Simple enough to use. Clear enough to trust.
The rules I design AI features by, and how AI fits into my own way of working.
Five rules I keep coming back to.
The model is the easy part to demo. The hard part is what happens when it is unsure, wrong, or talking to someone who has never used an app like this.
- 01
Show what the AI understood.
People only trust a result they can check. Voice reports keep the original voice note attached, so a manager can hear what the worker actually said, not just the AI's English summary.
Original voice note on every report - 02
Let people correct it.
Voice notes are shown as structured bullets with a review step before anything is sent. Workers can fix a wrong word or a wrong label in one tap.
+60% trust after adding the edit step (user survey) - 03
Coach, don't audit.
In testing, a supervisor felt red "missed" points were marking him wrong in front of his crew. Now points show "not yet covered" during the talk, and the missed list waits for the summary afterwards.
Changed after supervisor feedback - 04
Always have a way out.
AI will fail on a dusty site at 6am. When face check-in can't verify someone after two tries, they check in manually and their supervisor verifies it, so nobody loses a day's pay to a model.
Manual attendance fallback for supervisors - 05
Speak their language.
Most crews were not comfortable writing in English. Letting people talk in their own language, and handing managers English, is what made the feature actually get used.
82% field adoption by week 3, up from 12% (user survey)
I use AI for the first draft. Decisions stay with me.
a. Research synthesis
- Zaper
Interview and support notes went into Claude to group into themes: attendance issues, shift confusion, language barriers. I checked each theme against the original quotes before prioritising features.
- YellowZap
AI summarised app store and competitor reviews. Complaints about booking clarity and price transparency stood out, and shaped the booking-flow redesign.
b. Exploration
- Petty cash agent
For a WhatsApp-based petty cash agent I'm building, AI drafted conversation flows and edge cases like missing receipts and wrong amounts. I refined the strongest ones.
- Designs by Labeeb
For client work, AI moodboards let me test three or four visual ideas before opening Figma.
c. Prototyping
- clickfor.in Bazar
I worked with Claude to turn my design into a working, responsive page.
- Motion reels
Motion pieces built as code with AI help, instead of exported from a timeline.
AI speeds up the first draft. I still check the output against real user data and make the final design decisions myself.
my rule of thumb