Lab Notes

AI Article

The six places an AI feature can fail its users, and how to catch each one before launch.

A user trusts a wrong answer. Someone can’t stop an action. The AI does something nobody approved. Moments like these fall into six areas, and each one gives your team something specific to check before launch.

Where the Risk Comes From

Traditional software follows rules the team wrote, so its expected behavior is easier to describe and test. AI features work differently, whether they are a chatbot, an assistant, or an agent that takes actions. The same request can produce different answers, and people may use it in ways your testing didn’t cover. You can’t list every response in advance, so testing needs real-world scenarios, edge cases, and unexpected inputs.

So there are two questions. One is model quality, which your engineers or your vendor can measure. The other is product behavior: what the feature is allowed to do, what users believe it can do, and what happens when it gets something wrong. This article is about product behavior.

The Six Places an AI Feature Can Fail People

Boundaries

What should this AI never do? An assistant built to summarize documents might start giving legal advice, or an agent with account access might take an action nobody approved. Write down what the feature must never do, and build those limits into the product through permissions and approval steps.

Expectations

What will people wrongly assume? A polished answer or a score can look more reliable than it is, for example an AI candidate score that feeds a hiring decision even though the product presents it as advice. Give people a way to check the result before they act on it.

Trust & Clarity

Can people understand and verify what the AI produces? Show what the answer is based on: its sources, a short reason, or a clear note when the system can’t verify the result.

Control & Autonomy

Can people steer, pause, or override it? Ways to stop, undo, or change an action let people correct a mistake before it spreads.

Safety & Misuse

What could harm real people or data? This covers inputs designed to get around the AI’s instructions or reveal information it shouldn’t, sensitive data showing up in responses, and outputs that could hurt vulnerable users. Testing these cases before launch, including deliberate misuse, reduces the risk.

Failure & Recovery

What happens when the AI is wrong? AI features can give wrong or incomplete results, even when they did well in testing. Showing when it is unsure, and giving people a clear way to fix the result, keeps the experience working.

Why These Six Matter

These six areas shape how much people trust an AI feature, how much support it needs, and, in sensitive areas such as hiring, health, or money, whether it meets legal requirements.

Compliance Is Part of the Picture

The EU AI Act sets transparency rules for certain AI systems, including telling people when they are interacting directly with an AI in the cases it covers. Check which rules apply to your product, and from when.

The European Commission sets out the full timeline on its AI Act page. A UX risk review and a compliance review both start from a clear picture of what your AI does and how people experience it, but they check different things.

Catch the Risks While Behavior Can Still Change

Run this evaluation before launch, while you still have room to shape how the AI behaves. At that stage, changes can still go into the requirements, design, and build. After launch, the same changes usually take more work.

A structured audit takes your product requirements, specifications, and flows, maps each AI behavior against the six categories above, and returns a ranked list of what needs attention before users ever see the feature.

Four Questions to Ask Your Team This Week

You can start without a full audit. Ask your product team these four questions:

  • What should this feature never do, and where is that written down?
  • How does a user check whether the AI's output is correct?
  • Can a user stop, undo, or override an action the AI took?
  • When the AI is wrong, what does the user see, and who hears about it first?

The answers show where your specifications and flows have gaps.

Frequently Asked Questions

What is an AI UX risk audit?
An AI UX risk audit is a structured review of how an AI feature behaves toward the people using it. It evaluates the product against defined risk categories, including boundaries, expectations, trust, control, safety, and failure recovery, and it produces a ranked report of gaps with recommendations. It looks at product behavior and sits next to technical model evaluation, without replacing it.

When is the right time to run one?
The best time is when specifications and flows exist but the behavior can still change, which usually means before or during the build. The audit also works on live products, where it helps explain adoption problems and support load and gives you a prioritized fix list.

Our model vendor handles safety, so do we still need this?
Vendor safety work covers the model. It does not cover your product: what your feature is allowed to do, what your users will assume, whether they can verify and undo, and how your interface handles a wrong answer. Those decisions belong to the team that designs and builds the product, and that is where users feel the risk.

We Audit AI UX Risk at Creative Glue Lab

We built the six-category framework in this article through our client work on AI products, and we use it to audit AI features before they launch. Our services cover both AI Product Design and AI Development, with strategy, UX, and engineering under one roof, run by senior practitioners on every project.

If you want a clear read on the risks in an AI feature you are building, we'd love to have a chat. Happy to talk whenever you're ready. Book a clarity call

We bring the glue.
You bring the spark.

Ileana Marcut, co-founder of Creative Glue Lab
I’m Ileana, co-founder at Creative Glue Lab. Got an idea? Let’s chat.