Making AI-assisted audit work reviewable—without taking judgment away.
In a five-month internship, I led Select, Upload, and Report through production, and contributed to AI Results and Cortea’s first shared design system.
Role
Product Designer Intern · 5 months
Direct ownership
Select, Upload & Report
Collaboration
Senior designer, audit stakeholders & engineering
Product principleAI can surface evidence. Auditors remain accountable for decisions.
The design challenge was to make system work legible, keep review meaningful, and preserve a path to correct mistakes.
End-to-end audit workflowWorkflow map · not product UI
01
Auditor · Led
Select
Set framework and scope
02
Client · Led
Upload
Submit and match evidence
03
Auditor · Contributed
Review
Inspect AI findings
04
Auditor · Led
Report
Share reviewed results
Company work is kept private. This workflow map explains my contribution; it is not a screenshot of Cortea’s product.
3 stagesled through to production
7 auditorsacross 3 firms in beta
~15% lowerSOP scan time in a directional comparison
The scan-time and CSAT results are small-sample beta signals, not proof of broad or design-only impact. See evidence and limitations.
01 / Cortea internship
The hard part was accountability, not automation.
Cortea helps audit teams collect evidence, identify compliance gaps, and prepare reports with AI-assisted analysis. The product spans two roles and several handoffs; a file must stay connected to its request, review, and eventual finding.
01 · System legibility
Show what the AI is doing
Auditors need to understand document processing and findings before deciding what to trust or inspect next.
02 · Human judgment
Make review consequential
Confidence can help order attention, but it cannot stand in for an auditor’s review or approval.
03 · Recovery
Keep correction possible
When a classification or evidence link is wrong, people need a clear way to correct it instead of working around a hidden error.
Design questionHow might the product accelerate analysis without making accountability invisible?
02 / Cortea internship
Clear ownership across a connected workflow.
I led three stages from design through beta and production. I contributed to the AI-results experience and the shared system alongside a senior designer. The distinction matters: this was collaborative product work, not an end-to-end solo redesign.
My scope · 5 months
Led some. Partnered on some.
01I led this
Select
Framework, audit period, and scope.
02I led this
Upload
Bulk submission, matching, and correction paths.
03I contributed
Results
Review AI findings with evidence and confidence.
04I led this
Report
Turn reviewed findings into a shareable report.
Three stages I led reached production.Results and the first design system were collaborative contributions.
The workflow connects two roles: auditors ask for evidence and review it; clients respond to requests in their assigned area. The interface has to carry context across that handoff so a file, its request, and the eventual finding do not become separate pieces of work.
Evidence can include security-test results, ICT stress tests, incident-response exercises, and a compliance self-assessment. Each request needs a clear owner and a relevant document.
01RequestAuditor
Help auditors ask for the right evidence.
I worked on a guided request flow: choose a framework and audit period, select domains, assign owners, set a deadline, and preview the message before sending it. This makes the scope and responsibility clear from the start.
Design response
Let auditors check the request and its owners before sending it.
CORTEA · CONCEPTUALA
NEW AUDIT REQUESTSet up your audit
FrameworkISO 27001
Audit periodSelect period
OwnerAssign owner
DomainsChoose audit scope
Reconstructed structure · not original product UI
Retrospective trade-off
More structure asks the auditor to make setup decisions up front; it can reduce ambiguity for the client after the request is sent.
Clear guidance
Explain the audit purpose, the recipient’s assigned area, and who to contact for help.
Organised requests
Group evidence by audit area, with a description, due date, owner, status, and linked files.
02UploadClient
Make file submission easy to follow.
I worked on the client flow for reviewing assigned requests and uploading several documents at once, by drag-and-drop or file selection. The interface shows upload progress, scanning, and classification before linking files to requests.
Design response
Show what is happening to a file and give people a way to correct it.
CORTEA · CONCEPTUALC
ASSIGNED REQUESTS / EVIDENCEUpload evidence
Add documentsDrop files here or browse
Evidence document.pdfProcessing · category being identifiedPROCESSING
Second document.pdfNo request match yetAssign
Review the match before continuing
Reconstructed structure · example filenames and labels
Retrospective trade-off
Per-file status adds information to the screen, but makes processing visible and gives clients a chance to catch a wrong match.
Matching with a fallback
When a document cannot be matched automatically, the client can assign it manually or correct its link.
Reuse evidence
Link an existing upload to another request instead of uploading it again. Show its filename, request, type, upload date, and owner.
03ReviewAuditor
Keep the request, evidence, and discussion together.
I worked on views that help auditors track request completion, completed domains, and uploaded documents. Search and status filters separate outstanding work from completed requests, while a document viewer keeps the evidence available during review.
Design response
Help auditors follow up on a specific request without losing its context.
CORTEA · CONCEPTUALA
AUDIT RESULTSFindings to review
Finding AHIGH
Finding BMED
Finding CLOW
AI-IDENTIFIED FINDINGReview the evidence
ConfidencePriority signal · not a verdict
LINKED EVIDENCE
Reconstructed structure · not original product UI
Retrospective trade-off
Confidence can help order attention, but risks being mistaken for correctness; keep human review required for every finding.
