Tiffany Ho
← Back

AI evidence collection

Catching the wrong file before it’s sent: AI checks evidence against the request, so auditors don’t have to chase.

Role
Lead Product Designer
Company
Optro (formerly AuditBoard)
Period
Q4 2025
Team
Tiger team: product, design and engineering leads, plus two to four engineers
Evidence request task with AI File Checks marking population-file.xls as Needs Review, explaining that required columns are missing and 26 of 254 records fall outside the requested period.
AI File Checks flags what’s wrong with an uploaded file, and why, before it goes back to the auditor.

Overview

During an audit, auditors request evidence from stakeholders: people outside the audit team who hold the files, like invoice reports or access logs. Collecting it is slow, and the wrong file often comes back.

We added AI file checks to the evidence request. The person uploading sees right away whether their file matches the request, and can fix it before the auditor ever has to chase.

My role

  • Design lead on a tiger team: three leads (product, design, engineering) plus two to four engineers
  • Designed both sides: the auditor’s requests and the stakeholder’s upload experience
  • Kept the scope small and the design minimal, so we could build and test quickly

The problem

  • Slow communication: auditors spend hours following up on requests for files
  • The wrong files come back: each mistake means another round of back and forth
  • Repetitive work: the same requests repeat monthly, quarterly and yearly, taking time from high-value work

How I did it

Scrappy and fast, by design

Tiger team

Three leads (product, design, engineering) plus two to four engineers.

Small scope

Small design changes to do one job well, not a whole module redesign.

Build to test

Working code in front of users, not static mockups.

Prove the value

Show it can work before investing more.

The auditor’s side

Task cards, not a redesign

To test the idea fast, we added simple task cards to the bottom of each test phase: population, sample list, evidence collection and annotation. Minimal design and build, enough to learn.

Control test page with task cards for Population (file validated by AI), Sample List, Evidence Collection (request sent) and Annotate Test Samples.
Each card is one step of the control test, from requesting the population to annotating samples.
Generate Sample File dialog with methodology and sample size pre-filled from the test details.
Sampling methodology and size pre-filled from the test details.
Sample selection file generated, with a confirmation that the test sheet was prefilled.
Samples ready to test, with the test sheet prefilled.

The stakeholder’s flow

Completing the task correctly, the first time

The person who receives an evidence request uploads their files, and AI checks them against the request for scope, structure and quality.

Evidence request task asking for a report of all invoices paid from Q1 to Q3 2025, with an empty file upload area.
1 · Request received: the task shows exactly what’s needed, the sample period and the control.
AI File Checks marking the uploaded file as Needs Review, with the reasons listed.
2 · AI flags issues: missing columns and out-of-period records, explained in plain language.
AI File Checks showing Validated, with the file marked Matched.
3 · Validated: the corrected file is marked Matched and goes back to the auditor.

Key design decisions

Inside the existing task

The checks sit inside the evidence request people already use, not a new tool.

People stay in control

Results are suggestions for reviewers to act on, never automatic decisions.

An audit trail

A downloadable run log records what AI checked, which matters in a regulated setting.

Outcome

  • Launched to customers behind a feature flag, built by a small tiger team
  • It proved we could do more, so the work continued into the next quarter

What I took from it: speed matters. Building small and fast proved the value quickly, without months of investment up front.

Back toAll work See all →