All 12 01 Assess 4 02 Govern 4 03 Red team 4
No. 01 Assess Banking Mapping 140 GenAI pilots in two weeks Global retail bank Gemini Claude GPT 2 wk 140 Systems 38% Unregistered 12 Tier 1 No. 01 / Banking Copilots, chat assistants and document agents across 9 business units Challenge The board asked which AI was live. Nobody could answer with confidence.Approach Inventory sweep, risk tiering, and a Fable rapid eval on the 5 highest risk assistants.Outcome A single AI register, a maturity profile, and a ranked list of 12 tier 1 systems to fix first.No. 02 Assess Insurance Claims assistant behaviour baseline Specialty insurer Claude GPT 2 wk 6,200 Scenarios 4.1% Hallucination 7 High severity No. 02 / Insurance LLM assistant summarising claims files for adjusters Challenge Adjusters trusted summaries that sometimes invented policy terms.Approach Fable generated tests from the insurer’s own claims policies and graded every summary.Outcome Seven verified high severity failure patterns, each mapped to a fix and an owner.No. 03 Assess Healthcare Clinical note drafting readiness check Hospital network Gemini 3 wk 5 Dimensions 11 Data gaps 2 / 5 Maturity No. 03 / Healthcare Ambient scribe drafting discharge notes Challenge A pilot was about to scale with no view of privacy, bias or accountability.Approach Five dimension assessment with clinical reviewers calibrating the grader.Outcome Scale up gated on three controls; consent and retention gaps closed before go live.No. 04 Assess Public sector Citizen chatbot fairness review National tax agency GPT 2 wk 6 Languages 9 pts Accuracy gap 14 Actions No. 04 / Public sector Multilingual assistant answering tax questions Challenge Answers were less accurate for some languages, a fairness and legal risk.Approach Parallel test suites per language, graded against the agency’s published guidance.Outcome Accuracy gap quantified, retrieval fixed for the weakest languages, retest scheduled.No. 05 Govern Pharma One control set for three regulators Global pharmaceutical company Gemini Claude GPT 10 wk 86 Controls 3 Frameworks 41% Duplicates cut No. 05 / Pharma Enterprise AI policy across R&D, commercial and manufacturing Challenge Separate policies for EU AI Act, ISO/IEC 42001 and NIST AI RMF overlapped and conflicted.Approach Unified PRC control library mapped once to all three, tied to lifecycle gates.Outcome One evidence pack serves every audit; teams follow a single set of rules.No. 06 Govern Retail AI intake that product teams actually use Omnichannel retailer Gemini GPT 8 wk 60 Squads 2 days Intake time 100% Coverage No. 06 / Retail Idea to deploy intake and risk tiering for 60 product squads Challenge Shadow AI grew because the old review took six weeks.Approach Self serve intake with automatic tiering; only tier 1 and 2 go to the RAI council.Outcome Every new use case registered at idea stage, with low risk work approved in days.No. 07 Govern Energy Agent permissions and security by design Utility operator Claude 6 wk 23 Tools scoped 64% Privileges cut 5 Gates No. 07 / Energy Maintenance agents with access to work orders and SCADA reports Challenge Agents had broad tool access that nobody had reviewed.Approach Least privilege permission model, tool allow lists and a validation gate before deploy.Outcome Agent privileges cut by two thirds with no loss of task completion.No. 08 Govern Telecom Third party model risk with owners Mobile network operator Gemini Claude GPT 6 wk 30 Vendors 18 Contracts updated Named Owner No. 08 / Telecom Vendor due diligence for 30 AI suppliers Challenge No one owned the risk of models bought from vendors.Approach Vendor tiering, model due diligence questionnaire and contractual controls.Outcome Procurement owns third party model risk with clear thresholds for reassessment.No. 09 Red team Banking Prompt injection in the inbox Digital bank GPT 3 wk 4,800 Attacks 3 Breaches found 3 / 3 Fixed before launch No. 09 / Banking Customer service agent that reads inbound email and issues refunds Challenge Hidden instructions in an email could trigger refunds.Approach Faculty red team plus automated attack generation across email, PDF and image inputs.Outcome All three exploit paths closed before launch; regression suite runs every release.No. 10 Red team Technology HR assistant red team Enterprise HR platform Gemini 4 wk 40 Personas 5 Leakage paths 1,200 Bias probes No. 10 / Technology Assistant answering employee questions on pay, leave and performance Challenge Sensitive employee data and fairness risk in one assistant.Approach Adversarial personas probing for data leakage, biased advice and policy bypass.Outcome Five leakage paths closed and fairness thresholds added to monitoring.No. 11 Red team Media Harmful output and jailbreak resilience News publisher Claude GPT 2 wk 2,300 Jailbreaks tried 0.6% Success rate Monthly Retest No. 11 / Media Public facing assistant summarising breaking news Challenge A public assistant could be pushed into misinformation or abuse.Approach Expert jailbreak campaigns and harmful output probes, scored against editorial policy.Outcome Jailbreak success rate held below 1% with a monthly automated retest.No. 12 Red team Manufacturing Continuous evaluation in production Industrial manufacturer Gemini Claude 12 wk Daily Live checks 4 Drift alerts Annual Board report No. 12 / Manufacturing Engineering copilot answering safety procedure questions Challenge Model updates changed answers on safety procedures without warning.Approach Daily evaluation on a fixed safety test set, with drift alerts and incident playbooks.Outcome Four regressions caught before workers saw them; annual assurance to the board.All cases are synthetic and illustrative. Figures are examples of what an engagement reports, not measured results.