Review in context
Show a file’s policy and domain links, plus its owner. Support adding requests, reassigning owners, and replacing evidence.
Comments & activity
Keep questions about missing or required documents next to the request, alongside ownership and status changes.
Low-fidelity screen reconstructions. They explain interaction structure; they are not original product screens or recovered design files.
Access is scoped to the organisation and assigned documents. The screen studies illustrate interaction structure; sample statuses and content are not product data or measured outcomes.
04 / Cortea internship
Two decisions made the AI easier to review.
These decisions connect a documented problem to a specific interaction response. They show where I led the work, where I contributed, and what the solution could not promise.
01Upload · I led this
Make AI processing observable.
Observation
Beta auditors could not tell what the AI was doing while their files were being processed.
Decision
Show each document’s live category and processing status as it moves through the workflow.
Trade-off
More status information adds surface area, but gives auditors an early point to catch and correct a bad classification.
02Results · I contributed
Use confidence to prioritise review—not replace it.
Observation
Auditors did not have a clear place to start when reviewing a long list of AI-identified gaps.
Decision
Use confidence tags to order attention while still requiring review of every finding.
Trade-off
Confidence is a prioritisation signal, not proof of correctness; the interface must not let a high-confidence result bypass human judgement.
Interaction principleShow the system’s work, preserve the person’s judgment, and make correction a first-class path.
This connects upload status, evidence review, and recovery into one workflow rather than treating trust as a confidence badge.
05 / Cortea internship
Learn the domain before simplifying the interface.
Before shaping the experience, I learned how framework requirements connect to checks, domains, company policies, and evidence. That domain model helped me design navigation and handoffs around auditors’ work rather than treating the dashboard as a set of isolated screens.
My design process
1
Understand the audit
Learn how frameworks, checks, and company documents connect.
2
Design the flow
Map auditor and client tasks, including missing files and unmatched uploads.
I worked with a senior designer and audit stakeholders to understand terminology and workflow, then used beta observation to identify where users hesitated. Product-specific screens and internal research materials remain private.
06 / Cortea internship
Beta feedback surfaced friction; the measured signal is directional.
Seven auditors across three firms took part in beta feedback. The sessions surfaced two recurring points of hesitation. In an A/B comparison of the original and updated flow, the time to scan a company SOP against ISO 27001 was about 15% lower with the updated flow. Mean CSAT across the same seven auditors rose from 5/10 to 6/10.
~15%Lower SOP scan time
A/B comparison of original and updated flow; small-sample, directional beta signal.
5 → 6/10Mean CSAT
Across 7 beta auditors; directional signal, not a broad satisfaction claim.
7 · 3Auditors · firms
Small qualitative cohort; findings helped identify friction, not statistical certainty.
01
Beta observation
File processing felt opaque.
What it revealed
Auditors needed to see whether the AI was understanding and sorting the material before they could trust the next step.
Change made
Added live categorisation and visible status for each document as it was processed.
02
Beta observation
Results had no obvious starting point.
What it revealed
A long list of findings did not help auditors decide where to spend their attention first.
Change made
Proposed confidence tags to order the queue, while preserving review of every finding.
These are small-sample beta signals—not proof of a broad reduction in audit completion time or an effect attributable to design alone. The sessions used pre-computed audit results; at that stage, checks ran in the backend and the app was still moving toward a full self-service flow.
07 / Cortea internship
Make collaboration easier to scale.
Alongside a senior designer, I contributed to Cortea’s first design system by working on shared Figma components and joining design reviews with engineering. I can describe the collaboration and intended role of reusable patterns, but I don’t have permission to publish the actual system or handoff files. This was collaborative work, not a system I owned alone.
Shared foundations
Aa↗
Type · colour · spacing
Reusable patterns
Button
Input
Status
Table
Illustrative patterns, not Cortea’s actual components or design tokens.
08 / Cortea internship
Shipped work, with evidence boundaries intact.
The work moved from internship design into beta and production. These delivery milestones and project-level usage are useful context; neither should be confused with proof that my design alone caused business impact.
Week 10Owned stages entered beta
Select, Upload, and Report during the five-month internship.
Week 14Owned stages reached production
Delivery milestone for the three stages I led.
50+Project-reported audit reports
Across 5+ firms. This is team-level usage, not an effect attributed to my design alone.
Delivered
Select, Upload, and Report reached production; I contributed to Results and the first shared design system.
Still to measure
Time to first reviewed finding, correction rate by confidence band, upload overrides, and report comprehension would show whether the workflow improves judgment—not only throughput.
09 / Cortea internship
Design for the moment the AI is wrong.
05months
Auditing was a new domain for me. Understanding the terminology and how audit requirements, company policies, and evidence connect was challenging at first. I learned that designing for a complex domain starts with understanding people’s work before deciding what belongs on the screen—and with designing correction paths for when the system is wrong.
Lessons from my internship
What I would test next
I would measure 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 test whether the workflow supports accountable judgement, not only faster processing.
Working in a startup
Working with a senior designer, auditors, and developers helped me understand how different perspectives shape the product, and how feedback guides design decisions as the product develops.