AI products are non-deterministic. Design like it

A normal app keeps a promise: the same input, the same answer, every time. An AI product cannot make that promise. Ask the same model twice and you get two different drafts, and one of them will be delivered with full confidence and be wrong. Most AI apps are still built as if that variance does not exist, which is exactly why AI products fail at the interface, not the model. The model usually works. The surface around it does not.

The reliable part of an AI product is not the model. It is the interface that decides what the user sees when the model hesitates, when it is wrong, and when it needs them. Non-determinism is not an edge case to hide. It is the subject matter.

On this page

• Why do AI products need a different design?
• What does designing for uncertainty look like?
• How do users steer an AI product?
• What is the right failure state for an AI product?

Why do AI products need a different design?

Because a normal app has one fixed outcome per action, and an AI product has a range of outcomes: an excellent one, a mediocre one, and a wrong one delivered with total certainty. The user never knows which they get until they have already acted on it. That changes the design job from "guide a user down a path" to "keep a user oriented while the ground keeps moving."

ThoughtWorks' design team frames the work as designing AI as non-deterministic systems — deciding the context, the guardrails, and the escalation rules before any screen exists, because you cannot wireframe output that has not happened yet. The screen is still the product. But design effort moves to the edges: the boundaries, the escape hatches, the handoffs to a human.

What does designing for uncertainty look like?

Three shifts matter immediately.

First, shrink the space of possible answers before the user sees anything. The model is only as good as the context you feed it, and Anthropic's notes on context engineering make the point bluntly: curate the smallest set of high-signal context, and the model strays less. This is not an infrastructure detail. It is the difference between an app that drifts and an app that lands in the same place reliably.

Second, let the UI say what the model actually knows. Show "draft" instead of "done." Show alternatives so the user does not assume there was only one landing spot. Show where the answer came from when it matters.

Third, stop delivering AI output as final. If the interface cannot tell a draft from a verdict, it forces the user to do the confidence check, and users are terrible at it.

How do users steer an AI product?

If the model produces a draft, the user needs room to fix it. The pattern that moves results most in AI products is treating generated output as a starting point that the user edits, not as a result they receive. That prints into three affordances: edit the thing directly, regenerate a variant when the direction is wrong, and tell the model what to change rather than starting over.

The deeper pattern is drawing the right boundaries. The user should be correcting the model's decisions, not just its copy. Google's People + AI Research guidebook on calibrating trust is the best public reference here: it splits the work into planning where users should just accept output, where they should verify it, and where the system should always ask first. Whatever stays on that map matters. Whatever would stay correct on that map needs the design.

What is the right failure state for an AI product?

A path forward, never a dead end. The wrong model compresses into a "sorry" toast and an empty screen. The right one tells the user what the model could and could not do, how much it is sure about, and what to try next — including doing it manually. The user experiencing AI error is the user the product is about to lose, so the failure state deserves the full design budget, not leftovers.

Non-determinism is not a hunt you fix once. It is the medium. Design shows a screenshot — design that survives an LLM being wrong and confident in the same response wins. That is where the taste opinions live, and that is where users decide whether to stay.