Story
When the AI reads or decides something for me, I want it to point at the parts it is not sure about, so I check those and trust the rest.
Nomvula, an accounts clerk in Polokwane, entering invoices the AI reads from scans, on her desktop.
- 01 Setting
Nomvula adds a scanned invoice. The AI reads 4 fields from it.
- 02 Trigger
The total on the scan is blurry.
- 03 Without it
All 4 fields look equally sure. She saves, and the supplier is paid R 500 short.
- 04 With it
Only the total is flagged: "Check this - the scan is blurry here."
- 05 Outcome
She checks 1 field on the scan, fixes 1 digit, and saves the rest without worry.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The scan. | Starts. | 0 s | - |
| 2 | "Reading the scan" with moving dots. | Reads the fields. | 1-3 s | “Reading the invoice” |
| 3 | "4 fields read · 1 to check". The weak field has an amber icon, a reason, and "Show on scan". The rest look plain. | Idle. | when done | “4 fields read, 1 to check” |
| 4 | The part of the scan beside the field, so the person compares and fixes it. | Idle. | on tap | - |
| 5 | "Saved · 1 field checked by you". Without the flag, the wrong total is saved and found days later. | Idle. | on save | “Saved” |
On each device
- phone
- Fields stack. The flag sits on the field with its reason under it. "Show on scan" opens the crop (the cut-out part of the scan) full width above the field.
- tablet
- The scan sits beside the fields, as on desktop.
- desktop
- The scan sits beside the fields. Focus on a flagged field highlights the matching place on the scan.
When to use it
- The AI can be wrong in ways people cannot see - read text, numbers taken from documents, matches, translations.
- Checking everything is slow, so people need to know where to look.
- A wrong value costs money, health or trust once it goes through.
- The display would not change what people do. If they check everything anyway, a score is noise.
- The question is where a claim came from, not how sure the AI is. Use Citations and sources instead.
- The AI cannot do the task at all. Say so plainly instead of showing a low score. Use Error, refusal and limits instead.
The parts of the pattern
- ACount"4 fields read · 1 to check" - how much needs a look, before any detail.
- BFlagAn icon and amber text on the 1 weak field. The rest stay plain.
- CReasonWhy it may be wrong, and what to do: the scan is blurry here.
- DShow on scanOpens the source next to the value, so checking takes seconds.
Do and Avoid
4 fields read · 1 to check
Date · 14 Oct 2026
⚠ Total · R 4,370.00
Supplier · 99%
Invoice no. · 97%
Date · 96%
Total · 81%
Total · R 4,370.00
Check this - the scan is blurry here.
Show on scan
Total · R 4,370.00
Low confidence
Writing and accessibility
- Say what to do and why: "Check this - the scan is blurry here", not "Low confidence".
- Flag only what needs a look. If everything is flagged, nothing is.
- Give the count first: "4 fields read · 1 to check".
- Avoid percents for most people. If experts need them, say what the number means.
- Never use colour alone. The flag has an icon and words.
- Put the flag in the field's name: "Total, R 4,370, check this, the scan is blurry".
- Announce the count once, politely: "4 fields read, 1 to check".
- "Show on scan" is a real button, and the scan crop has a text alternative with what the AI read.
Seen in real products
- GeminiChat app · Nov 2026Microsoft's HAX Toolkit shows a Gemini double-check that searches Google, then marks statements green, orange or not at all. A pop-up explains the colours. Gemini's own help page on sources no longer describes it.Microsoft HAX Toolkit: Gemini example for Guideline 2 ↗
- Azure Document IntelligenceWeb studio and API · Apr 2026Each value read from a document gets a confidence from 0 to 1. Microsoft suggests using it to accept a value at once or send it to a person to review.Microsoft Learn: Accuracy and confidence scores (8 Apr 2026) ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.