Your AI is naked, and that's why everything it makes is mush

Sometime this year, your company got AI. There was an announcement, maybe a lunch-and-learn. Everyone got a chat box. And six months later the reports are still long, the decks are still ugly, and the smartest software in human history is mostly summarizing meetings nobody wanted to attend.

Meanwhile some tiny app your coworker found does one thing so well it feels like a magic trick.

Same technology. Wildly different results. The difference has a name, and the name is the fun part.

The horse race

Every few weeks there is a new headline about the smartest AI ever built. New model, new number, new record on a test you have never heard of. The companies behind them are in an actual arms race, and the coverage treats each release like a title fight.

Here is the thing the headlines skip: the models are all good now. Not identical, but close enough that for most everyday work, swapping one for another changes less than you would think. They are racehorses. Magnificent and expensive, and at a full gallop you cannot tell them apart.

So why does most AI at work still produce mush?

The part nobody tweets about

Engineers have a word for everything strapped around the model: the harness. The word is borrowed from horses, and the borrowing is the whole story. A harness is the gear strapped around the animal, the part that turns raw horsepower into a load actually pulled in a straight line.

In AI, the harness is every decision about how the model is used. What it gets told about your situation. What it is allowed to touch. What it is flatly forbidden to do. What happens to its output before you see it. None of it is intelligence. All of it is the difference between work you can use and work you have to redo.

A chat box is a naked racehorse. You climb on bareback, shout a destination, and hold on. Sometimes that goes great. You have seen the other times.

Same horse, different day

Ask a bare chat box to "make my quarterly update look good" and you get something that looks like everyone else's quarterly update, because the model is guessing at everything: your numbers, your audience, your taste, your definition of good.

Put the same model in a proper harness and the guessing stops. It knows the numbers because it was handed them. It cannot bury the important figure, because the rules it works under set numbers large. It cannot ship anything, because the harness routes every change past you for a yes. The model did not get smarter between those two paragraphs. It got dressed.

A confession

Here is a secret about foldout, the product this blog belongs to. People sometimes assume the pages it makes look the way they do because we have some exotic model in the basement. We do not. The model inside foldout is a perfectly ordinary workhorse.

What we spent our time on is tack. A design system the model cannot break, no matter how hard it tries to add a purple gradient. Slide geometry it cannot overflow. A rule that nothing it writes lands on your page until you have read the change and said yes. Take our model and put it in a bare chat box and you would get the same mush as anywhere else. The pages are good because the harness is strict.

We are not special in figuring this out. Every AI product you love is a modest horse in excellent gear. Every AI product that has disappointed you was probably a spectacular horse wandering a field.

What to ask instead

The next time someone at work evaluates an AI tool, notice the first question asked. It is usually some version of "which model is it?" That is a horse question, and it is the wrong one. Better questions:

• What does it know about my stuff? A tool that sees your actual document beats a smarter one guessing from a prompt.
• What is it not allowed to do? The forbidden list is where quality lives. No rules means mush with confidence.
• What happens on a lazy Tuesday? Judge the output when you give it three words, because that is how everyone actually uses these things by week two.

None of those questions require knowing anything technical. They are the same questions you would ask about a new hire.

The race will keep running

The model headlines will not stop, and the horses really will keep getting faster. Enjoy the race. Just remember that nobody ever got their groceries delivered by a racehorse.

The work gets done by the harness.