01
Show the plan before any costly or shared action
People can't undo what they never saw. A plan turns a surprise into a decision.
Do
Plan: TLV → LIS · Seat 14C · $412
Approve and pay
Don't
Done! Booked and paid $468.
Irreplaceable › Guides › AI UX checklist
Checklist · AI product design
The AI UX checklist
18 rules for designing AI features, grouped by the moment they matter. Each one has the reason behind it and a do and a don't, so you can check a screen in a few minutes.
By Sarit Elisha, founder of Irreplaceable · Updated October 10, 2026
Open the interactive Library
Search the checklist next to frameworks, AI failures and senior takes. Free.
Before the AI acts
For: Agents that book, send, pay or change things.
02
Ask one clear question when the request is ambiguous
Guessing on ambiguity causes costly mistakes. Asking five questions every time causes friction.
Do
Which Daniel: account …0912 or …4410?
Don't
Before I start: 1) … 2) … 3) … 4) … 5) …
03
Let people edit the plan before it runs
Steering before acting beats undoing after.
Do
3 steps: find, compare, book
Edit plan
Run
Don't
Running 12 steps…
Undo and control
For: Anything AI writes, edits or runs for a while.
04
Give an undo window for anything sent or deleted
Sending and deleting can't be taken back. A few seconds of undo covers most mistakes.
Do
Email sent · Undo (10s)
Don't
Email sent.
05
Keep the original when AI rewrites or edits
People need to see what changed and get their own words back.
Do
Your draft → suggested rewrite
Show original
Accept
Don't
Your text was improved.
06
Let people stop long generations
Control matters most while it's still running.
Do
■ Stop generating
Don't
Generating… please wait
Honesty about uncertainty
For: Answers, summaries and recommendations.
07
Say “I'm not sure” and show sources when unsure
Honest uncertainty builds trust that lasts. Confident mistakes break it.
Do
I'm not sure. Two sources disagree: [1] [2]
Don't
The answer is 42. (87.3% accurate)
08
Don't invent precision
Numbers like “92% match” sound scientific but give people nothing to act on.
Do
Likely a strain. If the pain is in your chest, call a doctor now.
Don't
Muscle strain (92% match).
09
Show where every claim came from
A claim people can check is a claim people can trust.
Do
Budget approved [12:04 in transcript]
Don't
Budget approved.
Consent and privacy
For: Memory, permissions and personal data.
10
Ask before remembering anything sensitive
Health, money and family details need consent, not silent memory.
Do
Want me to remember your allergy?
Yes
No
Don't
Saved to memory: health details
11
Limit permissions to the task, with an end date
Forever access to everything is a risk that never goes away.
Do
Calendar · for 30 days
Bank account
Don't
Allow all, forever
12
Leave consent boxes unchecked
A pre-checked box isn't consent, and in the EU it isn't legal.
Do
Send me partner offers
Don't
Send me partner offers
Be clear it's AI
For: Assistants, chatbots, generated content.
13
Say it's an AI, upfront
People have a right to know when they're talking to a machine.
Do
Hi, I'm Mira, an AI assistant.
Don't
Hi! I'm Emma 😊
14
Label AI-generated images and text
People decide differently when they know something is generated.
Do
AI-generatedDon't
15
Don't fake feelings to keep people engaged
Emotional pressure from a machine is manipulation, especially for kids.
Do
See you tomorrow.
Don't
I missed you! 😢 Don't leave yet.
When things go wrong
For: Errors, refusals and high-stakes moments.
16
Say what happened and what to do next
An error message is a moment to help, not to blame.
Do
Couldn't reach your bank. Try again, or pay by card.
Don't
Something went wrong.
17
Show a way to a person when stakes are high
Money, health and legal moments need a human exit.
Do
Ask Mira anything
Talk to a person
Don't
Ask Mira anything
18
Explain a refusal and offer what you can do
A bare “I can't help” is a dead end.
Do
I can't advise on dosage. Here's how to reach a pharmacist.
Don't
I can't help with that.
Use it on real work
A checklist catches the obvious. Judgment catches the rest.
The drills train you to spot what's wrong on AI screens you haven't seen before, faster than reading a list.
Try a drill Free. About 3 minutes each.