The curve before the gains
Two-thirds of marketers say AI has given them more time for creative work. The only question is what they'll do with it. I spent mine learning AI. That reads like a joke, but it paid off. I'm genuinely faster now, just not in the way that statistic suggests. Because there's a curve in front of it.
If you want to build something genuinely useful with AI, you're considerably slower at the start. You're figuring out how something works. Building it. Testing it. Throwing it away. Building it again. Meanwhile, the person who approved the efficiency gains has started asking about them. The acceleration shows up later. Take a system that automatically turns one campaign into dozens of format adaptations. Once it works, the savings can be enormous. Getting the output consistently good enough to actually use can take weeks if not months.
Economists have a name for this: the productivity J-curve. New technologies don't create their full value simply by being dropped into existing workflows. First you have to invest in the processes, skills and systems around them. Productivity can go down before it goes up. Which is pretty much where we are with AI. Almost everyone is using it: 88% of organizations, in at least one function. Only 7% have scaled it across the company. And Gartner reports that at least half of all GenAI projects were abandoned after the proof of concept.
But a J-curve is not a promise. Some of those projects weren't early. They just weren't worth building. Those are enterprise numbers, of course. The same mistake is just easier to see at desk scale. During my creative technologist training we built a lot of workflows, and afterwards we'd regularly end up asking the same question: would the result have been any worse if we had just used the chat? Often, the answer was no.
In 1966, Abraham Maslow put it neatly: "I suppose it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail." AI may be the most powerful hammer we've ever had. And right now, we're discovering an astonishing number of nails. An agent for this. An automated workflow for a task somebody does once a month. The thing is, you pay a version of that J-curve every time: figuring it out, building it, connecting it, testing it, maintaining it and getting people to actually use it.
So the interesting question isn't: Can we build an AI system for this? Of course we probably can. It's: Does this problem occur often enough, cost enough and matter enough to justify the curve?
