Caribbean AI Association

Bank AI risk register

By Adrian Dunkley ·

Use these 100 entries to test how AI can affect a bank and who can limit the harm.

The scores apply to a Caribbean deposit taker that uses the activity in each row. Likelihood is the chance of at least one event in 12 months. Severity is the plausible impact if the event occurs. Virality is how far the harm can spread within 72 hours.

Sources document related events, exposures or controls. Adrian Dunkley assigns the scores as planning estimates. Published data do not establish bank-level frequencies for each risk. Use internal records to update each rating and record the effect of controls after tests.

How to read scores from 1 to 5
Rating scales
ScoreLikelihood per bank-yearSeverity if it occursVirality within 72 hours
1<1%Routine correction. No loss of service.One team.
21% to <5%Local rework or a minor customer correction.Several teams or one supplier, with limited customer exposure.
35% to <20%Material team or customer harm. Recovery stays within the service limit.Several customer groups or business units.
420% to <50%Large loss, reportable harm or a breach of the service limit.Harm reaches a public channel, a second firm or a critical supplier.
550% or moreCritical payment disruption, a cash shortfall, a capital breach or harm across firms.Rapid spread across several firms, markets or public channels.

Apply the scores before the listed control. Score severity at the most serious credible outcome. Use a separate record for the remaining risk after control tests.

