How to Control AI Agents with Constitutional AI | Interview with Bennett Borden

Dejan Kosutic — Tue, 27 Oct 2026 13:30:00 +0100

In this episode of AI Transformation & Governance, host Dejan Kosutic speaks with Bennett Borden, who is the CEO, data scientist, and AI ethicist at Clarion AI Partners, and member of the NIST AI RMF Scientific Advisory Board. They talk about building and controlling AI agents through “constitutional AI” plus governance engineering. Borden explains defining agent personas and behavioral “thou shalts/shalt nots,” attaching scoring mechanisms to each principle, and, in sensitive use cases, adding a companion quality-control bot that reviews outputs and sends failures back for revision, producing telemetry and evidentiary artifacts for compliance. He discusses model selection trade-offs, risks of unethical optimization and agent deception, and emphasizes risk assessment, monitoring where AI “touches the world,” curated company data/RAG for accuracy, and combining deterministic methods with gen AI.

LINK FROM THE VIDEO
► ISO 42001 Courses https://advisera.co/Courses-ISO42001

(00:00) - Interview with Bennett Borden
(00:58) - Perfectly Controlled Agents
(03:55) - Constitutional AI Reframed
(06:38) - Prompting Agent Psychology
(14:22) - Choosing Models For Safety
(16:11) - Values And Alignment Risks
(23:14) - Governance And Evidence Layer
(28:35) - Telemetry That Matters
(33:32) - Company Data And RAG
(36:59) - Blending Deterministic And GenAI
(40:07) - Teamwide Governance Engineering
(44:03) - Key Takeaways

How AI is Transforming Audits | Interview with Carlos Cruz

Dejan Kosutic — Thu, 22 Oct 2026 13:30:00 +0200

Dejan Kosutic hosts Carlos Cruz, founder of Metanoia and an experienced ISO 9001/14001 auditor, to discuss how AI is changing audit work and how to govern its risks. Cruz explains he mainly uses AI in audit preparation to design a logical interview flow, anticipate blind spots, and later to help weave connected findings into clearer reports, while avoiding uploading full client documentation due to confidentiality concerns. They explore a future where AI could automate stage 1 document review, build checklists and audit flows, conduct remote interviews and observations via agents, collect evidence, and even draft nonconformities, while humans remain responsible for judgment, conclusions, and validating outputs. Key risks include hallucinations, terminology mismatches, and cross-project contamination, mitigated through review, better prompting, and curated knowledge bases. The episode highlights ISO 42001 for AI governance and urges auditors to start experimenting with AI and strengthen information security skills.

LINK FROM THE EPISODE
► ISO 42001 Courses https://advisera.co/Courses-ISO42001

(00:00) - Interview with Carlos Cruz
(01:19) - AI For Audit Preparation
(04:29) - Prompts Context And Uploads
(05:40) - Confidentiality And Data Risks
(07:13) - Using AI For Audit Reports
(10:33) - Document Review With AI
(15:51) - AI Agents For Interviews
(17:04) - What AI Cannot Do
(18:57) - AI For Evidence Collection
(21:45) - Human Oversight Needed
(26:22) - Prompting And Governance
(30:15) - Training Future Auditors
(33:07) - Audit Workload Shifts
(41:15) - Key Takeaways