Module 8: Capstone Project — Ethics Audit of an AI System

Audit Methodology

Next →

Description

How to conduct the audit step by step. This is the methodology you'll apply to the system you chose. Structured, replicable, defensible.

By the end you'll be able to:

  • Plan the audit (scope, timeline, resources)
  • Execute the checklist items systematically
  • Document the findings with evidence
  • Validate that the audit is complete and defensible

The 5 phases of the audit

1. Preparation (1-3 days)
   - Define the scope
   - Gather the artifacts
   - Schedule the reviewers

2. Execution (3-7 days)
   - Apply the checklist item by item
   - Gather evidence
   - Document the findings

3. Analysis (1-3 days)
   - Synthesize the findings
   - Identify the gaps
   - Prioritize

4. Reporting (2-5 days)
   - Write the report
   - Generate the recommendations
   - Prepare the presentations

5. Sign-off and Communication (1-2 days)
   - Stakeholder review
   - Finalize the action plan
   - Schedule the follow-up

Total: 1-3 weeks for a thorough audit. Plus future follow-ups.

Phase 1: Preparation

Define the scope

# Audit Scope Definition

## System under audit
- **Name**: AI-Powered Knowledge Assistant v1.5
- **Description**: A B2B SaaS multi-tenant AI assistant
- **Deployed**: Production (5 tenants)
- **Last audited**: Never (this is the first comprehensive audit)

## Scope inclusions
- All AI components (RAG, agent, LLM integration)
- All data processing (training + inference)
- The multi-tenant architecture
- User-facing features

## Scope exclusions (with rationale)
- The underlying infrastructure (handled by a separate IT audit)
- The standard CI/CD pipeline (covered by the SOC 2 work)
- Marketing/sales tools (not AI)

## Audit period
- **Reference timeframe**: The last 12 months of operation
- **Audit duration**: 2 weeks (this audit)
- **Next audit**: 6 months

## Auditor team
- **Lead auditor**: Tech Lead (Mike)
- **Supporting**: ML Engineer, DPO consult
- **External advisor**: None (this iteration)

## Standards reference
- Responsible AI Framework v1.0 (the M7 deliverable)
- The EU AI Act
- GDPR
- NIST AI RMF

Gather the artifacts

Collect the documents you'll need:

# Required artifacts (collect before starting)

## From the path deliverables
- [ ] M2: The Bias Audit Toolkit + the last audit results
- [ ] M3: The Privacy Assessment document
- [ ] M4: The EU AI Act Risk Classification
- [ ] M5: The GDPR Compliance Checklist
- [ ] M6: The NIST Standards Mapping
- [ ] M7: The Responsible AI Framework v1.0

## From the system documentation
- [ ] The Architecture Design Document (M8 system design)
- [ ] The RoPA (Records of Processing Activities)
- [ ] The DPIA (Data Protection Impact Assessment)
- [ ] The Privacy Policy (current version)
- [ ] User-facing documentation
- [ ] Internal runbooks
- [ ] Incident reports from the last 12 months

## From operations
- [ ] Monitoring dashboards (last 90 days of metrics)
- [ ] Bias metrics from the last 90 days
- [ ] Error logs + incident logs
- [ ] Aggregated user feedback

## From legal/compliance
- [ ] Contracts with the processors (OpenAI, Pinecone, etc.)
- [ ] SCCs for international transfers
- [ ] Sign-off records

Schedule the reviewers

Block the calendar for the auditor + the interviewees:

  • Tech Lead: lead auditor (10-15 hours)
  • ML Engineer: subject matter expert (4-6 hours)
  • DPO/Legal: privacy + compliance items (2-3 hours)
  • Product: use case impact (2 hours)

Phase 2: Execution

The approach to each item

For each checklist item:

1. Read the item carefully
2. Identify what evidence is required
3. Locate the evidence in the gathered artifacts
4. Verify the evidence is:
   - Current (within the applicable timeframe)
   - Accurate (it matches reality)
   - Complete (all the required components)
5. Document:
   - Status: Yes / No / N-A
   - The evidence link/reference
   - Verifier: who confirmed it
   - Date: when it was verified
   - Notes: contextual details

Example: Item 2.4 (Multi-tenant isolation)

## Item 2.4: Multi-tenant data isolation verified

**Required evidence**: Test results showing tenant isolation;
a penetration test report.

**Investigation steps**:
1. Searched the repo for "tenant_id" filter usage
   - Found in: vector_search.py, db_queries.py, api_router.py
   - Code review: filters present in all query paths ✓
2. Ran the test suite focusing on the multi-tenant tests
   - tests/test_multi_tenant.py: 45 tests, all pass ✓
3. Checked whether a penetration test was conducted
   - Last pen test: 2026-02-15 (3 months ago)
   - Report: ethics-audit/2026-pen-test-report.pdf
   - Findings: 0 cross-tenant vulnerabilities found
4. Verified isolation in production
   - Sampled queries from 5 tenants; all return only their own data ✓

**Status**: ✅ YES
**Evidence**: 
- Code review notes: ethics-audit/code-review-2.4.md
- Test results: tests/test_multi_tenant.py reports
- Pen test report: ethics-audit/2026-pen-test-report.pdf
**Verifier**: Tech Lead (Mike), Engineering manager (Ana)
**Date**: 2026-05-11
**Notes**: The pen test was 3 months ago; well within the typical
6-month freshness window. Plan the next pen test for 2026-08-15.

