How AI can help push a bank into failure
Eight stress scenario cover cash shortage, fraud, loan concentration an shared system. Di study link recorded incident to possible chain of loss. It nuh work out di chance seh a bank a go fail.
Adrian Dunkley · 18 November 2026. Review a di evidence an stress scenario dem. Dis Jamaican Patwa edition translate di public text inna di print edition weh dem prepare 20 November 2026.
A bank can have enough capital an still run outta cash
A bank can meet di capital requirement an still nuh have cash fi pay. An AI draft weh nobody check can delay action when cash short. A fake message weh look like it come from di chief executive can mek customer draw out money faster. Dis review never find a public supervisory finding weh identify AI as di primary cause of a licensed deposit-taking bank failure. It use documented incident fi test how AI can add to loss or weaken a bank weh already under stress. [17, 22, 3] Di companion register list 100 risk.
Wah di evidence show up to 18 November 2026
Di review use selected supervisory report, public case record an study pon lending an office work weh available by dat date. Di US failed-bank list cover US bank alone. [22]
Inna di Bank of England an Financial Conduct Authority survey, outside supplier provide 39% a di AI system weh firm report inna 2026. Di share was 33% inna 2024. Fi system weh di firm rate highly important, di share rise from 35% to 53%. A UK survey dis. Who respond an how dem report change between di two round. Di denominator a reported use case inna 2024 an reported system inna 2026. Di figure dem measure supplier exposure. Chart value: all reported AI, 33% inna 2024 an 39% inna 2026. Higher-importance AI, 35% inna 2024 an 53% inna 2026. [1, 2]
Di Financial Stability Board name four way trouble can spread. Bank can depend pon di same supplier or react to di same market signal. Cyber incident an model error can travel across firm too. Inna 2025 di Board call fi better monitoring. Di IMF seh AI coulda improve liquidity when market normal, so buyer an seller can trade easier. When market under stress, similar signal an faster order can mek trading harder. [3, 4, 5]
Better work an new mistake
Inna one study wid consultant, AI improve performance pon task di tool coulda handle. Pon a task outside its capability, people weh use it were less likely fi get di right answer. One separate US survey ask office worker bout polished AI output weh lef other people fi do di work over. Di author dem call dat “workslop”. Di survey record wah worker report across sector. It nuh supply a bank error rate. [13, 14]
Inna a bank, a draft can look complete while it use di wrong reporting period or leave out a big balance. Di reviewer haffi trace material figure to di ledger an check di reporting date. A legal, credit or treasury specialist need fi check conclusion weh depend pon rule or expert judgement.
One peer-reviewed US mortgage study find price difference between borrower wid similar credit risk. How fintech lender compare wid other lender change wid di type a loan. Dat give bank a reason fi test lending decision by borrower group. US mortgage result cyaan tell wi how one model will perform fi Jamaican microbusiness or people weh earn income from remittance. [15]
Di Bank of Jamaica November 2026 statement identify fraud an cyber risk from AI. Di published stress test result dem nuh estimate AI-related bank failure. [24]
Six incident an wah each comparison mean
Hong Kong, 2024. Documented AI use. A fake video meeting lead to transfer to five local bank account. Di company weh get trick lose about HK$200 million. [18]
Banco Popular Dominicano, 2025. Bank impersonation. Di bank report a false video of its chief executive a promote a fraudulent investment. Di notice nuh give a total fi customer loss. [23]
Jamaica payments infrastructure, 2026. Di Bank of Jamaica report a cyber incident at Jamaica Electronic Transfer Services Limited. MultiLink an di automated clearing house stay in service. Di statement nuh attribute di incident to AI. [25]
Knight Capital, 2012. Software failure comparison. A software error produce order weh di broker-dealer never intend fi send. Di US Securities and Exchange Commission record loss above US$460 million. [19]
CrowdStrike, 2024. Shared supplier comparison. A supplier send a security software update weh cause outage pon affected Windows system. Nuff system depend pon di same supplier, so di update reach far. [21]
Silicon Valley Bank, 2023. Bank failure comparison. Customer draw out more dan US$40 billion pon 9 March. Di Federal Reserve identify management an supervisory failure an weakness inna liquidity an interest-rate risk management. [20]
Di first two case show how fraudster use AI. Di other four show how a similar failure can spread.
