Discovering Shadow AI at an Acquired Bank Under Current Model Risk Guidance
Treat shadow AI discovery in a bank acquisition as a continuous evidence and control process, not a one-time inventory. Identify AI surfaces across the acquired bank, classify each use by data sensitivity and risk, and map findings to your model risk management and third-party risk processes. Then enforce policy at runtime (allow, redact, block, escalate for human approval) so that discovery becomes a controlled transition with records examiners can review.
Sequence the work across pre-close, Day 1 and post-close
- 1
Pre-close: add AI to technology due diligence
Technology and third-party risk teams request the target's AI uses, vendor AI features and model providers. Treat answers as a hypothesis to verify, not a finished inventory.
- 2
Day 1: establish observation
Security extends visibility to the acquired bank's users and systems. The combined inventory owner is named.
- 3
Day 1: apply a baseline policy
Risk and security set one policy for both institutions, starting with the highest-sensitivity data.
- 4
Post-close: classify and map
Model risk and third-party risk teams rate each observed use and place it in the right process.
- 5
Post-close: resolve each use
Business owners sanction, restrict, replace or retire each use. Exceptions are approved by named individuals.
- 6
Ongoing: review and re-baseline
Internal audit samples the evidence. New uses trigger classification instead of waiting for the next survey.
Questions to ask when evaluating runtime enforcement
- Which AI surfaces can you observe and act on in the execution path, and which need other discovery methods?
- Can one policy be set once and applied to both the acquirer's and the acquired bank's users and systems from Day 1?
- What decision record does each allow, redact, block or escalate event produce, and can it be exported for examiners and internal audit?
- How are human-approval escalations routed, recorded and timestamped?
- Can reporting separate sanctioned from unsanctioned use?
- What retention controls and data residency options apply to evidence and to inspected or redacted content, and where does it run?
Start with the current guidance, not SR 11-7
Many integration plans still cite SR 11-7 and OCC Bulletin 2011-12. As of November 2026, that is out of date. On April 17, 2026, the OCC issued updated model risk management guidance that rescinds prior issuances, including OCC Bulletin 2011-12. The Federal Reserve has issued SR 26-2, revised guidance that updates its model risk management guidance based on supervisory experience since SR 11-7. Do not cite the 2011 documents as current requirements in examiner-facing material.
Both current documents stress risk-based, tailored practices commensurate with size, complexity and extent of model use. Both are expected to be most relevant to banking organizations with more than $30 billion in total assets. Check where the acquirer, the target and the combined institution fall against that level, and confirm which agency guidance governs each legal entity. The OCC, Federal Reserve and FDIC have also said they plan a request for information on model risk management that considers AI, including generative and agentic AI, so expectations may shift.
Two other hooks matter for integration. Appendix F to Part 225 requires a bank holding company to ensure each subsidiary is covered by a comprehensive information security program, and to adjust that program for changes including mergers and acquisitions. Earlier supervisory text also called for a specific party to maintain a firm-wide inventory of models. That text is historical, but a single named owner for the combined inventory remains sound practice.
Identify the AI surfaces in the acquired bank
Questionnaires and point-in-time scans miss most shadow AI because use changes faster than surveys are completed. Combine several methods, and expect each to cover only part of the picture. Typical categories are below. The methods listed are general starting points that your security team will adapt to its own tooling.
- Public generative AI sites: Employees pasting work content into browser-based chat tools. Web egress and identity records are the usual starting point.
- Coding assistants: Editor and repository extensions used by developers. Review developer tooling and endpoint inventories.
- Embedded vendor AI features: AI added to existing vendor products. Review contracts, release notes and vendor attestations. Industry commenters note banks often have limited visibility into models inside vendor products.
- Direct model API usage: Unsanctioned API keys to model providers. Review secrets stores, code repositories and expense records.
- Agents, tools and MCP servers: Tools connected to internal data. Review the services they call and the credentials they use.
Classify by data sensitivity and risk, then map to existing processes
Not every discovered use is a model in the model risk management sense, and current guidance asks for proportionate effort. Start with two questions. What data does the use touch, and what decision or output does it influence? A coding assistant used on internal tooling and an unsanctioned API key feeding customer data to a model provider call for very different responses. Mapping uses to business functions such as customer service, fraud, anti-money laundering, credit underwriting or back-office processing helps show where consequences are highest.
Then route each finding. Uses that qualify as models belong in the firm-wide inventory with validation effort scaled to risk. Vendor-embedded features go to third-party risk, where due diligence should extend to fourth-party models and to the vendor's data, methodology and performance. Uses that only create data exposure belong with information security under the program Appendix F requires to be adjusted for the acquisition. The NIST generative AI profile offers an optional lens, including its categories for value chain and component integration and for data privacy.
To avoid a stalled backlog, give every finding one of four dispositions: sanction, restrict, replace or retire. Preserve the authority to restrict use, because an inventory without that authority only documents the problem.
Enforce policy at runtime so discovery becomes a controlled transition
Discovery shows what exists. Enforcement determines what happens next. If policy is only reviewed after the fact, unsanctioned use continues during the months a bank is being integrated. Enforced at runtime, a decision is made on each interaction: allow sanctioned use, redact sensitive data, block unsanctioned destinations, or escalate for human approval when a case needs judgment.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalate-for-approval actions. Its scope is individual interactions and tool calls. It does not replace model validation, and it does not trace entire agentic workflows end to end. Policy can be set once and applied to both institutions' users, which supports a single baseline from Day 1.
The same enforcement produces audit-ready evidence. Examiners and internal audit will want to see what was permitted, redacted, blocked or escalated, when, under which policy, and who approved exceptions. Retain the decision record for each event, the policy version in force, and the approval record for each escalation. Reporting should separate sanctioned from unsanctioned use, so the inventory owner can show the trend toward sanctioned use. Set retention and data residency to match the acquirer's information security program, particularly for inspected or redacted content that may include customer information. Trussed deploys self-managed in the customer cloud, with configurable retention and data residency. Its production experience is in regulated sectors such as insurance, pharma, healthcare and education, and it makes no claim of bank deployments.
From inherited AI use to a controlled transition
Identify
Find AI sites, coding assistants, embedded vendor features, API keys and agent or tool connections.
Classify
Rate each use by the data it touches and the risk it carries.
Map
Place findings in the model risk management inventory and third-party risk process.
Enforce
Allow, redact, block or escalate at runtime, and keep the decision record.
Plan shadow AI discovery before the integration timeline sets it for you
Talk through your acquisition scenario, the AI surfaces you expect, and what evidence your examiners will ask to see.
Talk to an Expert