Seven Controls for AI in Credit Union Liquidity and Deposit Pricing

An AI-assisted treasury process should preserve three separations: authoritative balance-sheet data from model features, approved risk and pricing limits from recommendations, and management decisions from automated output. Every forecast should be compared with actual behavior and remain connected to a tested contingency path.

Liquidity forecasting and deposit pricing are attractive uses for machine learning because the institution already has frequent data: balances, maturities, transaction flows, rate changes, product mix, channel activity and member behavior. A model may estimate deposit runoff, non-maturity deposit decay, rate sensitivity, cash needs or the likely response to a promotional rate. It can also compare scenarios faster than a spreadsheet cycle.

But a more precise forecast is not the same as a safe funding decision. The NCUA’s 2026 Supervisory Priorities say examiners will evaluate governance, contingency funding, strategic decision-making and whether credit unions use sound modeling, reasonable assumptions and appropriately tiered scenarios. The agency’s AI resource, updated April 28, 2026, says existing technology-neutral requirements apply and highlights internal controls, monitoring and third-party due diligence.

Those sources do not prescribe an AI architecture. The controls below translate the supervisory objectives into a practical operating design. Each institution should adapt them to its charter, products, size, complexity, risk appetite and professional advice.

1. Establish an authoritative data layer

Define the system of record for deposits, loans, investments, borrowings, cash and off-balance-sheet exposures before building a forecast. Reconcile that layer to the general ledger and core reports on a set schedule. Document how the process handles late files, duplicate accounts, product migrations, mergers, missing maturity dates and intraday changes.

Keep model features separate from authoritative balances. A derived “stable balance” or “likely runoff” variable is an estimate, not a ledger fact. Label it, version it and make its transformation reproducible. If data quality falls below an approved threshold, the forecast should stop or visibly fall back to a conservative method.

Evidence: source inventory, reconciliations, feature definitions, freshness limits, exception log and a trace from reported balance to model input.

2. Put risk appetite and pricing authority outside the model

The board and management should retain control of liquidity limits, concentration limits, minimum operating cash, borrowing authorities, product floors and ceilings, and who may change a deposit rate. Encode those decisions in a versioned policy or rules layer the model cannot rewrite.

A recommendation that breaches a limit should be rejected or routed to a named approval process. Do not silently clip it to the nearest permitted value. That hides the frequency and size of failed recommendations and weakens the evidence needed to evaluate the tool.

Evidence: approved limit table, pricing authority matrix, effective dates, machine-readable rules, boundary tests and exception approvals.

3. Use tiered scenarios and challenger assumptions

A single forecast path creates false confidence. Run at least a base case, an adverse institution-specific scenario, a market-wide stress and a combined event. Vary deposit betas, decay rates, early withdrawals, uninsured and concentrated balances, collateral availability, borrowing capacity and the speed at which management actions take effect.

Use a challenger method that does not share the same failure mode. A simpler cohort model, historical range or expert-defined stress can reveal when a complex model is precise but fragile. The NCUA’s interest-rate-risk supervisory framework is a useful reminder that risk measures, assumptions and governance must be appropriate to the institution rather than mechanically uniform.

Evidence: scenario definitions, assumption owner, rationale, challenger result, sensitivity table and management actions mapped to each threshold.

4. Reconcile every forecast to actual behavior

Measure forecast error by time horizon, product, member segment, channel and material concentration. Compare predicted and actual inflows, outflows, renewals, early withdrawals, promotional migration and rate-sensitive movement. A monthly average can conceal a model that fails on the days when liquidity matters most.

Set investigation thresholds before launch. Review both direction and magnitude: systematic underprediction of outflows deserves a different response than noisy but unbiased estimates. Keep market regime, campaign activity, operational outages and one-time events in the explanation record.

Evidence: forecast-versus-actual report, error definitions, thresholds, root-cause review, recalibration decision and owner with a due date.

5. Govern deposit-pricing recommendations as decisions

A rate recommendation should show the eligible product and member population, approved range, funding objective, assumed balance response, expected duration, cannibalization estimate and cost under alternative outcomes. Separate an estimate of member response from authority to publish or offer a rate.

Monitor what happened after the recommendation: balance gained or retained, migration from existing products, concentration, total funding cost, early runoff, complaints and exceptions. Avoid optimizing only for volume or retention. The scorecard should show whether the pricing action improved the stated funding need without creating an unexamined member or balance-sheet consequence.

Evidence: recommendation record, permitted range, decision owner, assumption set, final rate, affected population, campaign dates and post-action review.

6. Connect the model to contingency funding—and test the manual path

The forecast should inform the contingency funding plan, not replace it. Map warning levels to specific actions, decision rights, communication steps, collateral checks and funding sources. Incorporate model outage, vendor loss, stale data and a cyber event that makes normal channels unavailable.

The NCUA’s contingency-funding guidance emphasizes access to diverse funding sources and testing. Run a timed exercise that produces the core liquidity view without the AI service. Confirm people can obtain inputs, apply conservative assumptions, escalate limits and document the decision.

Evidence: warning-level map, contact and authority list, collateral and funding-source verification, outage runbook, timed exercise result and corrective actions.

7. Preserve change control, decision records and rollback

Track changes to data sources, features, assumptions, model versions, thresholds, rate tables, vendor services and user permissions. Classify materiality before release and require regression tests proportional to the impact. A forecast can change because of a data pipeline or pricing rule even when the model code is unchanged.

For each consequential recommendation, retain the inputs and versions, forecast ranges, limit checks, alternatives, decision owner, final action and later result. Define rollback triggers for missing evidence, unexplained forecast error, limit breaches, data drift or an unavailable contingency path. Roll back to a known manual or prior-version process—not merely to another untested model.

Evidence: change ticket, impact assessment, approvals, regression results, release record, decision log, rollback test and post-incident correction.

A minimum weekly evidence packet

Treasury, finance and risk leaders should be able to review one compact packet containing:

• authoritative balances, reconciliations and data-quality exceptions;
• base, adverse, market-wide and combined liquidity scenarios;
• forecast error by horizon, product and material concentration;
• deposit-pricing recommendations, approvals and post-action results;
• limit and warning-level breaches, overrides and unresolved actions;
• current funding capacity, collateral readiness and fallback status; and
• model, assumption, rules and data changes since the prior review.

If the packet cannot show what changed, why the forecast moved and who authorized the response, the process is not ready to operate at decision speed.

Run five pre-launch tests

1. Reconciliation test: rebuild the model input from source systems and reconcile it to the approved balance-sheet view.
2. Boundary test: submit recommendations at, below and above every liquidity, concentration and pricing limit.
3. Scenario test: combine rapid runoff, higher betas, unavailable funding and delayed management action.
4. Backtest: replay calm and stressed periods, compare against a challenger, and investigate the largest misses.
5. Fallback test: lose the model or vendor, produce the manual liquidity view, make an authorized decision and preserve the record.

The NIST AI Risk Management Framework organizes work around governance, mapping, measurement and management. For treasury, that means approved authority, explicit scenario context, forecast and outcome measurement, and a response path that works when the model does not.

AI can shorten the time between balance-sheet change and management insight. It should not shorten the control chain. Keep policy and authority outside the model, force recommendations through visible limits, measure forecasts against actual behavior and maintain a contingency path that management can operate without automation.