"We launched the assistant in March. Usage was great for three weeks, then it fell off a cliff. Is the model wrong?" We hear a version of that question from product leads fairly often, and the model usually isn't the main problem.
What we tend to find is an interface that never told users what the assistant was for, gave them no cheap way to fix a bad answer, and had no exit to a person. That's an AI UX design problem. A feature that's right most of the time but fails unpredictably feels less dependable than a dumber tool that behaves the same way every time.
Below are the patterns we use when designing chatbots and copilots. They draw on published guidance from Apple, Google and Microsoft, and on what Australian privacy law will soon expect.
Probabilistic features break old design habits
Traditional software is deterministic: same input, same output. AI features aren't, so they'll sometimes be wrong and users need help deciding when to rely on them. Google's People + AI Research (PAIR) guidebook calls this calibrating trust: users shouldn't trust the system blindly, but should rely on it in some situations and double-check it in others.
Apple's Human Interface Guidelines for generative AI ask apps to say where they use AI and never to trick someone into thinking they're dealing with a human. Microsoft's Guidelines for Human-AI Interaction, published at CHI 2019, set out 18 guidelines that still apply neatly to today's copilots.
There's a regulatory angle too. The OAIC confirms that from 10 December 2026, privacy policies must cover computer programs that use personal information to make or inform decisions that significantly affect individuals, and that this isn't limited to AI. Our guide to automated decision-making disclosure rules covers what to disclose. Good interface design puts that disclosure where users actually make decisions, not just in a policy page.
AI UX design patterns, mapped to the user journey
Microsoft's HAX Toolkit groups its guidelines into four phases: initial interaction, during interaction, when the AI is wrong, and over time. We use the same phases to plan an AI feature.
| Moment | Pattern | What it looks like |
|---|---|---|
| Before first use | State capabilities and limits | "I can answer questions about your orders and returns. I can't change payment details." |
| Before first use | Offer starter prompts | Three or four example requests you've tested and know work well |
| During use | Label AI output | A clear "AI generated" marker and a note that content may contain errors |
| During use | Show the basis for an answer | Linked source documents or the records the answer used |
| When wrong | Make correction cheap | Edit, Undo, Retry and Adjust controls beside the output |
| When wrong | Hand off to a human | A visible "Talk to a person" path that carries the conversation across |
| Over time | Collect feedback | Thumbs up or down, optional comment, never required |
| Over time | Announce changes | Tell users when the model or its behaviour changes |
The empty text box is your first impression
An empty prompt box invites people to test the edges, and the first failure sets their opinion. Apple recommends setting clear expectations about what a feature can and can't do, flagging known limitations up front, and offering curated suggestions for open-ended inputs. Microsoft's first two guidelines say the same thing: make clear what the system can do, and how well it can do it.
In practice that's a two-line capability statement in the empty state, starter prompts built from real support tickets or tasks, and plain language about scope. If your support bot only knows the help centre, say so. A gotcha we see: starter prompts written by the product team that the model handles badly. Run each one through your evaluation set before it ships.
Explain more when more is at stake
PAIR's advice is to match explanation to the stakes. For a playlist suggestion, a label is enough. For a credit limit, a claims triage outcome or a medication reminder, users need to see why. PAIR also warns that numeric confidence scores are risky because they assume users understand probability, and suggests testing alternatives such as categorical labels ("high", "medium", "low") or showing the top few options. Source citations often do more for trust than any score.
Fixing an answer should beat starting again
Microsoft's guidelines include "support efficient correction" and "support efficient dismissal". Apple asks for controls like Edit, Undo, Retry or Adjust near generated content, and a clear signal when a correction has taken effect. For copilots that draft content, never lose the user's original: let them compare, revert, and accept changes section by section rather than all or nothing.
A way out to a real person
Human in the loop design isn't only a back-office review queue. It's also the user's exit when the AI can't help. PAIR's chapter on errors and graceful failure says that when the system fails, letting the user take over is often the easiest path forward, as long as they have the information they need to do it.
