The Deposit Run That Wasn't (Until the Chatbot Agreed)
Northbridge Community Bank (NCB) has $1.4bn in deposits, of which 38% sit in accounts above the deposit-insurance cap. On a Friday afternoon, a regional newspaper publishes a story questioning NCB's commercial-real-estate concentration: 310% of total capital, against a regulatory guidance threshold of 300%. By 4pm, NCB's newly deployed AI customer-service assistant has handled 1,900 chats. Its scripted escalation rule fires only when a customer types the word 'fraud.' Customers instead type 'is my money safe' and 'should I move my savings.' The assistant answers each with a generic, reassuring response drafted for routine balance inquiries. Social media screenshots of those replies circulate by 6pm. Branch staff report unusual Saturday-morning withdrawal requests. NCB's treasurer notes it holds $210m in unencumbered liquid assets and $90m of committed but undrawn Federal Home Loan Bank capacity. The CEO wants to know whether to issue a public statement before Monday's opening.
1. Assess NCB's liquidity position under a plausible stress scenario and state whether a public statement before Monday is warranted. Show your numbers.
Read the worked answer
Liquidity math. Uninsured/above-cap deposits = 38% × $1.4bn = $532m. A realistic acute stress is not 100% flight but 15–30% of uninsured balances in the first days, amplified by social-media contagion.
- 15% outflow: $79.8m
- 25% outflow: $133m
- 30% outflow: $159.6m
Available liquidity = $210m on-balance-sheet + $90m undrawn FHLB = $300m, but FHLB advances require collateral, and NCB must also fund normal settlement and any collateral haircuts. A prudent usable figure is perhaps $260–280m.
Conclusion on coverage. Even the 30% case ($159.6m) is covered, but coverage ratio falls to roughly 1.6–1.9x, and the 15% case consumes over a quarter of usable liquidity in days. The trigger is not solvency - it is confidence and velocity. A statement is warranted if and only if it is factual, specific, and pre-cleared with the primary regulator. Vague reassurance ('your money is safe') invites the same mockery the chatbot attracted.
Recommended statement content: actual liquidity figures, insured-deposit coverage, CRE concentration with the remediation plan, and a named human contact. Silence is itself a message; a data-free statement is worse than silence.
2. The chatbot's escalation rule failed. Diagnose the design failure and propose a control that would have caught it before 6pm.
Read the worked answer
Diagnosis. The rule was built on a keyword ('fraud') rather than on intent. Customers expressing deposit-safety anxiety rarely use compliance vocabulary; they use plain language ('safe,' 'move my savings,' 'should I worry'). The system also lacked a confidence and sensitivity classifier: it treated a high-stakes, emotionally charged query as a routine balance inquiry because the retrieval layer found similar-looking text. Finally, there was no volume-anomaly tripwire - 1,900 chats in hours on a normally quiet Friday should itself have paged a human.
Controls to add.
- Intent taxonomy with a high-stakes tier covering safety, solvency, fees, and complaints; anything in that tier routes to a human or a pre-approved, legally reviewed response - never free generation.
- Semantic similarity guard: if a query's embedding sits near the 'safety/run risk' cluster, escalate regardless of keywords.
- Volume and sentiment monitoring with automatic human review when chat volume exceeds a rolling baseline by, say, 3x, or when negative-sentiment share crosses a threshold.
- Kill switch and canned-hold mode the duty manager can trigger without engineering.
- Post-incident logging so the 1,900 transcripts can be reviewed and, where needed, customers contacted proactively.
Governance point. The failure is not a model failure; it is a risk-tiering failure. The bank classified the channel as low-risk and applied low-risk controls to a high-risk use case.
3. The CEO proposes publishing the full CRE remediation plan and liquidity figures. The CFO warns this invites competitors and depositors to scrutinize weaknesses. Advise.
Read the worked answer
Frame the trade-off. The CFO's concern is real: disclosure can crystallize the very run you fear, and granular CRE loan-level data could be exploited. But the alternative - opacity during a confidence event - historically produces worse outcomes because depositors fill the information vacuum with the worst available narrative.
A workable middle path.
- Disclose aggregates, not loan-level detail: total CRE exposure, the percentage above guidance, the dollar amount of maturities in the next 12 months, and the specific actions (paydowns, syndications, charge-off reserves).
- Disclose liquidity in regulatory language: on-balance-sheet liquidity, unused secured borrowing capacity, and insured-deposit percentage - figures already reported to regulators.
- Do not disclose individual borrower identities, appraisal assumptions, or the internal watchlist.
Process conditions. Pre-clear with the primary regulator and, if NCB is public, coordinate with counsel on selective-disclosure rules. Publish simultaneously to all channels so no group gets a head start.
Decision rule. If the statement can be specific, verifiable, and consistent with regulatory filings, publish before Monday. If it cannot, do not publish a vague version - instead activate direct outreach to the largest uninsured depositors, who will call anyway.