Three days before launch, Mintel's new AI product didn't work.
The team had built a tool to answer questions over their research. It was 2023. The answers weren't good enough, and the launch date was fixed. OpenAI released GPT-4 Turbo. Three days later the product went out, and it worked.
Jason Thomson, who leads innovation at Mintel, told that story at aiblLIVE in London on 20 October. It was one of several on a slide he might as well have titled "not yet".
Two and a half years of no
The best one was a database agent. Mintel runs GNPD, a huge database of product launches from around the world. The idea was an AI agent that could dig through it and answer real questions, the kind an analyst might spend a day on.
It failed, and failed again. For two and a half years, by Jason's account, it didn't work. When a new model came along, it "just worked", and it launched in July 2026.
His line, which every business owner should hear: "Often when they don't work, it's just that they don't work yet."
Why it keeps changing
There's a reason for this, beyond the hype.
Epoch AI, a research group that tracks how AI is developing, published a study in September 2026. It found that the cost of getting a given level of AI performance has fallen about 47% every quarter for three years. That's around 13 times cheaper every year. Jason quoted an even bigger number on stage. Epoch's own figure is impressive enough.
When something gets that much cheaper and better that quickly, last year's "can't be done" is often this year's "of course".
What changes when it works
Jason didn't run his demo live. Instead he showed a recording of the finished agent working through a question about high-protein fizzy products. By his account it ran 16 searches and 13 analyses over five years of product launches, in three minutes, and wrote four pages.
And he said the thing that stuck with me most. "Maybe 1 in 4 of our users could have done this analysis by hand, and those that could would only have bothered if it was a really important question."
And: "When something gets that much easier, it stops being the same thing." As he put it, "You get new questions asked."
That's the real prize. People ask questions they'd never have bothered with before, because the cost of asking has dropped to almost nothing.
A list worth keeping
I like this idea because it fits something I've believed for a long time. Done beats perfect. You try, you learn, you go again. The bit people forget is the "go again".
Plenty of us try an AI tool once, get a so-so result and decide it's not for them. Some of that is fair. A lot of it is timing.
Here's a simple habit that helps. Keep a "not yet" list. Every time you try something with AI that doesn't quite work, a proposal, an analysis, a tidy-up of your customer list, write it down with the date. Every three months, try the top three again.
What matters is paying attention when "not yet" quietly becomes "now". That way you get there while everyone else is still telling each other it doesn't work.
Jason admitted the pace is tiring. "It's exhausting," he said in the Q&A, because what's possible "changes every week." He's right. A list makes it manageable. You don't have to chase every change, just check back on the things you already wanted. What's on your "not yet" list?
