Raj Villuri

Generative AI ,Digital Leadership & Strategy

[Raj Villuri @Rajvilluri]

5/5
Boards: if a regulator asked who decides what AI agents can do in your customers' accounts, would you give a name or "don't know"? 👇

#AI #Banking #AIGovernance #RiskManagement #FinTech

[Raj Villuri @Rajvilluri]

1/5
27% of bankers in a new Dallas Fed survey couldn't say whether their bank lets customers' AI agents into online banking.

Not "no". "Don't know." Often the rule sits with a vendor's default.

That's not caution. It's exposure you can't see. 🧵

[Raj Villuri @Rajvilluri]

4/5
Case study: on 9 Nov, Northern Trust and Microsoft unveiled an onboarding agent that cut data prep 83% (2.5 weeks to 3 days) at a 96.5% first-pass score, and still routes uncertain items to analysts.

Fast and safe came from one design.

[Raj Villuri @Rajvilluri]

1/5
42% of banks have no defined kill switch for an AI model that crosses its limits. 45% take 6+ months to get a model live.

Boards see two problems. It's one: banks that can't stop AI safely don't dare ship it fast. 🧵

2/5
Move 1: Tier by consequence.
Sort every AI use case by what it can do without a human. Pre-approve a fast lane for low-risk uses. Save expert review for AI that moves money or decides credit.

3/5
Move 2: Build the controls once.
One control plane for limits, logging and the kill switch. Every new agent inherits it on day one. Drill the stop before launch.

Move 3: Measure the queue. Days from approved idea to safe production.

4/5
Case study: at its 9 Nov results, ANZ's CEO outlined an enterprise AI control plane that sets boundaries for agents, halts activity outside them and makes AI decisions traceable.

Brakes in the platform, not a committee queue.

5/5
Boards: is your AI governance the reason your bank is slow, or the reason it can afford to be fast? 👇

#AI #Banking #AIGovernance #RiskManagement #FinTech

[Raj Villuri @Rajvilluri]

5/ Case: the same report says Citigroup, BofA, Wells Fargo and JPMorgan blocked agents too; Schwab's CEO said it blocks 10–20 clients a week from using them.
Blocking buys time. Proof wins.
CEOs and boards: how do you know your agents only do what customers authorised? 👇

[Raj Villuri @Rajvilluri]

1/ The Telegraph reported on 8 Nov that HSBC, Nationwide, NatWest, Barclays and Lloyds have moved to block AI agents from customer accounts. Pre-emptive, not after a breach. The real gap: no one can yet prove who authorised the agent, or who pays if it goes wrong. 🧵

[Raj Villuri @Rajvilluri]

4/ Move 3: Prove the limits of your proof.
• Say where your assurance stops
• Name what would prove you wrong
• Bring one decision on where agents may act next
Safe for balance enquiries is not the same as safe for transfers.

[Raj Villuri @Rajvilluri]

5/ Redundancy on a diagram is a claim. A tested failover is the evidence.

Boards: when your CTO says "we're resilient", what evidence do you ask to see? 👇

https://t.co/WcQzXYdbKd

[Raj Villuri @Rajvilluri]

1/ 53% of CIOs are confident in their tech partners' cyber resilience. Only 17% have real-time visibility and automated resilience across it (Accenture, 300 CIOs).

Confidence is not capability. If you can't show when it was last tested, it's a hope, not a control. 🧵

[Raj Villuri @Rajvilluri]

4/ Volunteer the limits.
Say where the claim holds and where it doesn't yet. Name what would prove you wrong.
The Bank of England and FCA say they first learned of the outage through affected firms, not Google.

[Raj Villuri @Rajvilluri]

5/ Reports on CERT-In’s advisory describe repeat attacks after incomplete remediation. Stopping one payment does not prove access was removed.
Bank and insurer boards: who can restrict a compromised route this weekend—and what evidence permits its safe return?

[Raj Villuri @Rajvilluri]

3/ Make escalation actionable.
• Which route and exposure?
• What is contained, and what still runs?
• Who must decide, and by when?
Responders need delegated authority. Boards receive material escalations—not every suspicious transaction.

[Raj Villuri @Rajvilluri]

4/ Control the payment path.
• Risk-based beneficiary cooling periods and holds
• Maker-checker approval and reconciliation
• API authentication and authorisation
A shared IP allow-list is not sufficient identity assurance. Test containment and safe restoration.

[Raj Villuri @Rajvilluri]

2/ Detect the pattern, not just the amount.
• Correlate micro-payments and larger follow-ons
• Maintain out-of-hours response
• Investigate unexpected API activity
A tiny payment alone is not proof of an attack. Context matters.

[Raj Villuri @Rajvilluri]

1/ A one-rupee payment can be a probe before a larger theft. CERT-In’s 4 Nov warning describes CRAFTY FALCON, an AI-assisted campaign targeting financial firms’ payment APIs, including during weekends. The test: who can contain the threat? 🧵