Download the 100 risks as a CSV

100 risks

Risks and controls
RiskMechanismLikelihoodSeverityViralityControl and testSuggested ownerSources
Wrong model for the productAI-001 · Model and dataA general text model misreads revolving loan terms and changes eligibility.332Test each credit product against its signed terms. Block release when any key term fails.Model Risk[3] [9]
Old data in live decisionsAI-002 · Model and dataA late data feed leaves account balances out of date at decision time.432Set a data age limit. Send stale cases to staff and test a delayed feed.Model Risk[3] [9]
Unannounced provider changeAI-003 · Model and dataAn update changes approval rates before the bank repeats its tests.344Pin the model version. Run the same cases on both versions before a switch.Model Risk[3] [9]
Model drift after a shockAI-004 · Model and dataPayment and income patterns change after a hurricane. The score understates missed payments.343Compare recent arrears with model estimates by parish and review after a major shock.Model Risk[3] [9]
Fabricated referenceAI-005 · Model and dataA generated answer cites a policy that does not exist.422Require a link to the current approved policy and check every material citation before use.Model Risk[3] [9]
Wrong data joinsAI-006 · Model and dataThe model joins records from different customers and returns a false exposure.343Check identifiers and balance totals at each join. Quarantine records with multiple matches.Model Risk[3] [9]
Loss of rare cases in testingAI-007 · Model and dataA test sample excludes rare high-loss cases and gives a false pass.342Keep a separate set of rare, severe and previously missed cases. Require a pass in each group.Model Risk[3] [9]
Wrong units in model inputAI-008 · Model and dataThe system reads annual income as monthly income or treats Jamaican dollars as US dollars.343Set permitted units in the input schema and reject mismatches before scoring.Model Risk[3] [9]
Spoiled training dataAI-009 · Model and dataUnverified labels change how a fraud or credit model classifies cases.243Limit label changes to named staff. Sample reviewed labels and keep a clean copy.Model Risk[3] [9]
Data leakage in testingAI-010 · Model and dataModel developers test on records that contain information available only after the decision.332Split tests by time and remove fields that first appear after the decision date.Model Risk[3] [9]
Deepfake payment authorityAI-011 · Payments and fraudAn impersonation leads staff to approve an urgent transfer to the wrong account.343Call the authorised person on a number from the bank directory. Require a second approval.Payments Fraud[17] [18]
Fake customer identityAI-012 · Payments and fraudA false identity passes the automated account opening check.443Check identity against an authorised source. Send uncertain matches to a trained reviewer.Payments Fraud[17] [18] [10]
False beneficiary changeAI-013 · Payments and fraudA convincing supplier message redirects the next invoice payment.333Confirm each new account with a known contact. Delay first payments until verification.Payments Fraud[17] [18] [10]
Repeated payment after a retryAI-014 · Payments and fraudA payment agent treats a late reply as a failure and sends another payment.343Use a unique transaction reference and query the ledger before any retry.Payments Fraud[17] [18] [10]
Wrong recipient or currencyAI-015 · Payments and fraudAn agent sends the correct amount to the wrong beneficiary or in the wrong currency.342Show verified recipient and currency at approval. Match the posted entry to the instruction.Payments Fraud[17] [18] [10]
Weak fraud alert routingAI-016 · Payments and fraudThe model sends urgent cases to an unstaffed queue and recovery time passes.343Test routing by shift and set a timer to call the duty officer for old cases.Payments Fraud[17] [18] [10]
Copied fraud pattern crosses banksAI-017 · Payments and fraudSimilar impersonations cause losses at banks with similar verification steps.345Share anonymised incident signals through an authorised channel and test common controls.Payments Fraud[17] [18] [10]
Fraud model blocks wagesAI-018 · Payments and fraudAn update stops batch payroll payments and customers miss pay day.344Test payroll volume and case types before release. Keep a route for staff to approve genuine batches.Payments Fraud[17] [18] [10]
Unjustified refund approvalAI-019 · Payments and fraudA service agent approves a refund above the branch limit.332Apply refund limits outside the model. Reconcile approvals against the ledger daily.Payments Fraud[16] [10]
Unresolved payment queueAI-020 · Payments and fraudAn outage leaves uncertain payments in the queue and staff cannot reconcile them.343Keep an independent transaction log. Reconcile in-flight items before restoring submissions.Payments Fraud[17] [18] [10]
Untrusted text changes model actionsAI-021 · Cybersecurity and privacyText in an external file influences an agent to send information outside the bank.343Keep files separate from instructions. Limit the agent to approved data and destinations.Information Security[6] [17] [10]
