AI Glossary · Letter M

Model Deprecation.

When an AI vendor announces it will shut off an older model on a set date, forcing every workflow built on that model to move to a new one.

Also known as Model retirement, model sunsetting

What it is

A working definition of Model Deprecation.

AI labs release new models every few months and retire older ones on a published schedule. Deprecation is the announcement and the countdown: the model keeps working for a while, then requests to it start failing or are silently redirected to a newer version.

The replacement is usually better on benchmarks, but better is not the same as identical. Tone shifts, formatting habits change, length drifts, and prompts that were carefully tuned to one model can produce noticeably different output on the next. Teams that pin a specific model version and keep a tested fallback get to choose when they migrate. Teams that point at a generic “latest” alias find out when the output changes.

Why ad agencies care

Why Model Deprecation matters in agency work.

Agency AI workflows tend to be tuned by feel, which makes them fragile when the model underneath changes.

Brand voice can drift overnight. A prompt that reliably produced on-brand copy can start sounding different after a model swap. If no one is checking, the client notices first.

Migrations need time in the schedule. Retesting a dozen production workflows against a new model is real work. Agencies that track deprecation dates can schedule it between campaigns instead of mid-launch.

A test set protects the agency. Keeping a small set of approved past outputs for each workflow lets the team compare old and new models side by side and sign off on the switch with evidence.

In practice

What model deprecation looks like inside a working ad agency.

An agency uses a custom AI workflow to write product descriptions for an ecommerce client, roughly 400 a month. The vendor announces the underlying model will be retired in 60 days. The agency’s ops lead pulls 25 previously approved descriptions, runs the same product inputs through the replacement model, and gives both sets to the copy lead blind. The new model’s descriptions run about 30 percent longer and lean on the same three adjectives. The team adds a length limit and a banned-words list to the prompt, reruns the test, gets copy lead sign-off, and switches two weeks before the cutoff.

Build AI workflows that actually run through The Creative Cadence Workshop.

The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.