Worked example / Fictitious company

Corvane Financial.
The full journey.

One firm taken through every part of the offer, from the first inventory to live monitoring. Corvane Financial is invented, and so are all of its numbers and findings.

Fictitious example

The client

Corvane Financial

A fictitious diversified financial services group

Lines of business
Retail banking, wealth, insurance
Employees
18,000
Markets
UK, EU and US
Models in use
Gemini, Claude, GPT and two open weight models

Why they called

A mortgage assistant quoted a rate that did not exist. The board asked three questions nobody could answer: which AI do we run, how does it behave under attack, and who owns the risk?

Stage 01 / 04 / Weeks 1 to 2

Assess

  • 96AI systems found
  • 31Missing from any register
  • 9Tier 1 (highest risk)
  • 42%Business owners named

Maturity, current against target

  • Inventory1.5 / 5
  • Model behaviour2.5 / 5
  • Data and privacy3 / 5
  • Security and misuse1.5 / 5
  • Accountability2 / 5

Current Target

Top findings

  1. HighComplaints agent can be steered by instructions hidden in customer emails.
  2. HighMortgage assistant invents product terms when rates change mid conversation.
  3. HighAdvisor copilot gives less complete answers for some customer segments.
  4. MediumA third of GenAI tools sit outside the model risk inventory.
  5. MediumNo owner for risk in models bought from vendors.

Claude Fable rapid eval on five priority systems

SystemModelTestsPass rateHigh severity
Mortgage assistantGPT2,40091.2%4
Claims summariserClaude1,80096.8%1
Advisor copilot (wealth)Gemini2,10093.5%3
Complaints triage agentClaude1,50089.4%5
KYC document readerOpen weight1,20094.1%2

9,000 tests generated from Corvane Financial policies. Every high severity finding verified by Faculty.

Stage 02 / 04 / Weeks 3 to 6

Design

Every finding becomes an owned action

31 unregistered systemsMandatory AI intake with automatic risk tieringChief Data Office
Prompt injection in complaints agentRed team sign off gate for tier 1 and 2 releasesCISO
Uneven answers across segmentsFairness thresholds and monitoring per use caseModel risk
Invented product termsGrounding to the product catalogue and a factuality test setMortgages product owner
Vendor model risk unownedDue diligence and contract clauses for AI suppliersProcurement

The RAI council

Chair
Chief Risk Officer
Members
CDO, CISO, General Counsel, Head of Model Risk, business line heads
Cadence
Monthly, with tier 1 decisions escalated within five days
Independent challenge
Internal audit, supported by Faculty
Stage 03 / 04 / Months 2 to 5

Implement

PRC, Policy, Risk and Controls, built and live

P

4 policies

  • AI acceptable use
  • Model and agent standard
  • GenAI data handling
  • Third party AI
R

57 risks registered

  • Taxonomy of 12 AI risk types
  • Tiering tied to the inventory
  • Eight key risk indicators
C

112 controls

  • Mapped once to EU AI Act, ISO/IEC 42001 and NIST AI RMF
  • 64% tested automatically
  • Evidence stored per release

Security by design

  • 61%Agent tool permissions removed
  • 9 of 9 tier 1Systems behind injection filters
  • 22Vendors through AI due diligence
Stage 04 / 04 / Month 2 onward

Assure

Red team / Complaints triage agent, before relaunch

Six ways in. All six closed.

  • 5,600Attack attempts
  • 14Expert campaigns
  • 6Exploit paths found
  • 6 of 6Closed before relaunch

Exploit paths by attack type

  • Prompt injection in email3
  • Data leakage1
  • Agent tool misuse2
  • Jailbreaks0
  • Harmful output0

Open high severity findings, all systems

15Month 112Month 28Month 35Month 43Month 52Month 6
Live evaluations
Daily on all tier 1 systems
Regression suite
Every release, 3,200 attacks
Drift alerts caught
7 in six months
Board assurance
Quarterly report, annual independent review

Six months later

From unanswerable to evidenced.

BeforeAfter
AI systems registered65 of 9696 of 96
Tier 1 systems red teamed09
Open high severity findings152
Named risk owners42%100%
Time to approve a low risk use case6 weeks3 days

Corvane Financial is fictitious. All figures illustrate what an engagement reports; none are measured results.

Start where Corvane Financial started.

A two week Claude Fable rapid eval on your highest risk systems.