that tool. Customer and call data often contains personal data, so connect only what the task needs and follow your organisation's data rules for each tool.
What can AI agents do in a CMS?
Agents do best with tasks that are repetitive, follow clear rules and are easy to check:
- Content audits: find outdated, duplicated or thin pages and list what to fix
- SEO and metadata: draft titles, meta descriptions and alt text where they're missing or weak
- Translation and localisation: prepare language versions for a native speaker to review
- Tagging and structure: tag content consistently and spot pages in the wrong place
- Questions about content: "Which pages still mention our old product name?"
Work that needs judgement, such as positioning, legal wording or sensitive messages, still needs a person in charge. The agent can prepare the material.
Benefits and limits of an agentic CMS
Benefits:
- Less routine work: audits, metadata and tagging no longer have to be done page by page.
- Changes at scale: updates across many pages are prepared in one run and reviewed together.
- More consistency: the same rules and tone are applied everywhere.
- Better use of data: agents bring in analytics, CRM and customer data that editors rarely have time to check.
Limits and risks:
- Wrong facts: language models can state things that aren't true. Agents need a source of truth, and people need to review.
- Permission mistakes: an agent with too many rights can change more than you intended.
- Quality drift: output that's fine on each page can make the whole site blander over time.
- Running costs: AI usage is often billed separately, per request or per credit.
- Unclear ownership: if nobody owns an agentic workflow, nobody notices when it goes wrong.
When you don't need an agentic CMS (yet): if you publish a handful of pages a month, run a small site or your content isn't structured, the gain is small. Fix the content model first and use AI help in the editor.
What needs to be in place first?
Agents amplify what's already there. A messy content model gives you messy results, only faster. Check these before you start:
- Structured content: clear content types and fields, so an agent can change a meta description without touching the body text.
- Permissions for agents: treat an agent like a new team member. Give it its own user or role, with only the rights the task needs.
- Approval rules: decide what an agent may save as a draft, what needs approval and what it may never do.
- History and rollback: make sure you can see who (or what) changed something, and undo it.
- A source of truth: product facts, tone of voice and brand rules the agent can use, so it doesn't guess.
- An owner: a named person who is responsible for each agentic workflow and its results.
In Umbraco: document types give content its structure, permissions are set per user group, and saved versions are kept in the history so you can roll back a change. These are the same foundations an agent needs.
How to evaluate an agentic CMS
Ask vendors to show you a real task from start to finish, not a slide. These questions help:
| Question | Why it matters |
|---|---|
| Which AI tools and models can work with the CMS? Can we use the ones we already have? | Decides whether you're tied to one AI vendor |
| What can an agent change, and what needs approval? Can we set this per role? | This is how you keep control |
| Does an agent work through the same permissions as a human user? | Separate rules for agents are harder to audit |
| Can we see everything an agent did, and undo it? | You need a trail when something goes wrong |
| How does it handle structured content and multiple languages? | Agents work best with clear fields and language versions |
| Can an AI tool combine CMS content with data from our other systems, such as analytics, CRM or planning tools? | The most useful agentic work rarely uses the CMS alone |
| What happens if we change AI model or provider? | Models change often, and your setup shouldn't break each time |
| Who maintains the connection: the vendor, a partner or us? | Ongoing cost and responsibility |
How much does an agentic CMS cost?
There's no standard price, because the cost comes from several places:
- The CMS licence or subscription, and whether AI features are included or sold as an add-on
- AI usage: tokens, credits or requests, billed by the CMS vendor or by your AI provider
- Implementation: connecting tools, setting permissions and building workflows
- Maintenance: keeping prompts, connections and rules up to date
When you compare vendors, ask whether AI usage is billed separately and who pays when an agent runs.
How to get started
- Pick one task. Choose something repetitive and easy to check, such as missing meta descriptions or alt text.
- Set the boundaries. Give the agent its own user with draft-only rights, and decide who approves.
- Run it on drafts. Let the agent prepare changes, and have an editor review every one at first.
- Measure and expand. Compare quality and time spent, then widen the scope or add the next task.
Agentic workflows with Umbraco
Umbraco offers both routes: AI inside the CMS, and an open connection for the AI tools you already use. We use this ourselves: Umbraco's marketing team has built an agentic setup that works with the CMS behind umbraco.com.
- AI in the backoffice (Umbraco.AI): editors generate field content with one-click prompts that use the page's context and your brand tone, then review the result before publishing. You choose the AI providers and models, such as Anthropic, OpenAI, Google, Amazon Bedrock or Microsoft Foundry, and different teams can use different providers without code changes. Umbraco.AI and Umbraco.AI.Prompt are available now on the Umbraco Marketplace, and Umbraco Copilot and more agent capabilities are being rolled out. Read about AI for editors.
- Automation with human review (Umbraco Automate): a drag-and-drop automation tool in the backoffice, built from triggers and actions. You can make an automation available as a tool for Umbraco's AI agents, for example to generate alt text or summarise blog posts, and add an approval step so your team reviews AI output before it goes live. Explore Umbraco Automate or read the Automate documentation.
- Your own AI assistant, through MCP (CMS Editor MCP): connect an AI assistant such as Claude, ChatGPT or GitHub Copilot to Umbraco and work with your content in conversation. It can create and edit pages, organise media, translate, run SEO and content health reports and make bulk updates. It signs in as your own backoffice user, so it can only do what your permissions allow, and changes are designed to be confirmed before they run (how that looks depends on your AI tool). It works with Umbraco Cloud and with self-hosted sites. Because it uses MCP, the same assistant can work with your analytics, CRM or planning tools in the same conversation. See what the CMS Editor MCP can do or set it up on Umbraco Cloud.
- For developers (CMS Developer MCP and Engage Developer MCP): developers can use AI coding tools such as Claude Code, Cursor, GitHub Copilot or OpenAI Codex for almost any task they could do in the backoffice, from building content models and document types to cleaning up unused items. The CMS Developer MCP is open source on GitHub. The Engage Developer MCP adds Umbraco Engage analytics, A/B tests, personas and personalisation.
- Governance: profiles, guardrails, logging and tests apply across Umbraco's AI features, and you decide which AI providers your content is sent to. Read about AI governance.
Umbraco is open source and built on .NET, so your team or partner can extend any of this to fit your own workflows. For the technical side, see the agent-ready CMS and the Umbraco AI documentation.