ISO/IEC 42001 and Adversarial Testing for a Contact-Center AI Provider
Voice agents that can no longer be talked into reading back card numbers, and an AI management system heading for certification.
Client name withheld by agreement- Client
- Conversational AI provider for contact centers
- Region
- Seattle, Washington
- Year
- 2026 · Ongoing
- Duration
- 22 weeks
- Scope
- 4 voice and chat models · 27 CRM and payment integrations
- Practice
- AI, Cloud & Digital Trust
The challenge
What was at stake.
A conversational AI provider runs voice and chat agents that handle billing, card payments and account changes for utilities, airlines and insurers. Its enterprise buyers had started asking for ISO/IEC 42001 certification and evidence of adversarial testing in the same RFP.
The agents connect to 27 CRM and payment systems. A caller who could persuade an agent to skip verification or read back stored data would turn a convenience into a breach.
What we did
The approach, step by step.
- 01
Every model and integration registered
Four production models and 27 integrations were cataloged with their purpose, data access, risk tier and an accountable product owner.
- 02
Management system to the standard
AI policy, impact assessments, change control for prompts and models, and post-release monitoring were built to ISO/IEC 42001 and cross-referenced to the NIST AI RMF.
- 03
Voice-first red teaming
Our testers called the agents directly, using social engineering, injected instructions in CRM fields and adversarial audio to push them past verification steps.
- 04
Fixes in the call flow
Verification moved out of the model and into deterministic checks, payment data was masked before it reached the model, and high-risk actions now need a confirmed second factor.
Outcomes
What changed, in numbers.
- 22
- verification bypasses closed in voice and chat
- 27
- integrations restricted to least privilege
- 4
- models under documented governance
- 0
- card numbers exposed in final retest
Procurement used to stall on AI risk. Now we send the test results and the management system, and the questions stop.
Services delivered
Related engagements
More work in AI, Cloud & Digital Trust.
-
Denver, Colorado · 15 weeks
AI, Cloud & Digital Trust · 2025SOC 2 Type II for a Legal E-Discovery PlatformAn unqualified Type II report covering Security and Confidentiality, delivered before the Am Law 100 renewal cycle.
-
Houston, Texas · 24 weeks
AI, Cloud & Digital Trust · 2024Cloud Security Platform for a Genomics LaboratoryPetabytes of patient genomes locked behind one set of guardrails, with no slowdown in sequencing turnaround.
-
Huntsville, Alabama · 18 weeks
AI, Cloud & Digital Trust · 2025CMMC Level 2 for a Satellite Communications IntegratorEvery practice met at the C3PAO assessment, with no POA&M items carried forward.
Let’s talk about your version of this.
Share what you are up against. A senior practitioner who has delivered this kind of work will reply within one business day.