What AI-driven experimentation still can't decide

There’s a class of experiment platform that, by 2026, can do almost everything I used to manually orchestrate. Generate variant copy. Set up the test. Allocate traffic with a multi-armed bandit. Watch the metrics. Roll back the loser. Generate a brief explaining why the winner won.

The pitch is full-stack experimentation. The pitch is mostly true.

What it can’t do is decide whether you should have run the test in the first place.

Here’s the kind of thing that happens. A team spends six weeks A/B testing the color of a checkout button. They get statistical significance. Variant B converts 4% better. Then someone asks the question that should have been asked in week one: what would we actually do with another 4% conversion at this stage? The answer, after some honest math, is: not enough to matter. They were optimizing the wrong layer of the funnel. The whole experiment, from a business perspective, was a more rigorous version of the wrong question.

AI-driven experimentation widens this hole. The tools get faster. The signal gets cleaner. The number of experiments you can run goes up. None of which addresses the only really hard part of experimentation, which is deciding what’s worth measuring at all.

The new wave of platforms can propose hypotheses. LLMs scan your product analytics and surface “users who do X are 2.4x more likely to retain. Should we try a variant that nudges X?” That’s useful. It’s also potentially dangerous. A system that proposes hypotheses based on patterns it sees in your data is a system that can fluently propose hypotheses you would never want to test on principle. Manipulating users into outcomes you optimized for, but didn’t decide on, is the easiest thing to do in this stack now.

Three things I think still belong squarely with humans:

The success metric. The model can suggest what to measure. The model cannot decide what should be measured. That choice carries every value tradeoff downstream. If you outsource it, you outsource the company’s actual values.

The off-limits. Some experiments shouldn’t run. Dark patterns. Dependency-inducing nudges. Tests that wouldn’t pass a press release. Models don’t have an instinct for this. The humans in the room do, if they’re given permission to say no.

The interpretation. “Variant B won” is a fact. “Variant B won because we shifted users from X to Y, and Y aligns with the part of the product we want to grow” is judgement. The first is automatable. The second is the part the company is actually paying for.

A useful test: imagine a team where every experiment is generated, run, and analyzed by AI, with a human only signing off. After a year, what does the product look like? In practice, the answer is usually a slightly more profitable version of whatever was already there, missing the thing that would have made it actually better.

That’s the cost of automating away the judgement layer. You get more decisions, faster, at the cost of better decisions, slower.

The right shape is the inverse. Let the machine run the experiments. Let the human decide which ones to run, and what the result means. The interesting work happens at the question, upstream of the funnel.