Unannounced AI Features in Bank Vendor Software: A Third-Party Risk Response Plan
Treat unannounced AI features in vendor software as a change-management and evidence gap inside your existing third-party risk and model risk frameworks. Detect and inventory the features, triage them by data exposure and decision impact, reopen vendor due diligence, tighten contract and configuration controls, and enforce policy at runtime only where the bank routes AI traffic through its own policy layer. AI running inside a vendor's environment remains a contract, questionnaire and monitoring matter.
A staged response plan
- 1
1. Detect and inventory
Build a register of AI features by vendor, product, version, default state and data touched.
- 2
2. Triage
Rank by data exposure and decision impact, then set review depth and deadlines per tier.
- 3
3. Reassess the vendor
Reopen due diligence on the new functionality and decide whether model risk management applies.
- 4
4. Update contract terms
Close gaps on notice, data use, subcontractors and audit rights at the next opportunity.
- 5
5. Constrain
Disable or limit features by configuration, and enforce policy where the bank controls the path.
- 6
6. Monitor and evidence
Re-check on a schedule and file what was found, decided and enforced.
Evidence to keep for examiner review
- Inventory entry with detection date, source, version and default state
- Triage record showing data exposure, decision impact and model scoping decision
- Reopened due diligence file with vendor responses, kept apart from bank-generated evidence
- Contract review notes and any amendments on notice, data use, subcontractors and audit rights
- Configuration changes with approver and date
- Where the path is bank-controlled, enforcement records of each allow, redact, block or escalation decision
Why point-in-time review misses vendor AI
Core banking, CRM, document processing, collaboration and security vendors are adding generative AI and machine learning features through routine updates. Some arrive enabled by default and without explicit customer notice. A third-party program built on onboarding due diligence and a signed contract can pass every scheduled review and still miss the moment a product starts sending customer or confidential data to an AI function, including a vendor-hosted or subprocessor model.
This is not a new category of risk. The 2023 interagency guidance on third-party relationships (SR 23-4) describes risk management across all stages of the relationship life cycle, and SR 11-7 describes model risk management through development, implementation and use, validation, and governance. The gap is operational: the bank has no reliable trigger that reopens those processes when a vendor changes what its product does.
One point on status. In September 2026 the FDIC, OCC, FRB and NCUA issued a proposal to replace the 2023 guidance, focused on tailoring practices to the level of risk. As of November 2026 the 2023 guidance is the operative text, so tie policy language to it and track the proposal before hard-coding anything to the new wording.
Detect and triage
Vendors can introduce AI through default-on releases, feature flags, added model providers or subprocessors, and revised terms of service. Detection therefore needs several inputs rather than one. Practical candidates include vendor release notes and product bulletins, administrator console settings after each update, contractual notices, network and data-flow telemetry that shows new destinations, and direct questions to vendor relationship managers. None is complete alone, so assign an owner to each and review them on a fixed cadence.
Triage on two axes. Data exposure asks what the feature can reach: public content, internal material, confidential information or customer data, and whether that data leaves the bank's environment. Decision impact asks what the output influences: convenience tasks, staff work product, or customer-facing, credit or compliance outcomes. A feature that summarizes internal meeting notes and one that drafts customer correspondence from account data deserve different depth of review.
Also decide whether each feature falls within the bank's model definition. Generative functions may sit outside the scope that would otherwise bring specific validation and monitoring expectations, so document the scoping decision and the reasoning, whichever way it goes.
Reassess the vendor and update the contract
The 2023 guidance ties due diligence to the third party's overall risk management, including its policies, processes and internal controls, and their alignment with the bank's policies for the activity. A new AI capability changes that picture, so reopen due diligence on it. Ask whether bank or customer data reaches AI functions, whether those functions are vendor-hosted or run by a subprocessor, whether data is used beyond what the contracted activities require, and whether the feature can be disabled or limited and what its default state is after an update.
On the contract, the guidance expects terms covering the activities the third party performs, ancillary services such as software support and maintenance, and the use of the bank's information, including customer information. Effective contracts typically prohibit use and disclosure of bank and customer information by the third party and its subcontractors except as needed to provide the contracted activities or comply with law. That clause is the main lever for vendor-hosted and subprocessor AI. Seek advance notice of new AI features, model-provider and subprocessor changes, and terms-of-service changes, plus opt-in or opt-out control and audit rights.
Where a vendor model updates, consider an internal standard for advance disclosure of re-training or version changes, and for comparing the replacement against the current version on identical inputs using materiality thresholds. These are practices suggested by industry commenters, not requirements, and short deprecation notice can limit how much pre-production comparison is possible.
What runtime enforcement can and cannot do
Contracts, questionnaires and monitoring cover the vendor's side of the relationship. They do not stop a given interaction from carrying sensitive data to a model. Runtime policy enforcement does, but only for AI interactions and tool calls that the bank routes through its own policy layer. It does not reach AI executing inside a vendor's own environment. For those features, the bank's controls remain contractual terms, vendor configuration settings, audit reports, periodic meetings and testing.
The scope is wider than it first appears. Where the bank controls the path, such as its own applications calling models, agents using tools, or staff using AI sites and coding assistants, policy can be enforced on each interaction. Trussed operates in this layer: it enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit or escalate for human approval. Enforcement happens at runtime, not through review after the fact. It is relevant only where that path exists, and the triage step should record which features have it and which do not.
The tradeoff is coverage. Routing AI traffic through a bank-owned layer takes design effort and cannot be applied to features the vendor runs internally. Be explicit in the inventory about which controls are enforced by the bank and which rest on vendor assurances.
Staged response to vendor AI changes
Detect
Find AI features added to existing vendor products and record them in an inventory.
Triage
Rank each feature by the data it can reach and the decisions it can influence.
Reassess
Reopen due diligence and review contract terms on data use and subcontractors.
Constrain
Apply configuration controls and, where the bank owns the path, runtime policy enforcement.
Evidence
Keep bank-generated records separate from vendor attestations.
Show what your AI was allowed to do and what it did
Where your bank controls the execution path, Trussed enforces policy on AI interactions and tool calls and produces audit-ready evidence for examiner review. Talk to our team about where that fits in your response plan.
Talk to an Expert