Federated AI Governance Under Shadow AI and Approval Backlog
A bank has 14 business units and 37 registered AI use cases. The central AI review board meets monthly and takes an average of nine weeks to approve a use case; 12 use cases are stalled. An internal audit found nine unregistered AI tools, including three generative AI note-takers used in credit meetings and a marketing team's externally hosted chatbot processing customer segmentation data. Business units say the process is too slow; risk and compliance say they cannot delegate because the units lack AI expertise. The board wants a 'federated' governance model within six months. The bank operates in four jurisdictions with different AI rules, has 2,300 employees, no dedicated AI risk function, and one model risk policy written for traditional models. The CISO insists all customer-facing AI remain centrally approved. The audit committee has asked for a proposal that preserves accountability without adding headcount.
1. Design a tiered governance model with clear decision rights. Explain what is centralized, what is delegated, and how the bank avoids unmanaged shadow AI.
Read the worked answer
Use a federated-with-standards model, not pure decentralization. Central: AI policy, risk taxonomy, tiering criteria, high-risk approval, regulatory mapping across the four jurisdictions, vendor and model inventory, and incident response. Business units: use-case identification, data stewardship, first-line risk assessment, and low-risk deployment. A small central AI risk function, even without new headcount, can be formed by assigning named risk, legal, security, and data roles. Tier 1 low-risk: productivity tools with no personal data, no customer impact, and no automated decisions; register in an inventory, complete a short checklist, and deploy with unit sign-off. Tier 2 medium: personal data or decisions that affect employees or customers but with human review; central review with a 10-working-day service level and risk-function sign-off. Tier 3 high: customer-facing credit, employment, biometric, or cross-border AI; full central board, CISO, legal, and model validation. Shadow AI is addressed by a registry, network discovery, procurement gate, and an amnesty period; the nine tools should be triaged into tiers, not simply banned. The CISO's requirement is consistent with central approval for customer-facing AI.
2. Define the criteria the bank should use to assign a use case to a tier. Apply them to the three examples in the setup.
Read the worked answer
Criteria: (1) affected population and potential harm, including access to credit, employment, or essential services; (2) sensitivity and volume of personal data; (3) degree of automation and human review; (4) customer-facing or internal; (5) legal and cross-border exposure; (6) model provenance and vendor access; (7) reversibility of harm and explainability needs. Score each and require the higher tier when uncertain. The three generative note-takers in credit meetings are medium-to-high, because they process customer financial data and credit discussions, even if they only summarize; require central review, recording controls, retention limits, and no use for credit decisions. The marketing chatbot processing customer segmentation data is medium if data are pseudonymized and no automated decisions are made; it is high if it uses special-category data or makes eligibility decisions. The unregistered tools generally should be registered first, then tiered. Applying the CISO rule, any customer-facing AI is at least medium-high. This prevents business units from labeling a credit-related tool as low risk simply because it is a note-taker.
3. What metrics and reporting would show whether the federated model is working within six months?
Read the worked answer
Measure speed, coverage, and risk outcomes. Speed: median and 90th-percentile approval time by tier, with targets such as 5 days for tier 1, 10 days for tier 2, and 30 days for tier 3; backlog and stalled cases. Coverage: percentage of known AI tools registered, percentage of new procurements passing the AI gate, and number of shadow tools found per quarter. Risk: percentage of high-risk use cases with completed assessment, model validation, and human-review evidence; number of incidents; adverse-impact findings; vendor change notices reviewed. Quality: override rates, audit findings, and business satisfaction. Report to the board and audit committee monthly, with a six-month review that compares central capacity and tier distribution. If tier 3 approvals dominate, the bank may need more central expertise; if units bypass the process, simplify tier 1 and enforce procurement. A falling approval time with stable incident rates indicates success; a falling approval time with rising unregistered tools indicates that the process is being circumvented.