I let an AI into the CMS. Nothing burned down. Yet.

vivenu builds ticketing infrastructure - the platform underneath the box office. It's the kind of software where an outage isn't an inconvenience, it's an empty stadium gate. My corner of it is the marketing site: the front door, in three languages.
The way it started is ironic in itself: their recruiter found me through an LLM. Somewhere out there, a model read what the web has to say about me and decided I was worth a call. I've since spent a good part of the engagement on GEO - generative engine optimization, making vivenu just as legible to machines as I apparently was. An AI recommended me to them; now I make sure AIs recommend them to everyone else. Fair trade.
The first contract was three months of cleanup. The site's technical infrastructure had grown complicated over the years - security, platforms, tooling - and it needed auditing, updating, and new functionality of the kind no out-of-the-box solution covers. It took two months instead of three. And since I've never been good at working only the ticket in front of me (I think about the whole product whether anyone asked me to or not) by then the conversation had already moved on from fixing things to where the site should go next. The contract got extended by another two months, and the job changed with it: future-proofing instead of repairing. Efficiency tooling for the people who update the site, a tighter connection to their CRM, the SEO and GEO work, and eventually the AI workflow this diary is named after.
When I wrote about my three years at TROX, I ended with why I left - more projects, more contexts, more rooms where things are still being decided. This engagement is what I meant. New team, new domain, and a very different kind of surface: not an app behind a login, but the public face of a company, indexed and measured, where every shortcut is visible to anyone who takes a look.
A marketing site sounds like the easy end of the stack. I've come to think it's one of the more honest ones. There's nowhere to hide.

The work had one big idea in it, and a lot of quiet craft around it.
The big idea starts with a truth every marketing team knows: copy rots. Positioning shifts, a stat goes stale, a competitor renames the thing you compare against, and the fix is spread across dozens of pages in three languages, each edit too small for anyone to own. The traditional answer is a content backlog that nobody ever burns down.
The answer I built instead: an AI agent that reads the live site and proposes edits, plus a review surface inside the CMS where an editor sees each proposed change, adjusts it where the agent got it almost right, and decides. Publish or reject, change by change. Nothing goes live without a person saying yes.
That last sentence is the design, not a caveat. The instinct with agent tooling right now is to celebrate every human you remove from the loop. Sometimes that's fine - there's plenty of work where a machine can just do it and nothing bad happens if it's wrong. But there's a whole lot where the risk doesn't disappear with the human, it just goes unsupervised, and a company's public voice sits firmly in that category. The agent is tireless and well-read; the editor is accountable. So the whole tool is shaped around keeping the human decision cheap, clear, and reversible. An agent that proposes well is worth ten that publish fast. It was the biggest thing I shipped this spring, and the most fun. The first time I've built with AI as a material rather than just working alongside it.

The quiet craft is everything a visitor never consciously notices. Search engines now get told about every page in their own structured language, generated from the content itself so it can never drift out of date. The site got measurably faster - the kind of performance pass where no single change is impressive and the sum is the difference between a site that feels engineered and one that feels assembled. And the site was made deliberately legible to machine readers, because the near future of the web has as many of those as human ones, and pretending otherwise is not a strategy.
Then there's the everyday work, which was most of the commits: the audits, updates, and patches the first two months were made of, then content modules, forms, animations - the connective tissue between a website and the marketing team that lives in it daily. Building blocks editors can compose into whole pages without a developer in the room. None of it is portfolio-glamorous. But I wrote in the TROX diary that my favorite decisions are the ones whose value shows up six months later in a clean PR - this is the same idea, applied to content: every block an editor can use on their own is a future ticket that never gets filed.

At TROX, AI entered my workflow somewhere in end of year one / two, once the tools stopped being a party trick. At vivenu it crossed a line: it's not just in my toolkit anymore, it's in the product. And building with it taught me something I didn't expect.
The hard part was never the agent. Getting a model to read a page and draft a better version of it is nearly free now, and it'll be freer next year. The hard part is deciding what it's allowed to touch - which decisions can be handed over entirely, and which need to stay with a person whose name is on the outcome. That's not an engineering question, it's a product question, and no model answers it for you. I suspect drawing that line well will become a bigger part of this job than writing the code around it. Fine by me. The job keeps getting more interesting.
The engagement wrapped in July - two contracts, round about four months, and a website that now drafts its own maintenance. They have other priorities for the moment, which is exactly how a freelance engagement should pause: nothing broken, improved product, door open in both directions.