Verity — AI Trust UX

01
Situation

A sustainability manager, 40 requirements, 6 weeks to audit

40

compliance requirements across multiple frameworks

15+

suppliers sending documents in different formats

6 wks

to collect, verify, and submit to external auditors

A sustainability manager at Muller GmbH collects certificates, energy reports, and supplier audits from 15 sources and submits them to external auditors every year. Today this happens over email and shared drives. One missing document fails the entire audit.

Verity centralises this. It uses AI to verify whether uploaded documents satisfy each compliance requirement. The product problem is solved. The design problem is not.

02
Problem

The AI is sometimes wrong. Acting on a wrong verdict in compliance has real consequences.

A failed audit. A regulatory penalty. Reputational damage for the company. The naive solution is a confidence score. But a number without explanation is not trustworthy. It asks the reviewer to trust a verdict without understanding the reasoning.

Trust is not a UI component.
It is a system of patterns that work together.

The answer is six patterns that together make AI reasoning transparent, override natural, and the audit trail automatic. Remove any one of them and trust breaks somewhere in the flow.

03
Approach

Three principles, every decision

Show evidence, not verdicts

Surface the raw document text the AI used. Reviewers verify the source, not just the conclusion.

Override is first-class, not an escape hatch

Human judgment is the final word. Overriding the AI is a legitimate action with its own workflow and permanent record.

Name what the AI does not know

When uncertain, the system names the exact gap and links directly to where to look.

04
Design

Context before the AI verdict

Plain language: what this requirement asks for. A checklist: exactly what valid proof must include. This frames every AI verdict before it appears.

Requirement 4.2 detail page with plain-language description, valid proof checklist, and AI review showing all 5 criteria matched

Req 4.2 — Scope 1 and 2 Emissions Certificate · verified · all 5 criteria matched

Evidence, not verdicts

Each criterion in the AI review card is expandable. A click reveals the exact text extracted from the document, the page it came from, and one sentence explaining what the AI found. The verdict doesn't have to be taken on faith. The evidence is right there.

The key decision

Every expansion is labelled "WHAT THE AI FOUND" not "WHY THIS MATCHES." Honest about what the AI did. Whether the evidence is sufficient is a human judgment.

All 5 criteria expanded, each showing extracted text, page reference, and plain-language AI explanation

All 5 criteria expanded · extracted source text · page reference · plain-language finding per criterion

Close-up of one expanded criterion showing extracted German text, page 2 section 1.1 reference, and AI finding

Extracted text in German · page 2, section 1.1 · jump to page link

The reasoning is visible, not hidden behind a spinner

Criteria resolve one by one as the AI reads the document. By the time the verdict appears, the reasoning has already built in plain sight. The result is a summary of what was already visible, not a black box output.

AI working through criteria sequentially, with extracted text appearing for each, ending in an amber flag on the methodology criterion

Live analysis · criteria resolve sequentially · amber flag emerges on methodology

When the AI is uncertain, it says exactly what it could not confirm

Not "verification failed." The exact gap: ISO 14064 is referenced but the document does not specify whether market-based or location-based Scope 2 accounting was applied. The criterion turns amber. A direct link points to the section where the AI looked.

Needs review state: 4 green criteria, 1 amber methodology flag with specific uncertainty text and two action paths

Needs review · 4 of 5 verified · methodology flagged · specific gap described · two action paths

Overriding the AI requires a reason, and that reason becomes the record

Confirmation requires a written reason. That note is logged alongside the original AI flag, in sequence, with timestamps. The AI finding is never deleted. Both layers are visible to anyone who later needs to verify that this decision was made carefully.

Certify Manual Approval modal showing the AI system flag, verification notes field, responsibility checkbox, and disabled Confirm and Approve button

Certify manual approval · verification notes required · responsibility checkbox · audit trail logs both AI flag and human decision

05
System

Six patterns. Remove one and trust breaks.

01

Extracted source text

The raw evidence, not just the conclusion. Verifiable, not assumed.

02

Page and section citation

Exactly where to look. No manual searching.

03

Graduated states

Verified, Needs review, Wrong document. Each with its own weight and call to action.

04

Specific uncertainty

Names the gap precisely. Not a status. An explanation.

05

First-class override

A legitimate action, not an emergency exit. Its own workflow.

06

Preserved AI reasoning

Override adds a layer. It never erases. Both live in the trail.

06
Question

Does the form produce the judgment, or just the record?

Requiring a written reason slows the reflex and gives the audit trail meaning. The failure mode: a reviewer who writes "verified against source" and clicks through. The log entry exists. The judgment doesn't. Design can make carelessness harder. It cannot make someone care.