For a conversational UI, that means:
- Offer a person before the user has to ask twice, and always after two failed answers in a row.
- Pass the transcript to the agent so the user doesn't repeat themselves.
- Tell the user what happens next and roughly how long it'll take.
- Keep a non-AI path (a form, phone number or email) for anyone who'd rather not use the bot.
- For decisions with significant effects, have a person review the outcome before it reaches the user.
Plan for the errors nobody reports
PAIR treats errors as a chance to learn: build feedback options into error states and normal output alike. It also points out that users rarely report errors neither they nor the system noticed, so you need your own quality checks to find them. Apple adds that feedback should always be voluntary and placed where it doesn't interrupt the task. Microsoft's "scope services when in doubt" fits here: when the system isn't sure, ask a clarifying question instead of guessing.
Five mistakes we keep finding
- A chat box as the only interface, when a button or form would finish the task faster.
- No sign that content is AI generated, which costs trust when users find out.
- Fluent, confident answers with no sources, even on pricing or policy questions.
- No route to a person without abandoning the conversation.
- Feedback buttons wired to a table nobody reads.
Each is cheap to fix in design and expensive to fix once users have stopped trusting the feature.
What a JRC Software AI design engagement looks like
Our UX and product design team works next to our AI engineers, so interface decisions and model behaviour get designed together. In week one we map every decision the AI feature influences and how much each one matters to the user. That map tells us where you need explanation, review steps and a human handoff, and where a simple label will do.
Then we prototype the expectation, correction and handoff flows and test them with real users, including failure cases we trigger on purpose (out-of-scope questions, ambiguous requests, a retrieval source that's down). You end up with reusable patterns and components, ideally inside a small design system, plus an accessibility check, since a copilot panel that traps keyboard focus fails users before the model is even involved.
What we'll push back on: adding a chat interface where a well-designed form would do the job.
If you're planning or rescuing an AI feature, talk to a product designer about your use case.
Frequently asked questions
What is AI UX design?
AI UX design is designing interfaces for features powered by machine learning or generative models. It deals with problems ordinary software rarely has: outputs that are sometimes wrong, behaviour that changes over time, and users who need help deciding when to trust a result. Core patterns include stating capabilities, labelling AI output, cheap correction and a clear route to a person.
How do you build trust in a chatbot or copilot?
Set accurate expectations before first use, label AI-generated content, and show sources where the stakes are high. Make results easy to edit, undo or retry, and offer a human handoff when the AI can't help. Google's PAIR guidebook calls the goal calibrated trust: users rely on the AI in some situations and double-check it in others.
What is human in the loop design?
Human in the loop design puts a person at defined points in an AI-assisted process, either to review outputs before they take effect or to take over when the system can't help. In the interface, that means clear review steps for high-stakes decisions and a visible handoff that carries context across, so neither users nor staff start again.
Do Australian businesses need to disclose AI use to users?
Some will. From 10 December 2026, privacy policies must cover certain automated decisions that use personal information and significantly affect individuals, whether or not AI is involved. Our separate article on automated decision-making covers the detail. Clear interface labels are a sensible complement to the policy. This is general information, not legal advice, so check the OAIC guidance and get advice for your situation.
UX & Product Design
Talk to a product designer
Tell us where you're stuck and we'll come back with practical next steps.
Get in touchSources
- OAIC: New resources on transparency for use of AI and automated decision-making
- Google PAIR: People + AI Guidebook, Explainability + Trust
- Google PAIR: People + AI Guidebook, Errors + Graceful Failure
- Apple: Human Interface Guidelines, Generative AI
- Microsoft Research: Guidelines for Human-AI Interaction (CHI 2019)
- Microsoft: HAX Toolkit, Guidelines for Human-AI Interaction
- Microsoft: HAX Toolkit, Design Library