Case studies / AI governance
Simulated case study: fictional company (Northstar Financial Technologies), synthetic data, no real audit or client.
4. AI governance and use-case risk assessment
AI-01 support assistant: conditionally approved (limited pilot). AI-02 document analyser: not approved: must not proceed. 16 approval gates; 0 evidenced.
Why it matters to the business
The business context. In July 2026 two AI requests reached Northstar's governance process. Customer Support Response Assistant would draft replies to customer tickets for about 60 support agents. Policy and Procedure Document Analyser would let about 70 employees ask questions about internal policies, contracts and board papers. Both depend on the same third-party model provider, whose terms on data retention and training are unclear. The business wants the benefit but cannot yet answer three simple questions: where does the data go, who is accountable, and what happens when the tool is wrong.
What is at stake. The register scores “Incorrect AI-generated customer response” (R-08: 12 out of 25 before controls, recorded residual 9) and “Sensitive data exposed through an AI tool” (R-09: 15 out of 25 before controls, recorded residual 12). Northstar's stated appetite is low for putting AI in front of customers without human review and a documented risk assessment. A blanket ban wastes a useful tool. A rubber stamp exposes confidential data or sends customers wrong answers under the company's name. Neither is a decision a risk function should be proud of.
Who relies on the work. The AI Governance Committee takes the decision. The Head of Customer Support and the Chief Financial Officer own the two use cases. The Privacy Officer cares where personal data goes, and the Chief Risk Officer accepts any residual risk above the threshold.
What the business needs from this work.
- A decision framework that can say yes with conditions or no with reasons, with owners and dates on every condition.
- A risk assessment that looks past the model to the supplier, the data, the human review and the monitoring (11 risk areas here).
- A plan for the day something goes wrong: incident response and a way to switch the tool off.
What this shows about how I work. It applies a recognised framework (the NIST AI Risk Management Framework) without claiming conformity, gives a clear and different answer to two different requests, and keeps the AI work consistent with the vendor work, since the vendor-linked gates read their status from the third-party tracker. It is the discipline AI governance needs, rather than the vocabulary.
Approach
Each use case was assessed across 11 risk areas (purpose and affected people, data, human oversight, accuracy, security, vendor dependency and others) with three ratings: inherent, residual as proposed, and target after gates. Controls map to the NIST AI Risk Management Framework functions Govern, Map, Measure and Manage. Approval is a human committee decision, never automatic, and vendor-linked gates read their status from the Project 3 tracker so the two projects cannot disagree.
What is automated, and what is not
Automation here means scripts that run the same way every time on synthetic data. It is not a live connection to any real system. The table separates what a script does from what still needs judgement.
| What is automated | How | Where in the repository | What a person still does |
|---|---|---|---|
| Vendor-linked approval gates | Each gate that depends on the vendor reads its status from the third-party tracker, so the two projects cannot disagree | scripts/build_p4.py | Deciding what must be true before go-live |
| Risk assessment and register | 22 assessments (two use cases across 11 risk areas, each with inherent, proposed and target ratings) feed a risk register, control mapping, gates and monitoring measures | scripts/build_p4.py, p4_config.py | The ratings themselves, which are structured judgement |
| Policy, playbook and decision memo | Generated as PDFs and Markdown from one source | scripts/build_p4.py | The committee's decision and the legal review the playbook says it needs |
| NIST AI RMF extract | All 72 subcategories were extracted from the NIST document by script | scripts/extract_airmf_core.py | Interpreting how each applies |
Re-run it. bash scripts/build_all.sh rebuilds everything from the synthetic data in 22 steps and runs 92 automated tests. The single script is named in each row.
Decisions
| AI-01 Customer Support Response Assistant | Conditionally approved (limited pilot) | High | High | Medium | 0 of 8 |
|---|---|---|---|---|---|
| AI-02 Policy and Procedure Document Analyser | Not approved: must not proceed | High | High | Medium | 0 of 8 |
No rows match the current filters.
AI-01 may run as a limited pilot (at most 10 agents, approved knowledge-base content only, no automatic sending) once its 8 gates are evidenced. AI-02 is not approved: uploading confidential contracts and board papers to an external tier with no technical control, and relying on users to verify summaries, leaves residual risk too high.