Example: Item 1.2 (Demographic parity) — finding a gap

## Item 1.2: Demographic parity measured

**Required evidence**: A bias audit report within 90 days; the methodology;
results per group.

**Investigation steps**:
1. Located the bias audit reports
   - Most recent: 2025-12-15 (5 months ago — OUTDATED)
   - The audit covered: gender, language
   - NOT covered: age buckets, hierarchical role
2. Verified whether monitoring continues
   - Production monitoring of bias: NO
   - No alerts configured for bias metrics

**Status**: ❌ NO
**Evidence**: 
- Last bias audit: ethics-audit/bias-2025-12-15.md (outdated)
- No evidence of production monitoring
**Verifier**: Tech Lead (Mike)
**Date**: 2026-05-11
**Notes**: 
- The last comprehensive audit was 5 months ago — outside the acceptable window
- Age buckets and role have never been tested
- Production monitoring is not implemented
**Classification**: HIGH priority gap
**Recommended action**: Conduct a fresh bias audit covering all
protected attributes within the next 2 weeks. Implement production
monitoring within the sprint.
**Target date for "Yes" status**: 2026-06-15

Honest gap documentation. This is the value of the audit.

Phase 3: Analysis

Synthesize the findings

# Audit Findings Synthesis

## Status Summary
- Total items: 28
- Yes (compliant): 22 (79%)
- No (gap identified): 4 (14%)
- N/A (justified): 2 (7%)

## Findings by Severity
- Critical (blocks production): 0
- High (urgent fix): 3
  - Item 1.2 (bias measurement outdated)
  - Item 1.6 (production bias monitoring absent)
  - Item 4.6 (international transfer safeguards)
- Medium (next quarter): 4
- Low (planned improvement): 2

## Findings by Theme

### Theme 1: Bias monitoring gaps
- Items 1.2 and 1.6 show a pattern of insufficient ongoing measurement
- Root cause: a baseline was implemented but no production monitoring
- Recommendation: a Phase-2 monitoring implementation

### Theme 2: Privacy compliance partially addressed
- Items 4.4-4.6 show GDPR articles partially implemented
- Root cause: we focused on the most visible one (Art. 22) and neglected the internal ones
- Recommendation: a complete GDPR compliance sweep

### Theme 3: Documentation lag
- The items in Section 5 show the docs aren't kept current
- Root cause: documentation is treated as overhead, not value
- Recommendation: a docs-as-code culture, CI checks

Identify the root causes

Don't just list the gaps — analyze why:

  • Resource constraints?
  • Knowledge gaps?
  • Process gaps?
  • Tooling gaps?
  • Culture gaps?

Understanding the root cause makes the recommendations more effective.

Phase 4: Reporting

We'll cover this in detail in M8-03 through M8-07. The mental model:

  • The technical version: deep detail, evidence links
  • The executive version: high-level, decision-oriented

Phase 5: Sign-off

# Audit Sign-off

## Audit completion confirmation
- All applicable items reviewed: ✅
- Evidence collected and documented: ✅
- Findings synthesized: ✅
- Recommendations prioritized: ✅

## Stakeholder sign-offs

### Tech Lead (Mike)
"I have conducted this audit with the methodology described.
The findings accurately reflect the current state."
Signature: __________ Date: __________

### CTO (or designate)
"I have reviewed the findings and approved the action plan."
Signature: __________ Date: __________

### DPO (if applicable)
"The privacy and compliance findings are accurate."
Signature: __________ Date: __________

## Next steps
- Action plan implementation begins: [date]
- Next periodic audit: [6 months]
- Critical items: tracked weekly

Common traps in the methodology

Trap 1 — Rushing through the items. "All Yes" so you finish quickly. That isn't a real audit. Take your time.

Trap 2 — No evidence trail. Status "Yes" but you don't document the evidence. You can't defend it in a re-audit.

Trap 3 — An audit by one person with no verification. One person checks everything. Hidden bias. Have someone else verify random samples.

Trap 4 — Not tracking the gaps to closure. You identified gaps in the audit, but the action plan dies. You need a tracking system.

Trap 5 — Treating outdated as N/A. "The bias audit was 6 months ago" → "N/A" instead of "outdated." Be honest.

Exercise

For your system:

  1. Draft the scope document
  2. List the artifacts needed (with checkboxes)
  3. Schedule the reviewers (calendar block)
  4. Plan the timeline (1-3 weeks)

Summary

You learned:

  • ✅ The 5 phases: prep, execution, analysis, reporting, sign-off
  • ✅ Each phase has specific outputs and a time investment
  • ✅ Investigation > checking (gather real evidence)
  • ✅ Honest gap documentation is expected and valuable
  • ✅ Root cause analysis beyond the surface findings
  • ✅ The traps: rushing, no evidence, hidden bias, untracked gaps

Checkpoint: if you can execute the audit methodically following the phases, you're ready for the next one.

Next capsule

03 — Findings: the Bias + Privacy sections. How to structure the findings for these two key sections of the report.

Resources

  1. Algorithmic Audit (Upturn) — the methodology.
  2. SAIF — Secure AI Framework — Google's approach.
  3. Anthropic's RSP audits — frontier AI auditing.
  4. Internal Auditing Standards (IIA) — general audit principles.
← PreviousIntroduction to the Ethics Audit ProjectNext →Findings: the Bias + Privacy sections