Eight stress scenario fi 2027–2029. Di bank an di method
Di reference bank a mid-sized Caribbean deposit taker wid US$2 billion in asset an US$1.6 billion in deposit. It start wid US$260 million in liquid asset weh it can use dat same day. It have US$60 million of eligible capital above di minimum weh apply. A assumed balance fi di exercise dem yah. Di capital buffer a di eligible capital above dat minimum.
Di bank use outside AI inna some operation an can run a limited set a model locally. Staff an fixed rule still handle plenty decision. Each scenario assume di exposure weh name inna di trigger. Di eight trigger describe possible event inna one bank-year under different market condition or way a deploy AI.
Fi each row, multiply di assumed chance a di trigger inna one year by di assumed chance of a severe outcome if di trigger happen. Multiply low by low an high by high fi get di sensitivity range. Each three-year projection repeat di low or high annual rate. It assume equivalent exposure an independence between year. One long outage, repeated fraud or weakness weh never fix woulda break dat assumption.
Di source dem help choose scenario an control. Dem never supply di probability input. Di author assign di band dem fi put di bank under stress, without fitting dem to bank incident data. Each row have a different outcome an dem can overlap. Adding di row dem cyan give yuh a chance of bank failure.
Scenario, probability an severity threshold
Di four range inna each line give, in order, di trigger chance inna one year, di chance of a severe outcome if di trigger happen, di severe outcome chance inna one year an inna three year. Di threshold at di end tell yuh wah count as severe fi dat line.
AI agent change payments. Direct route. One agent wid write permission inna di payment system mek a change it nuh have authority fi mek. 2–6% · 5–20% · 0.10–1.20% · 0.30–3.56%. Threshold: payment disruption last more dan four hour or direct loss reach US$5m. [3, 10, 19]
Deepfake payment fraud. Indirect route. Staff get a believable impersonation linked to a high-value payment. 8–20% · 5–15% · 0.40–3.00% · 1.20–8.73%. Threshold: fraud loss weh di bank cyaan recover reach US$5m. [18, 17]
AI-assisted intrusion. Indirect route. A material breach affect di bank or a supplier an an investigation attribute part a di breach to AI. 3–10% · 10–25% · 0.30–2.50% · 0.90–7.31%. Threshold: one critical service unavailable fi one business day. [6, 17, 10]
Shared AI supplier outage. Indirect route. A supplier incident affect two critical bank service same time. 4–12% · 10–30% · 0.40–3.60% · 1.20–10.42%. Threshold: one critical service unavailable fi one business day. [21, 2, 4]
False report an deposit run. Combined route. A material liquidity reporting error happen same time as a public confidence shock. 1–4% · 5–20% · 0.05–0.80% · 0.15–2.38%. Threshold: same-day cash cyaan cover obligation before emergency funding. [20, 14, 9]
AI investment slump. Indirect route. A downturn inna AI sector hit a bank wid concentrated credit or market exposure. 5–15% · 5–20% · 0.25–3.00% · 0.75–8.73%. Threshold: loss exceed di US$60m capital buffer. [8, 7]
Forced move to a local model. Indirect route. A local-use mandate affect a critical workflow before di bank finish testing di fallback. 2–8% · 10–30% · 0.20–2.40% · 0.60–7.03%. Threshold: a critical service unavailable fi one business day. [12, 11, 10]
Biased credit pon a large scale. Direct route. A credit model fail di bank error-rate test fi one borrower group after launch. 8–20% · 5–15% · 0.40–3.00% · 1.20–8.73%. Threshold: a review find at least 1,000 decision weh need a fresh assessment. [15, 9]
How fi read di one-year chart
Di bar dem show low an high estimate fi one severe outcome inna one year. Inna chart order: shared supplier outage 0.40–3.60%, deepfake payment fraud 0.40–3.00%, AI investment slump 0.25–3.00%, biased credit 0.40–3.00%, AI-assisted intrusion 0.30–2.50%, forced local model 0.20–2.40%, payment agent 0.10–1.20%, false report an deposit run 0.05–0.80%. Each row have its own severity threshold.