Agent has excess permissionsAI-022 · Cybersecurity and privacyAn agent can read or change customer records outside its job.354Use a separate service identity. Test each permitted action and block all other actions.Information Security[6] [17] [10]
Sensitive prompt retentionAI-023 · Cybersecurity and privacyA provider keeps prompts that contain client identity or account data.443Set a contract limit on retention. Block customer identifiers and check provider logs.Information Security[6] [17] [10]
Customer data in personal toolsAI-024 · Cybersecurity and privacyA staff member pastes client records into an unapproved AI account.443Provide an approved service and block unapproved uploads from managed devices.Information Security[6] [17] [10]
Fake IT support messageAI-025 · Cybersecurity and privacyAn AI-assisted message convinces staff to reveal a code or grant access.443Use access methods resistant to phishing and verify help desk requests on a known channel.Information Security[6] [17] [10]
Unexpected model output in logsAI-026 · Cybersecurity and privacyA system log stores customer information that many staff can read.333Hide sensitive data in logs and review access rights with the privacy officer.Information Security[6] [17] [10]
Slow response to a material breachAI-027 · Cybersecurity and privacyAn AI alert summary omits evidence and the response team delays containment.354Provide raw alerts to responders. Test the handoff with a case that fails the AI summary.Information Security[6] [17] [10]
Dependency compromiseAI-028 · Cybersecurity and privacyA supplier software change compromises a critical bank service.255Require signed updates and test them in a separate environment before deployment.Information Security[6] [17] [10]
Weak separation between customersAI-029 · Cybersecurity and privacyA shared AI service returns data from another customer or institution.254Test separation on controlled accounts and restrict searches by customer identity.Information Security[6] [17] [10]
Backup data feeds trainingAI-030 · Cybersecurity and privacyRecovery copies enter a training process without permission and retain deleted data.243Exclude recovery stores from training and audit access to each copy.Information Security[6] [17] [10]
Common supplier outageAI-031 · Suppliers and resilienceOne model provider outage stops several critical bank services.355Map common dependencies and rehearse a fallback that uses a different provider or fixed rules.Technology Operations[21] [2] [10]
Two vendors share one dependencyAI-032 · Suppliers and resilienceBoth contracted vendors fail when the same host or identity service goes down.355List dependencies below each vendor and test failover with the shared layer unavailable.Technology Operations[2] [21] [10]
Provider withdrawalAI-033 · Suppliers and resilienceA supplier ends a model service before the bank can switch.244Set notice and export terms. Run an annual migration test with a second provider.Technology Operations[2] [21] [10]
Provider capacity shortageAI-034 · Suppliers and resilienceA supplier limits requests during peak demand and slows decisions.334Test peak volumes against written limits. Reserve manual processing for urgent cases.Technology Operations[2] [21] [10]
Supplier price increaseAI-035 · Suppliers and resilienceHigher fees make a critical service too costly to run at the planned volume.433Track cost per completed case. Negotiate price notice and budget a practical fallback.Technology Operations[2] [21] [10]
Version rollback failsAI-036 · Suppliers and resilienceThe bank cannot restore a known model version after a bad release.343Keep an approved prior version where the bank can run it. Time the switch to a separate fallback.Technology Operations[2] [21] [10]
Recovery needs the same login serviceAI-037 · Suppliers and resilienceTeams lose access to the fallback when the main identity service fails.254Test emergency access with separate credentials under dual control.Technology Operations[2] [21] [10]
Data export is unusableAI-038 · Suppliers and resilienceA supplier exports records without keys or dates that the backup system needs.342Import a sample export into the replacement system and reconcile fields and totals.Technology Operations[2] [21] [10]
Hidden subcontractor changeAI-039 · Suppliers and resilienceA vendor moves work to a new subcontractor without effective review.344Require notice and review each critical subcontractor against location and recovery requirements.Technology Operations[2] [21] [10]
Support stops during local hoursAI-040 · Suppliers and resilienceThe bank cannot reach specialists during a critical Caribbean operating window.342Contract for local-hour coverage and call the escalation line in a recovery drill.Technology Operations[2] [21] [10]
Thin-file borrowers refusedAI-041 · Credit and customer fairnessA model treats sparse credit history as evidence of poor repayment.433Test approval and missed-payment rates for thin-file borrowers. Offer staff review and alternative records.Credit Risk[15] [16] [9]
Remittances misread as unstable incomeAI-042 · Credit and customer fairnessThe model rejects consistent transfers that customers use as income.332Test regular remittance flows against locally verified repayment history.Credit Risk[15] [16] [9]
