In 2023 I wrote in this newsletter that OpenAI’s new GPT framework represented “a strategic advantage for businesses”, on the grounds that generative AI “achieves its best performance when focused on a particular domain”. On 11 September, OpenAI announced that it will retire custom GPTs across every ChatGPT plan. They are scheduled to stop working on 11 December.

I have reread that piece with some care. The premise has held up well: domain focus is still where the value sits. What I backed was the container, and the container is going.

Google is doing the same. Gems, its version of the custom assistant, become “skills” for personal accounts from November, for Workspace business customers from March 2027, and for education from June 2027. OpenAI, meanwhile, has come full circle. When it shut down ChatGPT plugins in April 2024, it said GPTs had “full feature parity” with them. Two years later, GPTs are being replaced by plugins.

Focus On: Who Chooses Now

Strip away the branding and one thing changes. With a GPT or a Gem, a person chose the expertise. You opened the contract-review GPT because you knew you needed a contract reviewed, and its instructions governed everything that followed.

With a skill, the model chooses. Google’s own guidance says Gemini checks only a skill’s name and description to decide whether it is relevant, and opens the full instructions only when it decides they are. OpenAI says ChatGPT may use an installed skill automatically when its description matches the request. Several skills can apply in one conversation, both vendors let them run inside scheduled tasks, and both now accept skills written in the same SKILL.md format. That is why I’d call this a maturity step rather than a rebrand.

Take the conference slides away and this is what orchestration means in practice: delegated selection. A machine reads a short description and decides whether your expertise applies.

The First Mass Test

The GPT migration is the first time millions of users have handed that decision over at once, and the early reports are instructive. OpenAI’s migration turns each GPT’s instructions into a skill and copies its files across. Then it warns, with admirable candour, that “a migrated plugin may respond differently”, and gives users a checklist whose first item is whether the replacement selects the right skill.

Builders on OpenAI’s forum are reporting what that looks like. One user found instructions and knowledge files ignored until they typed an explicit instruction to follow the skill. Another watched responses drift back to ordinary ChatGPT after a few turns. Someone relying on GPTs for clinical documentation reported changed headings, reorganised sections and omitted details. A builder who ran several GPTs as a coordinated team asked how to preserve the hand-offs, and OpenAI’s support team wouldn’t promise the workflow would survive. These are self-selected complaints, and some will be fixed before December. They still point at the same mechanism.

The research has a name for it: confusability. In one controlled study, adding a single near-duplicate for each skill cut selection accuracy by between 7 and 30 per cent, and the main source of error was how alike the skills looked. Grouping skills into families before choosing recovered most of the loss, by up to 40 points on the smaller model tested. The libraries were synthetic and the models have since been superseded, so trust the direction more than the decimals. That direction is uncomfortable for marketing.

Why Marketing Should Worry

In the GPT era, sprawl was harmless, and we celebrated it as adoption. Moderna’s employees built 750 GPTs in two months, and the figure became a case study. A GPT nobody opened cost nothing.

In the skills era, a near-duplicate costs something. It sits in the same pool as every other skill and makes all of them harder to tell apart. The library has become a commons.

Marketing functions produce near-duplicates by design. A regional brand voice next to the global one. A local campaign template next to the master. A product-line version of the messaging framework. My expectation is that any federated marketing organisation will grow a library of plausible, overlapping skills within a quarter, and the failure will be quiet: the wrong skill runs and the output still looks fine.

CMOs know this problem under another name. It is brand architecture. A house of brands whose sub-brands customers can’t tell apart cannibalises itself, and a skills library whose descriptions a machine can’t tell apart does the same. The disciplines carry straight over: one job per skill, a description that names what makes it different, look-alikes merged, families grouped. The description field is copy written for a machine reader, and it deserves the care we give any other copy.

In July I argued that the defensible asset in enterprise AI is the learning loop. The skills library is that loop in its most concrete form, and it is the first version of it that can be degraded by its own growth.

What the Sceptics Get Right

Three objections deserve weight. A skill’s body is still prompt text, and a simple GPT really is just being renamed. Some builders argue, convincingly, that a deliberately chosen environment is the right control for regulated work, and that ChatGPT’s Projects are the truer successor. And the vendors’ motives are visible: Google’s skills currently sit behind its paid tiers where Gems were free, and OpenAI is consolidating around a format that also sells its agentic products.

All true. None of it changes who now makes the decision.

So Should Your Brand Skills Trigger Themselves?

Yes. Lock your brand and compliance skills to manual selection and you have rebuilt the GPT inside the new architecture, giving up the reason for moving. But the right to trigger should be earned, the way any delegated authority is. A skill earns it by being picked correctly on a fixed battery of test prompts, and loses it the moment a look-alike enters the library. The ten-prompt test battery I set out in August does the job with one change: test which skill fires as well as what it produces. Both vendors already let you switch automatic selection off, skill by skill. Keep the switch off until the skill has earned it.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.