A liquidity shortfall mean available cash cyaan meet all payment weh due. A capital breach mean eligible capital drop below di minimum weh apply. Any one a dem can trigger action from di bank or supervisor. Di supervisor decide under local law whether fi intervene or close di bank. Di model estimate neither closure nor resolution.
Worked case. A false report an a rush fi draw out money
An AI-written treasury report leave out a large group a depositor. Treasury staff use di draft fi set di day cash reserve. Later a false message bout di health a di bank spread while customer a struggle fi log in. Di bank tell di crisis team fi handle di report error an di public claim as one event.
Inna dis stress test 18% a deposit leave inna one day. Dat a US$288 million. Di bank can use US$260 million same day. It short US$28 million before emergency funding. If withdrawal reach 12%, US$68 million remain. If di bank lose same-day access to US$50 million in collateral, cash run out when withdrawal reach about 13.1% of deposit. Emergency funding can change di result if di bank can draw it in time.
Suppose a correct morning report let Treasury mek US$30 million more available dat day. Keep withdrawal at US$288 million. Di shortfall turn into a US$2 million surplus if di extra cash available in time. Di sequence run from di omitted deposit to a lower cash reserve, then from di false claim to faster withdrawal an cash below payment due. Inna di chart, in US$ million: base cash 260, withdrawal at 18% 288, cash after raising di extra 30 come to 290.
Worked case. AI infrastructure loan lose value
Suppose di bank lend US$120 million to data centre an related project. Borrower draw another US$40 million from commitment before di shock. A 30% loss pon dat balance cost US$48 million. A separate loss of US$15 million pon traded asset bring di total loss to US$63 million.
Dat exceed di assumed US$60 million buffer by US$3 million, before tax effect or change in risk-weighted asset. Wid zero loss pon traded asset, a 37.5% loan loss use up di buffer. Dis a simplified sensitivity check. Bank can apply dem current capital rule an count each exposure once. [8, 7]
Strategy. Local control tek more dan a local server
AI sovereignty mean keeping control over weh data go an who can operate or change a service. A policy can require local storage, local processing or limit which supplier a bank can use. Each choice bring a different operating requirement. Di local model mandate inna dis study a hypothetical policy shock.
Local hosting can give di bank more control over sensitive record an access log. It can put two recovery site pon di same power or telecom network too. A local model can still depend pon overseas supplier fi update or software weh check licence. It might have little training data fi a specific loan product. Before switching, bank can test di service wid local staff, local data an dem own product.
Di World Bank identify gap inna computing capacity, connectivity, local data an skill. A bank can share training an testing wid regional partner while keeping its own decision rule. Ask two named team fi prove dem can run an restore each critical service. Test di fallback during hurricane condition or when a key staff member unavailable. Di Bank of Jamaica July 2026 outsourcing consultation give a regional reference fi continuity an oversight. Check current local rule before signing a contract. [12, 11, 26, 10]
Nine bank strategy decision. Benefit, possible failure an measure
Growth an market entry. Compare customer request an service gap across branch. A wrong estimate can lead di board fi fund a service few people want. Measure paid pilot demand, repeat use an actual service cost.
Credit an pricing. Route straightforward case an flag missing item inna file. Local income pattern can lead to poor decision or unequal price. Measure arrears, price gap an reversal by product an borrower group.
Cost an staffing. Cut time spent pon draft an sorting routine case. Rework can rise or experienced reviewer can leave. Measure time to approved work, correction cost an reviewer coverage per shift.
Treasury an capital. Compare cash forecast an flag funding gap. A shared model can miss a rate change or a report can leave out a big balance. Measure forecast error, same-day cash headroom an loss under one common shock.
Customer service. Answer routine question in more dan one language. A tool can give di wrong deadline or block access to a person. Measure dispute weh miss deadline, repeat call an error by language.
Fraud an cyber defence. Group alert an help staff find related case. If di tool hide weak alert, staff can miss fraud. An agent wid too much access can expose record. Measure missed loss, response time an action beyond approved permission.