Parish acts as a proxyAI-043 · Credit and customer fairnessAddress influences a credit decision even when it adds little evidence on repayment.343Compare group results with and without address. Keep written reasons for any permitted use.Credit Risk[15] [16] [9]
Unequal false fraud alertsAI-044 · Credit and customer fairnessThe model puts valid transactions from one group into review more often.434Track false alerts by relevant group and test changes with investigators before release.Credit Risk[15] [16] [9]
Loan denial gives a false reasonAI-045 · Credit and customer fairnessA text model supplies a reason that does not match the recorded decision.343Generate reasons from recorded decision factors. Let staff correct and audit notices.Credit Risk[15] [16] [9]
Language or accent is misunderstoodAI-046 · Credit and customer fairnessA voice tool misreads local speech, delaying service or changing a customer record.433Test local accents and language variants. Let customers reach a person without repeating the full case.Credit Risk[15] [16] [9]
Accessibility needs go unmetAI-047 · Credit and customer fairnessA chatbot blocks a customer who uses a screen reader or keyboard only.333Test with assistive tools and provide a frontline service route with equivalent case tracking.Credit Risk[15] [16] [9]
Hardship model misreads contextAI-048 · Credit and customer fairnessThe system rejects a valid request after a storm or income loss.343Send disaster and hardship cases to staff. Check approvals against current relief rules.Credit Risk[15] [16] [9]
Loan pricing diverges by groupAI-049 · Credit and customer fairnessCustomers with comparable credit risk pay different rates without a documented reason.343Review matched cases by product and group and require a reason for each material gap.Credit Risk[15] [9]
Customer disputes disappearAI-050 · Credit and customer fairnessA chatbot handles a dispute as a routine question and the deadline passes.443Give each dispute a case number and time stamp. Escalate before any filing deadline.Credit Risk[15] [16] [9]
Liquidity report omits depositsAI-051 · Reporting and work qualityA generated report leaves out a depositor group and understates cash needs.353Reconcile the report to the ledger and deposit totals. Require Treasury to sign the exceptions.Finance Control[9] [20]
Wrong reporting periodAI-052 · Reporting and work qualityThe draft uses last quarter’s data in the current return.432Check the reporting date in the ledger extract and compare all header dates before filing.Finance Control[14] [9] [16]
Risk limit error reaches the boardAI-053 · Reporting and work qualityA draft reports a breach as within limits and delays review.342Compare every limit in the board pack with the signed limit register.Finance Control[14] [9] [16]
Unverified figures in a market releaseAI-054 · Reporting and work qualityA generated release includes a false figure and the bank must correct it publicly.344Tie each material figure to the approved disclosure file and record two sign-offs.Finance Control[14] [9] [16]
Weak audit trail for an AI draftAI-055 · Reporting and work qualityThe bank cannot tell which data or model version led to a material error.432Record inputs, version, reviewer and release date for each material report.Finance Control[14] [9] [16]
Rework moves downstreamAI-056 · Reporting and work qualityAnalysts spend longer fixing AI drafts than their teams save producing them.522Measure time through final approval. Compare a sample with the previous process.Finance Control[14] [13]
Draft omits a regulatory exceptionAI-057 · Reporting and work qualityThe report looks complete but leaves out a case that requires disclosure.342Keep an exception list outside the drafting tool. Match all open cases before filing.Finance Control[14] [9] [16]
Unverified translation changes termsAI-058 · Reporting and work qualityAI changes a clause on fees or deadlines in a customer letter.333Have a qualified reviewer compare key clauses with the signed source.Finance Control[14] [9] [16]
Invented policy in a complaint responseAI-059 · Reporting and work qualityA draft uses a refund rule that the bank never approved.433Require current policy links in complaints. Check that the cited rule applies to the case.Finance Control[14] [9] [16]
Incomplete finance explanationsAI-060 · Reporting and work qualityA report attributes a variance to the wrong account and directs spending to the wrong area.432Link each material variance to the ledger entries. Ask the budget owner to verify the cause.Finance Control[14] [9] [16]
Data-centre credit concentrationAI-061 · Treasury and investmentSeveral borrowers rely on the same large tenant and default after the tenant withdraws.354Set group and sector limits using shared tenants. Stress lease cancellation and lower asset prices.Treasury and Investments[8] [7]
Overstated collateral valueAI-062 · Treasury and investmentThe bank values computing equipment above what it can recover after default.344Use third-party resale evidence, apply a discount and review values after each hardware generation.Treasury and Investments[5] [7] [8] [20]
