When a product adds an AI feature, the question users ask changes. It is no longer only "Does this work?" but "Is this right, how did it decide, and can I undo it?" Trust in AI features is built on three things: being open about what the system is, giving people control over it, and making what it does reversible.
AI output is based on probability, not certainty. It can be right most of the time and still be wrong in ways that are hard to predict. Google's People + AI Guidebook puts it directly: explaining predictions, recommendations and other AI output to users is critical for building trust. For financial and other high-stakes products, getting this right matters even more. This article covers the three principles and how to apply them.
Why Trust Works Differently With AI
Traditional features behave the same way every time, so users learn what to expect. AI features don't. The same request can produce different results, and the reasoning is usually invisible. That creates two opposite risks:
- Over-trust. Users accept every suggestion because it looks confident, including the wrong ones.
- Under-trust. Users ignore a useful feature because they can't tell when it can be relied on.
The goal is calibrated trust: people rely on the system when it deserves it and double-check when it doesn't. Design has to give them the information to tell the difference.
1. Transparency: Be Clear About What It Is
Say when users are dealing with AI
People should not have to guess whether a message, summary or recommendation came from a person or a model. Regulation is moving the same way: Article 50 of the EU AI Act says providers must ensure that AI systems intended to interact directly with people are designed so those people are informed they are interacting with an AI system, unless this is obvious from the context. This is not legal advice, but clear labeling is good practice well beyond any single regulation.
Explain, at the right level
Not every output needs a full explanation. Google's guidebook asks when explanations are needed and how much to show. A useful rule is to explain more as the stakes rise.
- Show the main reason behind a suggestion in one short sentence, such as "Based on your last three payments."
- Link to the sources or data the result relied on, where possible.
- Offer more detail on request instead of showing it by default.
Be honest about limits
- Tell users what the feature is good at and where it can be wrong.
- Show uncertainty when it is meaningful, rather than presenting every answer with equal confidence.
- Say plainly when the system cannot answer, instead of producing something plausible.
Transparency is not about explaining everything. It is about giving users enough to decide how much to rely on the result.
2. Control: Let People Steer
The same guidebook has a chapter on feedback and control, asking what the right level of control and customization is for users. Control turns an AI feature from something that happens to people into something they use.
- Edit before it counts. Let users change AI-generated content, categories or settings before they are applied.
- Offer simple feedback. A quick way to mark a result as right or wrong improves the system and gives users a sense of influence.
- Allow adjustment. Let users decide how proactive the feature is, what it may use and when it should ask first.
- Keep manual paths available. Nobody should be forced through the AI route to complete an important task.
3. Reversibility: Make Mistakes Cheap
No AI feature is always right, so trust depends on how easy it is to recover. Following NN/g's principle of user control and freedom, users need a clearly marked way out of an unwanted action.
- Provide undo for anything the AI changes.
- Show a preview before actions that are hard to reverse.
- Keep a history of what the system suggested or did, so users can review it later.
- Mark actions that cannot be undone, and always ask before they happen.
What This Looks Like in a Financial Product
Financial products show why these principles matter. Take a feature that automatically categorizes spending and suggests a monthly budget:
| Principle | Weak design | Stronger design |
|---|---|---|
| Transparency | A budget appears with no explanation | "Suggested from your last six months of spending," with a link to the data |
| Control | Categories can't be changed | Tap any category to correct it, and the correction is remembered |
| Reversibility | Changes apply instantly and silently | Preview before applying, with a visible undo afterwards |
The stronger version asks for a little more design work, but it replaces a black box with something a careful user can understand, correct and rely on.
Common Mistakes
- Hiding the AI. Presenting generated content as if a person wrote it.
- Projecting false confidence. Showing every answer in the same assured tone.
- Making correction hard. Offering no way to fix a wrong result.
- Explaining too much. Long justifications nobody reads.
- Making AI the only route. Removing the manual alternative for important tasks.
Trust Checklist for AI Features
- Is it clear to users when they are dealing with AI?
- Does each important result include a short reason?
- Do users know what the feature can and cannot do?
- Can they edit or correct results before they take effect?
- Is there a simple way to give feedback on a result?
- Can every AI action be undone, or clearly marked as final?
- Is there always a manual way to complete important tasks?
Conclusion
Users don't need AI features to be perfect. They need to understand them, steer them and recover from their mistakes. To summarize:
- Be transparent about what the system is, why it suggests what it does and where it can be wrong.
- Give people real control over how it behaves and what it changes.
- Make every action reversible, or clearly flag the ones that aren't.
If you are adding AI to a product where trust matters, our UX Audit & Optimization service reviews the moments that decide whether users rely on your features, and redesigns the ones that put them off.