Approval gates
| AI-01-G01 | AI-01 | Written confirmation from Lumenvale of prompt and output retention, no model training on Northstar content, and a subprocessor list, for the API tier | Chief Product Officer | 2026-11-01 | Blocked by 3 open vendor issue(s) in Project 3 |
|---|---|---|---|---|---|
| AI-01-G02 | AI-01 | Incident notification clause with a stated time limit, content and named contact | Head of Procurement | 2026-11-01 | Blocked by 1 open vendor issue(s) in Project 3 |
| AI-01-G03 | AI-01 | Privacy impact assessment completed, including cross-border handling and redaction of identifiers | Privacy Officer | 2026-11-15 | In progress |
| AI-01-G04 | AI-01 | Technical enforcement that no reply is sent without agent approval | Head of Customer Support | 2026-11-15 | Not started |
| AI-01-G05 | AI-01 | Pre-deployment evaluation on a representative ticket set with documented thresholds met (material error rate at or below 2%) | Head of Customer Support | 2026-11-15 | In progress |
| AI-01-G06 | AI-01 | Prompt injection and data leakage tests passed | IT Security Manager | 2026-11-15 | Not started |
| AI-01-G07 | AI-01 | Logging of prompt, sources, draft, final reply and approver in place with a retention period | Head of Platform Engineering | 2026-11-22 | Not started |
| AI-01-G08 | AI-01 | Monitoring plan live and the AI incident playbook rehearsed with support supervisors | AI Governance Committee | 2026-11-22 | Not started |
| AI-02-G01 | AI-02 | Written terms for the document-upload tier: retention, no model training, subprocessors, incident notification; or an alternative provider | Chief Financial Officer | 2027-01-31 | Blocked by 3 open vendor issue(s) in Project 3 |
| AI-02-G02 | AI-02 | Classification gate that blocks upload of board papers, privileged and Restricted documents | Head of Platform Engineering | 2027-01-31 | Not started |
| AI-02-G03 | AI-02 | Per-user document isolation and upload limited to documents the user may access | Head of Platform Engineering | 2027-01-31 | Not started |
| AI-02-G04 | AI-02 | Technical traceability: every answer shows the source passage and requires acknowledgement before reuse in a record | Chief Financial Officer | 2027-01-31 | Not started |
| AI-02-G05 | AI-02 | Privacy impact assessment and legal review of privilege and confidentiality risks | Privacy Officer | 2027-01-31 | Not started |
| AI-02-G06 | AI-02 | Accuracy evaluation on a representative set of policy documents with documented thresholds | Chief Financial Officer | 2027-02-28 | Not started |
| AI-02-G07 | AI-02 | Upload and access logging with monthly access review | IT Security Manager | 2027-02-28 | Not started |
| AI-02-G08 | AI-02 | Monitoring plan and AI incident playbook adopted and rehearsed | AI Governance Committee | 2027-02-28 | Not started |
No rows match the current filters.
What management should do
Close the vendor retention and no-training confirmation first; it blocks both use cases. Adopt the monitoring plan before the pilot starts. Monitoring thresholds in the plan are proposals and are not derived from data.
Uncertainty and limitations
- ISO/IEC 42001 was read at abstract level only; there is no clause mapping and no alignment or certification claim.
- The incident playbook states no legal reporting deadlines; Northstar's counsel would need to supply them.
- Two use cases are a small sample; ratings are structured judgement.
Downloads
Every file below is generated by this repository's scripts and listed in the download index.
- AI governance policyPDF · 10 KB
- AI incident response playbookPDF · 8 KB
- Monitoring and periodic review planPDF · 7 KB
- Risk assessment: AI-01PDF · 11 KB
- Risk assessment: AI-02PDF · 11 KB
- Approval decision memoPDF · 8 KB
- AI governance workbookXLSX · 29 KB
- AI risk register (CSV)CSV · 8 KB
- Approval gates (CSV)CSV · 4 KB
- NIST AI RMF subcategories, verified extract (CSV)CSV · 14 KB
Sources
- S018: Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (NIST), verified
See the full source register.