Why automations break, and how to build ones that last

OpenAI's deprecations page records that it shut down a batch of its models on October 23, 2026, among them gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1 and o4-mini, plus some fine-tuned versions. For each one it names a recommended replacement (OpenAI deprecations page). The page is the authority here, so check it for the current list before you rely on this one.

I am not telling you this to pick on one company. Every AI vendor retires models, and the notice is usually short. If an automation of yours quietly calls a model that is switched off, it does not slow down. It stops, and unless you built a way to hear about it, nobody finds out until the work piles up.

A common bottleneck: automations that quietly stop working

Plenty of owners have built an automation, felt great for a month and then found out it stopped doing anything weeks ago. Nobody noticed, so the work went back to being manual, and the owner concluded that automation does not work.

It usually breaks for one of four reasons:

The process changed and the automation did not.
A tool changed its pricing, its interface or its connection.
The one person who set it up left, or forgot how it worked.
A failure was invisible. It stopped, and nothing told anyone.

How to spot a fragile automation

Ask three questions of anything you have already built:

If it stopped today, how would I find out? If the answer is "a customer complains," it is fragile.
Who owns it? A name, not "the team."
Could I move it to a different tool in a week? If not, you are renting your process.

How it gets built to last

Fix the process before you automate it. Automating a messy process makes the mess faster.
Write the rule down in plain words. If a new person cannot read it and understand it, it is not ready.
Make failures loud. When it breaks, someone gets a message the same day.
Give it one named owner. Review it on a schedule, even a short one.
Keep the pieces portable. Your data and your rules should be exportable. Treat any single AI model or vendor as swappable.

That last step is the one a model retirement points at. Do not pick your setup around one vendor's model. Before you sign anything, ask how you get your data out and what it costs to switch. If the answers are vague, that is the answer.

Why I do it this way

In my own products, including an AI business knowledge tool, what goes wrong in daily use is rarely the clever part. It is the silent failure and the missing owner. Every project gets a fixed quote before work starts, and I reply to every message myself.

What "fixed for the long term" means

The automation survives a change of tool, a change of model and a change of staff. When something breaks, you see it that day and you know who fixes it. That is a smaller promise than "set it and forget it," and a truer one.

What I would do on Monday

Pick one automation you already have. Write down how you would know if it stopped, who owns it and where its data lives. If any of the three is blank, fix that before you build anything new.

Show me yours

Bring the bottleneck, or the automation that stopped working. On the free call you describe your week, I look for what repeats and what is fragile, and you leave with one place to start. No obligation.

Show me yours. Text (512) 337-9530, email josh@joshdlange.com, or book the free AI assessment.

Source: OpenAI deprecations page. Shutdown date as listed there; checked October 14, 2026.