Introducing AI content management in Umbraco 17

When you first log in, Umbraco AI is yours to configure.

This is the foundation everything else in this introduction sits on. You install it and get a new AI section, including connections, Profiles, Contexts, Guardrails, and Settings you can adjust.

On its own, Umbraco AI doesn't do much (it's the plumbing, not the tap), so you’ll need to add features to it, such as:

• Umbraco Agents — your own configurable AI helpers, built on top of Umbraco AI.
• Umbraco Copilot — a chat assistant for editors in the Umbraco backoffice.
• Umbraco Prompts — one-click AI assistance for individual Umbraco CMS fields
• Umbraco Automate — automated workflows that run when something happens (similar to tools like n8n or Make, but built specifically for Umbraco).

There's loads more available from the Umbraco Marketplace. Your agency or developers can help install Umbraco AI and recommend the right features for your use case. If you would like to know more about the features you can access with Umbraco AI, talk to one of our Umbraco experts.

You set up one connection per provider, and Umbraco AI lets you use more than one at a time. You can use different providers for different jobs: a cheaper, faster model for routine internal tasks, and a stronger model for client-facing content that needs more care.

In this example, we used Anthropic's Claude, which only requires an API key. Other providers take a bit more setting up; for example, Microsoft Azure AI Foundry is a bit fiddly to fill in, but if your organisation already uses that system, it'll look familiar.

Key settings to note:

• Model — you select from whichever models your connection supports (we picked a Claude Sonnet model).

• Temperature — Thom's recommendation: reduce this down, just to keep the model on track and avoid hallucinations.

• System prompt — you can set one here, but as the next section shows, most of this instruction lives in Context instead.

Within your Context, you can set up content rules.

For our example, our rules were:

•Always use British English spelling and grammar
•No emojis
•Rules for specific content fields
•Image rules: always use WebP format, use descriptive file names, keep file size under a set limit
•Always include FCA-required risk wording where relevant 

That’s it. Set them once, and these rules are picked up automatically by everything you choose to add later, such as Umbraco Prompts, Umbraco Copilot, Umbraco Agents and Umbraco Automate.

In our demo, we set our Guardrail to "block" and set it for "pre-generate"

We then asked Copilot (which I'll introduce in a moment) to take the current page, change everything to US English, and sprinkle it liberally with swear words. It refused to do it. That's a relief 😅. 

In Settings, it's a good idea to set up a default chat profile. This is what Copilot uses if nothing else has been set up. It matters because it closes a gap: even an interaction nobody set up on purpose still runs through your approved connection, your chosen model, and your guardrails, so nothing runs unmanaged by default.

First, we made Bob, a simple example of a suggest-only Umbraco Agent. We wanted him to be like a newer member of the content team, or an external contributor wanting a second opinion on a page before it's published.

In our demo, Thom has full AI configuration control (as an AI Manager might have), so he handled adding Bob's instructions and permissions. I then logged into the CMS as a content editor and asked Bob to suggest some content changes for the page.

Bob came back with a long list of highly relevant issues, thanks to the brand voice Context we added during setup.

Most importantly, Bob's permissions are restricted, so you can action the list at your own pace. This gives content teams a completely risk-free way to get a second opinion, with no chance of an unapproved change going live, without having to switch between different tabs.

Next, Thom set up a second Agent, Dave. Where Bob only flags issues, Dave can act on them, updating any allowed fields like the title, subtitle, meta description and body copy.

Dave uses the same profile as Bob, and largely the same setup, but with different instructions and permissions: he has read and write access.

This kind of Umbraco Agent suits teams who trust their Agent's judgement, want to move faster through higher volumes of content, and are happy with a quick confirm-before-you-publish step.

Umbraco 17 helps make AI governance easier to manage. In the backoffice, every AI interaction is logged, so you can see what the Agent was asked to do and what it produced.

The logs and analytics let you regularly check how your AI Agents are performing, catch problems early, add additional Guardrails if needed, and confirm they're delivering the right results and behaving as expected.

You can see which model was used, how many tokens it used, the exact prompt it received, your content rules, and exactly what the Agent did with it - exactly the kind of detail regulated organisations and auditors expect to see when something needs reviewing. Super handy for keeping an eye on cost for when you’re running AI at scale.

Setting up an Umbraco Prompt to generate alt text

To show you how it works, we set up a prompt in the backoffice and called it "Generate alt text".

First, we added basic prompt instructions: write short, descriptive alt text, under 125 characters, and write it straight into the alt text field.

Since we were just looking for a prompt to generate alt text, we made this available on the right content type and property—in this case, the "image" content type and the "alt text" property. That's all the setup done.

Logged in as a content editor, I found three dots next to "Alt Text", showing me that a prompt is available to generate under the conditions of my Umbraco AI settings.

You then get handy, on-brand suggestions that match the image, already following the guidelines we set up earlier. 

You can see the image shows a calculator and a pen with a spreadsheet behind it, so I’m happy with the suggestion. If you want multiple suggestions to choose from, you can select that option.

Umbraco Prompts also work for any CMS field: SEO descriptions, meta titles, and so on. You could, for example, add a button on any text field in the content section that checks the content against your brand guidelines. Someone can do that in one click, without opening a chat with Copilot. It’s a super quick way to fine-tune your stale content.

You start by setting up your triggers and actions.

Here we've added:

1. Trigger: content is published (optionally scoped to specific content types).

2. Get content: bind to the content key returned by the trigger.

3. Notify editor: a pop-up notification immediately after publishing.

4. Send Slack message: a message to a chosen Slack channel, including the full published URL.

Here's how it's set up: a Publish Alert workflow, triggered whenever content is published, that gets the content, notifies the editor, and posts to Slack.

And there's our Slack message.

This is a simple example, but it shows where Automate is heading. It's really useful for big content teams and larger organisations. Automate can also connect to Umbraco Workflows and Umbraco Forms — handy if you want a form submission or a multi-stage sign-off process to trigger the same kind of notification, without anyone needing to check for it manually.