You can't judge what you don't know.

An AI-generated image comes up on the screen and it looks finished. Then somebody leans in, zooms in, and says the frosting is wrong and the logo sits a pixel off.

That somebody is Dominik Heinrich's boss at Coca-Cola. Heinrich is the company's Global Head of Design Intelligence. He also teaches AI design at Pratt and co-founded Creative AI Academy, and he spoke at AUTONOMOUS earlier this month.

Ten months before that he'd been on the same stage describing a design system built with Adobe, then still a proof of concept. It's now running in production across more than 200 markets. So when he describes what AI changed inside the work, he watched it happen at scale.

What changed is the ratio. Creative work used to run somewhere around 70 to 80 percent production and 20 to 30 percent crafting. Production covers the machine of making something, from the idea through the storytelling to the shoot. Crafting is everything after: the retouching, the color, the small decisions that separate finished from nearly finished. Heinrich puts the split now at 30 to 40 percent production and about 50 percent crafting.

Those numbers don't total a hundred, and he said that was on purpose. The gap is the time you actually save.

Heinrich described watching firms hire back talent they let go six months to a year ago, after finding that AI is expensive and still unreliable without someone doing the real work. He was pointed about the phrase “human in the loop,” because it has come to suggest a reviewer who approves output.

“The human needs to go in and needs to actually work and do the details.”

His advice to students and to the professionals in his courses is to learn the craft first, and his reason is that you can't judge what you don't know. Someone who never learned what good looks like will approve work that's almost right. Almost right, produced at volume, is how a brand thins out a thousand small decisions at a time.

The ceiling on what you can put out is how many people can tell good from nearly good. Most companies have never counted that number.

None of this is only about design. Anywhere the production step collapsed, the same thing is happening. A marketing team can turn out a hundred versions of a campaign and still needs the person who knows which one will move a customer. A commercial team can model twenty pricing scenarios and still needs someone who knows which assumption holds up with real accounts.

Heinrich put most companies in the first wave of this, where the conversation is still about efficiency. The second wave is where quality and craft start paying. His read is that we're touching the second without being in it. The productivity is real, and the P&L hasn't moved.

Everything I know how to do, I learned by volume. A lot of reps, feedback from the market, course correct, go again. I still work that way.

Judgment has always been built by doing the production work. You learned what good color looked like by getting it wrong a few hundred times, and which assumption the answer really depended on by building the model that broke in front of a client. The apprenticeship was hidden inside the grunt work, and nobody had to design it because it came free with the job.

AI just removed the grunt work. The same step that slowed down the people starting out was the step that turned them into the ones you rely on.

The production work was the apprenticeship.

Which means work process redesign needs a second question. We've been asking what can be automated and what it saves. Nobody asks what the step was teaching, or who stops learning when it goes.

Work process redesign is going to be everywhere through the rest of this year and into 2027. My view is that some share of the work should stay the old way on purpose, because that's how people build the judgment the automated version depends on. Find the steps that taught the most, keep those human, and budget for the inefficiency the way you'd budget for research or training and development.

Concentrate it. A few projects run start to finish the old way, with specific people on them, will build more than a slice of every project kept manual. Spread it thin and you slow the whole operation down without teaching anybody, because the learning came from carrying one thing all the way through.

There's a book by Henry Cloud called Integrity that shapes how I think about this. He writes about embracing the negative, going at the hard thing instead of around it, and about being oriented toward truth. Both of those get built in the arena. You try something, somebody tells you how it could be better, and you go again until you can see it yourself.

Somewhere along the way, mastery stopped being the point. Remote work took away the person leaning over your shoulder to say that's not right yet. Moving fast took the second pass with it. Now AI can make the attempt for you, so there's less of your own work sitting out there to be judged, and fewer moments when somebody tells you the truth about it.

I want to see us go back to building mastery with conviction, and being clear what it costs. Give people work that's theirs to carry all the way through, tell them clearly when it isn't there yet, enable them as needed, and let them bring back better versions on repeat.

It runs both directions. Leaders have to be willing to say the hard thing, and the person who wants to develop mastery has to go ask for it, because nobody develops on feedback they managed to avoid. Take your last piece of work to someone whose judgment you respect and ask how it could be better. Then sit there while they tell you, without helping them soften it.