Shipped workCortea · B2B SaaS · AI-assisted audit

Designing an AI-assisted audit workflow auditors could confidently stand behind.

During a five-month product-design internship, I led three stages of Cortea's audit workflow—Select, Upload, and Report—and contributed to Results. The work reached beta in week 10 and production in week 14.

Role

Product Designer Intern

Duration

Feb–Jun 2025

Ownership

3 of 4 workflow stages

Collaboration

Auditors · Design · Engineering

Product screens are proprietary. All models below use fictional content and abstracted states.

End-to-end audit workflow

My ownership is shown at each stage.

Auditor in the loop throughout

01Led

Select

Choose framework

02Led

Upload

Classify evidence

03Contributed

Results

Review findings

04Led

Report

Deliver decisions

01 · Design problem

AI performs the analysis. The auditor owns the answer.

AI could accelerate the audit, but it could not become an excuse for unreviewed decisions. The workflow needed three things.

Visibility

Show what the AI is doing while it processes evidence—not only the final output.

Judgement

Help auditors decide where to look first without suggesting that confidence equals correctness.

Control

Make correction and sign-off feel like real authority, not a ceremonial approval step.

“If the AI is wrong, who signed off on it?”
The question that guided the interaction model

02 · Key design decisions

Three moments where trust had to be designed—not assumed.

Abstracted models show the interaction logic without exposing proprietary screens.

Recreated interaction model

Live

Uploaded evidence

↥Access-control-policy.pdf

↥Incident-response.docx

↥Risk-register.xlsx

AI categorisation

Policy

Access-control-policy.pdf

Procedure

Incident-response.docx

Evidence

Risk-register.xlsx

01 · UploadI led this

Make AI processing observable.

Signal: Processing was invisible, so auditors could not verify how evidence was classified.
Response: Pair every source file with its live AI category as processing completes.

Value

Catch classification errors before analysis begins.

Recreated results queue

Potential compliance gaps

Ordered for review—not auto-approved

3 findings

Privileged access review

42% confidenceReview first →

Incident response testing

71% confidenceNeeds review →

Backup retention policy

89% confidenceNeeds review →

Every finding still requires auditor review before reporting.

02 · ResultsI contributed

Use confidence to prioritise review—not replace it.

Signal: A long results list gave auditors no sensible place to begin.
Response: Use confidence tags to order attention while requiring review of every finding.

Value

Prioritisation without automatic approval.

Recreated report structure

Executive summary

DORA readiness audit

Ready to export

68%

compliant

High priority4 actions

Medium priority7 actions

Compliant23 controls

Next action

Prioritise privileged-access review and assign an accountable owner.

03 · ReportI led this

Design the report for two reading speeds.

Signal: Executives and technical teams needed different levels of detail from the same report.
Response: Lead with risk and compliance distribution, then reveal findings and actions.

Value

Fast orientation with enough detail to act.

03 · Beta validation

Seven auditors changed what we shipped.

Two recurring hesitations in beta became two concrete product changes.

01Observation → change

Observed

File processing felt opaque.

Changed

Show live categories as each document moves through processing.

Guardrail

Preserve the original filename and status.

02Observation → change

Observed

Every finding looked equally urgent.

Changed

Use confidence tags to create a review starting point.

Guardrail

Every finding still requires approval.

Participants

7 beta auditors

Organisations

3 audit firms

Method

Task-based beta observation

04 · Delivery and impact

Work that moved beyond the prototype.

The strongest evidence is delivery and use. These are project-context figures—not causal design claims.

owned workflow stages

Select, Upload, and Report

Week 10

entered beta

During a five-month internship

Week 14

reached production

All three owned stages

50+

reports delivered

Using the report template across 5+ firms

Shared system contribution

30+ reusable Figma components.

I co-designed shared components used across the product and joined design reviews with engineering. The component work helped the team carry the same interaction patterns across workflow stages without redesigning common states each time.

05 · Reflection

What the project changed in how I design.

The challenge was not making AI look intelligent. It was making uncertainty, correction, and ownership easy to act on.

Learn the domain

Understanding the audit sequence made design decisions faster and reduced guesswork about what mattered.

Design the failure path

Low confidence, disagreement, and correction deserve as much attention as the successful AI response.

Observe hesitation

A slowed cursor or repeated check can expose uncertainty that a direct interview question never surfaces.

What I would measure next

Time to first reviewed finding, correction rate by confidence band, classification overrides during upload, and report comprehension across executive and technical readers. Those measures would show whether the workflow improves judgement—not only throughput.