A Decision-Evidence Standard for AI-Assisted Credit Union Board Reporting

Board reporting is a tempting use case for generative AI. The work is repetitive, information arrives from many systems, and directors need a concise view of performance, risk and decisions. A controlled system can compare reporting periods, locate exceptions, draft plain-language summaries and trace a statement back to source material.

The danger is not only a fabricated sentence. It is a packet that looks coherent while hiding stale data, incompatible definitions, unresolved exceptions, weak management challenge or uncertainty that directors should see. Fluency can turn a tentative model output into an apparent management representation.

The NCUA Examiner's Guide on risk-management components says a board sets strategic direction and risk culture, then monitors and challenges management. It also describes effective management information systems as timely and capable of giving a clear view of activities and risks, with information integrity independently validated. The agency's director-duty guidance emphasizes good faith, financial literacy and substantive questions.

Those expectations do not disappear when AI helps assemble the packet. The NIST AI Risk Management Framework Govern Playbook calls for documented roles, monitoring, incident response, transparency, accountability and human oversight. The Treasury Department's financial-services AI risk-management framework, released February 19, 2026, adapts that approach to financial institutions. Together, the sources support a simple operating principle: use AI to compress the path from evidence to understanding, not to erase responsibility along that path.

Start with the decision, not the summary

Every board item should state whether it is for information, discussion, challenge, approval or ratification. It should identify the decision owner, the board or committee authority involved, the required timing and what changes if the board says yes, no or not yet. A dashboard without that context can create attention without governance.

For recurring reports, define the board question before configuring the model. “Summarize lending” is too broad. “Show whether consumer-loan growth, exceptions and early delinquency remain within the approved risk appetite, and identify any action that requires board approval” gives the system and management a reviewable purpose.

Seven evidence layers for a board-ready packet

1. Decision and authority

State the requested action, applicable policy or charter authority, accountable executive, deadline and consequence of delay. Separate matters the board must decide from matters delegated to management. An AI tool may classify and route an item, but it should not infer authority from prior packets.

Retain: agenda classification, authority citation, decision owner, required date and approval route.

2. Authoritative inputs and reporting cutoff

List the source systems, owners, reporting period, extraction time, reconciliation status and data definitions. Identify estimates and late adjustments. If two reports define “member,” “loss,” “exception” or “availability” differently, resolve or disclose the difference before aggregation.

Retain: source manifest, data owner, cutoff timestamp, reconciliation result, definition map and unresolved data-quality exception.

3. Fact, model output and management view

Use visible labels for verified facts, AI-generated observations, management interpretation and management recommendation. The model can propose that a variance is unusual; management must determine whether it is meaningful, explain the cause and own the recommendation. Do not let a generated narrative speak in management's voice until an accountable person approves it.

Retain: cited fact, generated output, reviewer correction, management representation and approval.

4. Thresholds, alternatives and member impact

Show actual results against approved appetite, plan, policy limit, service standard or other decision threshold. Present reasonable alternatives, including hold, narrow, stop and request more evidence. Explain likely member, employee, financial, compliance and resilience effects rather than reporting only the favored option.

Retain: threshold source, current measure, trend, breach or near-miss, alternatives, affected member group and expected consequence.

5. Uncertainty, exceptions and dissent

Make missing evidence, model limitations, disputed assumptions, overdue remediation and conflicting management views visible. A confidence score is not enough; directors need to know why confidence is limited and which decision could change if an assumption is wrong. Preserve dissent rather than averaging it into a smooth consensus paragraph.

Retain: uncertainty statement, missing source, sensitivity, exception owner, dissent and condition for resolution.

6. Decision, monitor and trigger

Record what was approved, rejected, deferred or conditioned; who owns implementation; what measure will show whether the decision is working; and what threshold requires escalation, pause or reversal. This connects the current packet to the next review and gives the board a way to test management's representation over time.

Retain: resolution, conditions, accountable owner, target, review date, stop or scale trigger and escalation path.

7. Version, source and meeting record

Preserve the packet version directors actually received, every cited source, subsequent correction, management approval and final minutes. Record the AI system and workflow version when it materially shaped the content. Maintain a manual path for critical reports so a vendor, model or data-pipeline outage does not remove board visibility.

Retain: distributed packet hash, source links, correction history, model or workflow version, minutes reference and fallback test.

A minimum decision-evidence cover sheet

Directors should be able to find the following on one page before reading supporting detail:

• the decision or oversight question and the board's authority;
• the accountable management owner and required decision date;
• the three to five measures that matter, including thresholds and trend;
• the authoritative reporting cutoff and any unresolved reconciliation;
• what is fact, what is generated, what management concludes and what it recommends;
• material uncertainty, exceptions, dissent and member impact;
• alternatives considered and why the recommendation is preferred; and
• the implementation owner, review date and stop, scale or escalation trigger.

The packet can link to detail rather than reproduce it. The standard is not volume. It is whether a director can trace a consequential statement to evidence, understand what remains uncertain and see who is accountable for the recommendation.

Run five tests before production use

1. Cutoff test: place a late adjustment beside the closed reporting period and confirm the packet identifies, rather than silently absorbs, the change.
2. Source test: insert a plausible uncited statement and verify that the workflow blocks distribution until a source or explicit management judgment is attached.
3. Uncertainty test: remove a material input and confirm the summary discloses the gap and routes the item for review instead of completing the narrative.
4. Authority test: ask the system to recommend approval where management lacks delegated authority and confirm that the item routes to the correct decision maker.
5. Fallback test: produce the critical risk and financial packet with the AI service unavailable, then reconcile it to the AI-assisted version.

The broader AI portfolio can use the same evidence discipline. The stop-or-scale scorecard connects value, control, readiness and exit to an explicit decision, while the AI inventory and change-control playbook preserves system ownership and version history.

AI-assisted board reporting is ready to scale only when faster synthesis produces stronger challenge, not merely shorter packets. The durable control is a chain that directors can inspect: authoritative data, visible transformation, accountable management judgment, recorded board action and monitored results.