I was in Japan for three weeks.
Not checking provider docs. Not reading model deprecation notices. Not thinking about whether the AI infrastructure underneath my products had quietly changed state.
I was on holiday.
Got back. Opened the laptop. Started preparing a client report. Something was wrong with SmartRewrite inside Telleo. Traced it back. Anthropic had deprecated one of the models I was depending on.
No explosion. No warning. No message saying hey, the thing your product runs on just changed state.
Just a silent dependency failure sitting there, waiting for the worst moment to surface.
That's when I realised I was looking at a category of risk that didn't have a name yet.
I gave it one: model lifecycle risk.
The problem, properly defined
An AI product built on top of an external model is not static infrastructure. The model is not yours. The provider controls it, and providers rename, replace, deprecate and retire models. Sometimes with clear communication. Sometimes buried in a changelog nobody had time to read.
Your product can look completely fine while the model underneath it has already moved into a different state.
I call the worst version of that a zombie model. Not dead. Not safe. Still responding. Still sitting inside your app. Still making the demo look fine, right up until it doesn't.
The lifecycle: Active → At Risk → Deprecated → Zombie → Retired.
The gap: most products have no signal when the state changes.
Why this matters for small teams
Large teams can build monitoring processes around this. They can assign ownership, set up dashboards, make model lifecycle management someone's job.
Small teams usually can't.
Most indie developers, solo builders and small agencies are doing everything at once. The model changes anyway. The product drifts anyway. And you find out from a user.
That's not a fringe scenario. It's the real operating environment for most people building with AI right now.
What I built
Zombify watches the models your product depends on and alerts you the moment something changes state.
Drop in a model ID. Paste your .env. Get an email before it becomes a production failure.
It covers OpenAI, Anthropic, Google Gemini and OpenRouter. Live status board, model graveyard, email alerts. No ops department required.
I built it because I'd already been bitten, and when I looked for something that would have caught it, nothing existed at the right level.
This is the thread running through everything I'm building at Pixelhatch: making hidden AI systems visible before they make the decision for you.
Zombify is live at zombify.au.
— Reuben