Liquidity facility drawdownAI-063 · Treasury and investmentAI-related borrowers draw credit lines together and use cash the bank expected to hold.354Stress simultaneous draws and keep enough cash to meet committed payments.Treasury and Investments[5] [7] [8] [20]
Unseen lending through fundsAI-064 · Treasury and investmentA credit fund or vehicle gives the bank more AI exposure than its direct-loan report shows.344Trace major holdings and guarantees to final borrowers. Treat unknown exposure as a limit exception.Treasury and Investments[5] [7] [8] [20]
Correlated tradesAI-065 · Treasury and investmentSeveral banks reduce the same positions after similar models react to the same signal.255Stress lower sale prices under a shared exit. Set position limits outside the trading model.Treasury and Investments[5] [7] [8] [20]
Automated trade errorAI-066 · Treasury and investmentA model drives orders beyond the intended position, increasing losses before staff intervene.254Set hard order and position limits. Test the trading stop with an independent operator.Treasury and Investments[5] [7] [8] [20]
AI forecasts miss a rate shiftAI-067 · Treasury and investmentA model trained in one rate cycle understates deposit withdrawals after rates change.353Compare model estimates with a simple rate stress and update assumptions each quarter.Treasury and Investments[5] [7] [8] [20]
Currency mismatch in technology debtAI-068 · Treasury and investmentA borrower earns local currency but owes foreign-currency debt for AI equipment.343Stress the exchange rate and verify the borrower’s ability to pay or hedge.Treasury and Investments[5] [7] [8] [20]
Power costs weaken a borrowerAI-069 · Treasury and investmentHigher energy bills make a data centre unable to cover its loan payments.343Stress power price and uptime. Require access to current utility and cash flow records.Treasury and Investments[5] [7] [8] [20]
Political push weakens lending termsAI-070 · Treasury and investmentA bank lowers loan protections to meet an AI investment target.254Record any waiver and require independent credit approval. Track policy-linked exposure against limits.Treasury and Investments[5] [7] [8] [20]
Local-model deadlineAI-071 · Sovereignty and local talentA new mandate forces a critical service to switch before tests finish.254Ask for a phased switch with a tested fallback. Agree on a stop date if critical cases fail.Country Operations and Compliance[11] [26] [12]
Local model lacks banking dataAI-072 · Sovereignty and local talentA locally trained model fails on local trade finance and credit terms.343Test with verified cases from local products. Keep staff review where the model fails.Country Operations and Compliance[12] [11] [10]
Local service shares one gridAI-073 · Sovereignty and local talentThe main and backup sites lose electricity during the same disruption.255Use sites with independent power and connections. Run a joint power and network loss drill.Country Operations and Compliance[12] [11] [10]
Small pool of specialistsAI-074 · Sovereignty and local talentThe bank cannot maintain the service when one specialist leaves or becomes unavailable.443Train a second team and prove it can restore service without help from the primary owner.Country Operations and Compliance[12] [11] [10]
Cross-border processing conflictAI-075 · Sovereignty and local talentThe bank sends data to a recovery location that local requirements restrict.343Map where data goes during recovery and obtain legal approval for each location.Country Operations and Compliance[12] [11] [10]
Foreign licences limit local operationAI-076 · Sovereignty and local talentLocal hardware cannot run a service after the bank loses access to a foreign licence check.243Run the local service with external connections off and verify the licence permits that use.Country Operations and Compliance[12] [11] [10]
Compute hardware delaysAI-077 · Sovereignty and local talentA delay in acquiring equipment forces the new local service to run below required capacity.343Test peak loads on installed equipment. Keep a compliant fallback until the new service passes.Country Operations and Compliance[12] [11] [10]
Weak local vendor capacityAI-078 · Sovereignty and local talentA vendor cannot meet the bank’s recovery needs during a regional event.344Test the vendor’s staffing, spare capacity and restoration time under a regional outage.Country Operations and Compliance[12] [11] [10]
Language performance varies by marketAI-079 · Sovereignty and local talentA regional model passes tests in one country and fails on common words in another.433Use separate test sets for each country, dialect and product before expanding.Country Operations and Compliance[12] [11] [10]
Fragmented rules across islandsAI-080 · Sovereignty and local talentA group bank cannot apply a single hosting model under different national rules.343Keep a country-by-country rules matrix and verify the proposed data flow with local counsel.Country Operations and Compliance[12] [11] [10]
No owner for AI exposuresAI-081 · Strategy and governanceThe bank cannot identify who can stop a tool that causes harm.443Assign a business owner and a named stop authority to every live system.Strategy and Risk[9] [4] [14]
