Banks now say customers own what their AI agents do. Your board will ask who owns yours.
What happened this week
- Nov 16: Fed Governor Barr said major banks report AI agents trying to act in retail accounts without the bank being able to verify authenticity, consent and alignment with customer instructions. His guardrails: strict authentication, explicit consent verification for specific actions, technical limits on agent capabilities, and rules- and values-based constraints. VERIFIED C01
- Nov 18: Citi told customers they are responsible for activity conducted through third-party AI agents, including if the tool is hacked (as reported); JPMorgan and Bank of America announced similar policies last month; PNC changed its terms Nov 18; Citizens emailed customers. CITED C02
- Nov 17: two hijacked releases of Microsoft's agent-framework were live on PyPI for about 55 minutes with a credential stealer aimed at cloud and CI secrets. VERIFIED C05
Five questions for the board
- Which of our agents can move money or change an account, and who approved each by name
- How does an agent action prove customer consent for that specific action
- What limits stop an agent exceeding its authority, and who can shut it off CITED C25
- Can we retrieve every agent action for an examiner, by date and customer
- If a third-party agent acts in a customer account, what does our authentication show and what do our terms say CITED C02
Control map (Ariana.Digital view, opinion)
| Control | AEGIS pillar | Evidence to hold |
|---|---|---|
| Agent registry and risk tier | Pillar 2, AI System Inventory | Register with owner, purpose, data touched |
| Per-action consent and authentication | Pillar 4, Controls and Human Oversight | Consent record linked to each action |
| Capability limits, kill switch | Pillar 4 | Limits config, shutdown test log |
| Action log and outflow alerts | Pillar 6, Monitoring and Incident Response | Retrievable log, alert history |
| Customer terms match controls | Pillar 5, Transparency and Rights | Terms review sign-off |
| Board owner and reporting line | Pillar 1, Governance Architecture | Charter, quarterly board paper |
| Guidance tracking | Pillar 7, Regulatory Intelligence and Evolution | Comment log for the Dec 16 third-party risk deadline VERIFIED C12 |
Workforce-readiness box
Only 11% of companies have fully built AI's effects into workforce plans; 35% partly, 34% are discussing it, 13% are not assessing it. VERIFIED C07 AI terms now appear in 7.2% of Indeed postings, up from 3.6% a year ago. VERIFIED C08
The people who supervise agents are a control. myndQ is AI workforce intelligence for AI-ready enterprises: Skills Circuits (assess, coach, enrich the team you have), Hiring Intelligence (job requirements become interview rounds, with human review and an evidenced decision), and TalentHub (verified specialists you cannot grow in time). PROPRIETARY C30 See the BLS role-requirement scorecard PROPRIETARY C31 and the AI readiness diagnostic. Nothing here states that any process is legally compliant.
Your next step
Read the full Friday edition at ariana.digital and bring the five questions to your next board risk committee.