AI trust UX·2026·Concept

No Black Box

An AI reviews compliance documents, and it is sometimes wrong. So it shows the text it read, names what it couldn't confirm, and never overwrites a human decision.

Role
Research framing, UX strategy, UI
Domain
ESG compliance, enterprise SaaS
Built with
Figma, React, shadcn/ui
Type
Concept work

Context

One missing document fails the whole audit

A sustainability manager gathers certificates, energy reports and supplier audits, then hands the set to external auditors once a year. Today that happens over email and shared drives.

40Compliance requirements
15+Supplier sources
6Weeks to submit
1Gap fails the audit

Problem

A score is not an explanation

Verity's AI checks whether a document satisfies a requirement, and it is sometimes wrong. A confidence number asks the reviewer to believe a verdict they can't check. In compliance, acting on a wrong one fails the audit.

Solution

Put the evidence between the verdict and the decision

Take the middle away and the decision is a click, not a judgment.

The machine
  • A verdictSometimes wrong
On screen first
  • The exact text it read
  • The page it came from
  • What it couldn't confirm
The human
  • A decisionAccept, or override

Designed

Every requirement on one surface

Status, owner and due date for all forty, so a gap is visible long before the deadline is.

The requirements dashboard: totals for in review, need action and fulfilled, then every requirement with its status, documents, due date and owner.

Designed

Evidence, not verdicts

Every criterion opens to the exact text the AI read, the page it came from, and one line on what it found. The label says what the AI found, never why it matches. Whether the evidence is enough stays a human call.

A requirement's criteria opened: for each, the text extracted from the document, the page and section it was found on, and what the AI found.

Designed

The reasoning arrives first

Criteria resolve one at a time as the document is read. By the time the verdict lands, it's a summary of what was already on screen, not the output of a spinner.

Designed

It names what it can't confirm

Not “verification failed” but the exact gap: ISO 14064 is cited, and the document never says whether Scope 2 was market-based or location-based. The criterion is flagged for review and links to the section the AI read.

A document marked needs review: four of five criteria verified, and the methodology criterion opened with the text it found and the gap it couldn't confirm.

Designed

Override is first-class

Going against the AI takes a written reason. The note is logged beside the original flag, timestamped, and the finding is never deleted. Both layers stay visible to whoever checks this decision later.

The certify manual approval dialog: the unverified criterion, a required verification note, and a confirmation checkbox, over the document review.

Result

Six patterns. Remove one and trust breaks.

  1. Extracted source textThe raw evidence, not the conclusion.
  2. Page and section citationExactly where to look, no searching.
  3. Graduated statesVerified, needs review and wrong document, each with its own weight.
  4. Specific uncertaintyNames the gap. An explanation, not a status.
  5. First-class overrideA legitimate action with its own flow, not an emergency exit.
  6. Preserved AI reasoningOverride adds a layer and never erases one.

The open question: the written reason slows the reflex, but it can't make someone care.