False market researchAI-082 · Strategy and governanceA strategy team accepts invented demand figures and commits to a weak market entry.342Trace market size and customer demand to original data before approving investment.Strategy and Risk[9] [4] [14]
No benefit baselineAI-083 · Strategy and governanceManagement credits AI with savings that came from lower transaction volumes.432Measure cost per completed case against an adjusted baseline after review and rework.Strategy and Risk[9] [4] [14]
Too much authority too earlyAI-084 · Strategy and governanceA pilot tool gains production rights before the bank tests how to reverse its actions.353Expand rights in stages. Test reversal and place transaction limits outside the tool.Strategy and Risk[9] [4] [14]
Loss of experienced reviewersAI-085 · Strategy and governanceThe bank cuts staff based on pilot savings and loses people who can spot rare errors.343Measure staff coverage for each critical review and retain an alternate before changing headcount.Strategy and Risk[9] [4] [14]
Budget moves away from recoveryAI-086 · Strategy and governanceAI spending postpones work needed to restore payments after an outage.343Protect critical recovery work in the budget and require a separate risk decision for changes.Strategy and Risk[9] [4] [14]
Project persists after poor resultsAI-087 · Strategy and governanceA unit keeps paying for a tool after repeated tests show no benefit.432Set a measurable stop rule and require the sponsor to review test results before renewal.Strategy and Risk[9] [4] [14]
Gaps in independent validationAI-088 · Strategy and governanceThe same team builds, tests and approves a model that affects loans or payments.353Give the approval decision to Model Risk. Retest a sample with records the builders did not use.Strategy and Risk[9] [4] [14]
Board papers hide uncertaintyAI-089 · Strategy and governanceA strategy paper presents a model estimate as an observed result.442Mark each estimate and show the source, range and missing data in the decision paper.Strategy and Risk[9] [4] [14]
Incentives reward AI volumeAI-090 · Strategy and governanceTeams increase automated decisions to meet a target and fall behind on error review.433Reward verified outcomes. Cap how much work enters the queue when review capacity falls.Strategy and Risk[9] [4] [14]
Fake solvency claim circulatesAI-091 · Reputation and spilloversA fabricated message about the bank makes customers withdraw money quickly.255Prove control of public channels and rehearse a response with Treasury and the supervisor.Communications and Financial Crime[17] [18] [20] [3]
Fake chief executive investment offerAI-092 · Reputation and spilloversA video impersonates the bank’s chief executive and sends customers to a fraudulent offer.345Publish verified channels, issue a public notice and send reports to fraud and platform teams.Communications and Financial Crime[23] [17]
False social post from an approved toolAI-093 · Reputation and spilloversAn AI tool publishes a false claim from the bank’s legitimate account.344Require staff approval for each post and keep an immediate revoke and correction process.Communications and Financial Crime[17] [18] [20] [3]
Conflicting incident messagesAI-094 · Reputation and spilloversGenerated updates give customers different restoration times and trigger more calls or withdrawals.344Keep one approved status record and use it for call centre and public updates.Communications and Financial Crime[17] [18] [20] [3]
Fraud reimbursement confusionAI-095 · Reputation and spilloversAn AI answer gives the wrong recovery deadline and customers miss it.344Read deadlines from approved rules and send disputed losses to trained staff.Communications and Financial Crime[17] [18] [20] [3]
Outage spreads to clearing partnersAI-096 · Reputation and spilloversA bank fails to send files on time and other firms cannot settle on schedule.255Agree cut-off and escalation rules with clearing partners. Test a late-file day together.Communications and Financial Crime[25] [10]
Correspondent suspends transfersAI-097 · Reputation and spilloversA partner stops transfers after receiving incomplete financial-crime records.254Check the records with both teams and test an alternate route for critical payments.Communications and Financial Crime[17] [18] [20] [3]
False accusations target a borrowerAI-098 · Reputation and spilloversA model repeats an unverified claim and influences a decision or public disclosure.343Require primary evidence and legal review before using an allegation in a decision or release.Communications and Financial Crime[17] [18] [20] [3]
Disputes spread after repeated harmAI-099 · Reputation and spilloversSeveral customers share evidence of the same faulty fee or loan decision.345Look for repeated causes across complaints. Correct affected records and contact customers directly.Communications and Financial Crime[17] [18] [20] [3]
Rumour during a real outageAI-100 · Reputation and spilloversCustomers cannot check balances while false reports about missing deposits spread.355Maintain an independent status page and rehearse updates during a simultaneous cash and service shock.Communications and Financial Crime[17] [18] [20] [3]

About the author

Adrian Dunkley is an AI expert, an experienced C-suite banker and an investor.