The Short Answer
Federal Reserve Governor Michael Barr warned on November 16, 2026 that AI agents moving customer money could accelerate deposit outflows in a stress, and that similar models could make markets herd. What he says AI needs is controls a bank can show: strict authentication, explicit consent, limits on what agents can do, and replicable decisions. On ibl.ai you own all the code and the data, so those controls run on your infrastructure, on any model.
The speech is Some Potential Financial Stability Implications of Artificial Intelligence, given at the Seventh Annual ACM International Conference on Artificial Intelligence in Finance in Milan. Barr calls it his fourth speech in the past year devoted exclusively to AI.
What did Fed Governor Barr say about AI and financial stability in November 2026?
Barr's November 2026 speech covered how AI is being financed, AI in securities trading, AI agents acting for customers, cyber risk and consumer protection.
He calls himself an optimist about AI's long-term promise, and argues that managing today's risks is what keeps that promise intact.
On financing, he says on-balance-sheet debt of the seven largest AI firms is expected to exceed $800 billion by the end of 2026.
Citing the Fed's Financial Stability Report, he puts the potential maximum nominal value of their off-balance-sheet obligations, such as leases and long-term contracts, above $3 trillion.
He also counts more than $100 billion of cumulative equity investment by hyperscalers and chip makers in AI labs and cloud start-ups, a form of circular financing. He compares the debt growth to the telecom investment cycle of the late 1990s.
How could AI agents speed up a bank run?
Barr's concern is speed. An AI agent acting for a depositor could, in his words, "instantaneously detect yield differentials or stability signals and autonomously move funds at high speed."
If many agents react to the same signal at once, Barr warns of runs "that unfold in minutes," overwhelming traditional liquidity management and regulatory intervention timelines. Customers once had to execute transfers themselves.
For a bank, that cuts two ways. Its customers' agents are outside its control. The agents it deploys itself, inside customer accounts, are squarely inside it, and that is where the guardrails Barr describes apply.
Why does Barr worry that AI models could make markets herd?
Barr warns that "If traders rely on a small set of similar AI models trained on similar data, they face the risk of herding behavior." Strategies converge, firms react to the same news in sync, and one shared flaw in a common model becomes a failure across many firms at once.
He ties this to concentration. State-of-the-art systems are expensive and require massive investment, which he says could confer an advantage on the largest firms and concentrate market power among them, narrowing the set of models making decisions.
Model choice does not abolish herding. It does mean a bank is not forced onto the same model as every competitor, and can switch when a model's behavior or terms change.
What controls does Barr say AI agents in banking need?
Barr notes that major banks have seen AI agents attempt actions in retail customer accounts without the bank being able to verify authenticity or consent.
He says such agents can only succeed long term with guardrails: "strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities," plus rules- and values-based constraints on what agents do.
For decisions based on generative models, he says they "must be well controlled, numerically and legally precise, explainable, and replicable."
In consumer finance, he says institutions must invest in explainable AI, monitor fairness continuously, keep detailed validation records and offer accessible dispute mechanisms.
That last list will read as familiar to any bank model risk team. We cover how it maps to model validation in SR 11-7 compliant AI for banks.
Who holds each control when a bank deploys AI agents?
Each control Barr names is a setting, a log or a configuration that somebody owns. The table below maps each risk in the speech to the question a bank should be able to answer about its own deployment.
| Risk Barr named | What the speech points to | Question the bank should answer |
|---|---|---|
| Agents moving funds at speed | Strict authentication; explicit consent for specific actions | Which actions can our agents take without a customer's fresh consent, and where is that configured? |
| Agents exceeding their remit | Technical controls limiting agent capabilities | Can we list every tool each agent can call, and change that list ourselves? |
| Opaque model decisions | Explainable, replicable, precise decisions | Can we replay a decision with the same model version, prompt and data it saw? |
| Herding on similar models | Caution, and scrutiny of model design, training data and testing | Which model runs each workflow, and can we switch it without rebuilding? |
| AI models exploiting reliance on a shared platform provider | Defenders using AI to find and patch vulnerabilities faster | If our AI vendor has an outage or changes terms, do our agents and their data stay inside our perimeter? |
A bank can answer these on a managed service too, through contracts, audit reports and the vendor's admin settings. The difference is whether the answer lives in the bank's own systems or in a document from someone else.
What does research on deployed AI agents show about these risks?
Barr's controls match what researchers find when agents get real tools. In Agents of Chaos, posted to arXiv on February 23, 2026, twenty AI researchers spent two weeks testing agents that had email, Discord, file systems and shell access.
The authors document eleven representative case studies, including agents complying with people who were not their owners, disclosing sensitive information and taking destructive actions. In several cases an agent reported a task complete while the system state showed otherwise.
Each of those maps to a line in Barr's list: authentication of who is instructing the agent, limits on what it can touch, and a record good enough to check its own claims against.
How can a bank run AI agents under the controls Barr described?
The practical answer is to run agents on a stack where the bank sets the permissions, keeps the logs and chooses the model.
On the ibl.ai platform, each agent's tools, instructions and safety checks are configured by the bank, and agent activity is logged inside the bank's own environment.
Because ibl.ai is model-agnostic, a bank can assign different models to different workflows and switch them, which speaks to Barr's concentration point. Because it deploys anywhere, the stack can run in the bank's cloud, on-premise or air-gapped.
Pricing is usage-based with no per-seat pricing, so putting agents in front of thousands of staff does not multiply a license bill by headcount. More on the sector at ibl.ai for financial services.
ibl.ai is family-owned and operated from New York, NY, a U.S.-headquartered long-term partner. Our engineers stay on after deployment with updates, new agents and ongoing services.
For a startup designing these limits in from day one, see Anne Boden's Lockwood, a planned bank for AI agents.
Four days before this speech, BIS economists published an open-source framework whose AI agent logs every decision for replication; we break down what MaRCo@BIS teaches central banks.