AI already builds faster than we can review. Give an agent a brief and it sets up a CMS in minutes: content types, fields, compositions, test content. Impressive, until you realise someone has to check all of it. On every site, after every change. The reviewer has become the bottleneck, and nobody dares to let the AI deliver on its own.

Every CMS team knows the agreements that make a site "ours": SEO on its own tab at the end, global content in one place, the same field limits everywhere. They live in people's heads and wikis. The AI receives them once in the prompt, context drifts, and on site seven the SEO fields end up under Content. Nobody notices until an content editor does, and who is to blame?

In this session I show a different approach. Don't try to make the AI predictable: sandwich it between two parts that are fixed.

A brief in plain language goes in, with an example of the result. AI turns it into a fitness function: rules a computer can check, reviewed once by a human. Your team's agreements become house rules: written once, applied to every project. Then the AI configures the CMS however it likes. How doesn't matter. Finally, a script with no AI in it compares the configuration and the delivery API against the rules. Thousands of checks in seconds, same answer every time. Red goes back to the AI as a precise instruction. Green goes live.

And then I break it, live on stage. Watch the checks catch what a human reviewer would miss, and the AI fix it without anyone explaining what went wrong.

You'll walk away with:
- A way to let AI deliver CMS work on its own, without the review work growing with it
- How to turn your team's unwritten CMS conventions into house rules that are checked automatically
- Why a headless CMS is a testable CMS, and how the brief becomes your frontend contract
- A pattern that isn't tied to one CMS: shown on Umbraco, also runs on any other

Target audience: CMS developers, solution architects, tech leads and agency teams who build and maintain many CMS sites. No AI experience needed; experience setting up content models helps.

Format: talk with live demo. Preferred length 45–50 minutes; a 30-minute version is possible.

Language: English or Dutch are both possible.

Demo: Umbraco 17 with uSync and the Content Delivery API, runs fully local, no internet required. Pre-recorded fallback available.

Demo is a working proof of concept: an AI agent configures a headless CMS from a plain-language brief, validated by four deterministic checks against the exported configuration and the delivery API. Includes a real team spec run through the method unchanged. At our company we use this in practice for a big account, not just a proof of concept.

Speaker background: not experienced but not new at speaking and attendee at meetups, previously spoke at DF23 and twice at Umbraco Community Day.

Concept: this talk is not only for CMS, the same is true for building operation business-driven APIs. Would like to transform the same concept on any topic on request. DDD, data, DevOps, try me ;).

First public delivery is not yet done, would like to try this talk at a meetup.