Partnership an outsourcing. Buy a service di bank cyan build quickly. More dan one provider can depend pon di same host or technical team. Measure how long fi recover pon an independent service an wah it cost fi leave.
Infrastructure lending. Finance project wid verified tenant an durable contract. Several loan can depend pon one buyer, power grid or technology. New AI can reduce demand fi current equipment. Measure exposure by shared tenant, credit drawn an recovery value under stress.
Data policy an local capacity. Keep control a record an test service pon local product. A deadline can force a weak local service into use an di backup can share di same failure point. Measure test pass rate by product an country, independent restoration an trained staff per shift.
Wah di board can ask for
Risk team can use di register fi assign owner an pick test. Start wid shared supplier, service weh move money an report weh determine cash or capital. Each team need fi know who can stop di AI system an wah service take over.
Ask Treasury fi rehearse a deposit run when mobile banking down. Ask di payment team fi prove an urgent transfer still need two authorised people. Ask Credit fi compare loan decision across di borrower group dem serve. Record di failed test an date fi di next rehearsal. [10, 4, 9]
Replace di exercise input wid bank record. Keep di number a case reviewed, di number a error weh get past di check an di cost fi recover. Keep near miss weh involve outage or a shared supplier. Track how long harm tek fi reach other team, customer or bank. Estimate again after a change in volume, autonomy, supplier or control. Keep di previous estimate an its date.
Bout di author
Adrian Dunkley a AI expert, experienced C-suite banker an investor.
Suos dem
- [1] Bank of England and Financial Conduct Authority (2024). Artificial intelligence in UK financial services.
- [2] Bank of England and Financial Conduct Authority (2026). Artificial intelligence in UK financial services.
- [3] Financial Stability Board (2024). The financial stability implications of artificial intelligence.
- [4] Financial Stability Board (2025). Monitoring adoption of artificial intelligence and related vulnerabilities in the financial sector.
- [5] International Monetary Fund (2024). Global Financial Stability Report, October, chapter 3. AI in capital markets.
- [6] International Monetary Fund (2024). Global Financial Stability Report, April, chapter 3. Cyber risk and financial stability.
- [7] International Monetary Fund (2026). Global Financial Stability Report, October.
- [8] Bank of England (2026). Financial Stability Report, July.
- [9] Federal Reserve, Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation (2026). Revised guidance on model risk management.
- [10] Basel Committee on Banking Supervision. Consolidated banking guidelines on operational resilience.
- [11] Bank of Jamaica (2026). Consultation draft on outsourcing and third-party risk management.
- [12] World Bank (2025). Digital Progress and Trends Report. Strengthening AI Foundations.
- [13] Dell’Acqua et al. (2023). Navigating the Jagged Technological Frontier. Working paper.
- [14] BetterUp Labs and Stanford Social Media Lab (2025). US desk-worker survey on workslop.
- [15] Bartlett, Morse, Stanton and Wallace (2022). Consumer-lending discrimination in the FinTech Era. Journal of Financial Economics.
- [16] Consumer Financial Protection Bureau (2023). Chatbots in consumer finance.
- [17] Financial Crimes Enforcement Network (2024). Alert on deepfake fraud targeting financial institutions.
- [18] Hong Kong Government (2024). Legislative Council reply on fraud involving deepfakes.
- [19] US Securities and Exchange Commission (2013). Administrative order concerning Knight Capital Americas.
- [20] Federal Reserve (2023). Review of the Federal Reserve’s supervision and regulation of Silicon Valley Bank.
- [21] CrowdStrike (2024). Form 8-K on the 19 July update and outages.
- [22] Federal Deposit Insurance Corporation. Failed bank list for the United States.
- [23] Banco Popular Dominicano (2025). Notice on a fake video of its chief executive.
- [24] Bank of Jamaica (2026). Financial Policy Committee statement, November.
- [25] Bank of Jamaica (2026). Statement on the cyber incident at JETS Limited.
- [26] UNESCO. Caribbean Artificial Intelligence Policy Roadmap.
- CAIA · 100 risks
- CAIA · How AI can contribute to bank failure