Many organizations have started setting rules for employee use of ChatGPT, Microsoft Copilot, and other artificial intelligence tools. These policies may tell employees which platforms are approved, what information should never be entered, and when human review is required.
That is an important step, but it addresses only part of the risk.
AI may already be operating inside the software, security platforms, customer service systems, reporting tools, and workflow applications your organization buys from outside vendors. A provider can add an AI feature through a routine software update, sometimes before customers have fully evaluated what changed.
For community banks and other regulated businesses, that creates a practical vendor-management question: Does your organization know which third parties are using AI, what those systems can access, and how the added technology changes your risk?
The Independent Community Bankers of America recently identified vendor-embedded AI, shadow AI, fourth-party dependencies, contract terms, incident response, board oversight, and cyber insurance as areas community banks should examine. Its guidance recommends building on existing third-party risk management processes rather than waiting for a separate regulatory mandate.
AI May Already Be Inside Your Vendor Environment
A vendor does not have to sell an “AI product” to introduce AI-related risk.
AI features can appear in:
- Core business applications
- Online banking and financial platforms
- Customer support systems
- Email and collaboration tools
- Security monitoring products
- Meeting transcription services
- Document-management platforms
- Accounting, legal, or healthcare software
- Automated reporting and analytics tools
- Workflow and approval systems
Some features provide summaries or recommendations. Others can generate customer communications, change records, trigger workflows, or take actions within connected systems.
The level of risk depends on what the AI can see and do. A tool that drafts a generic meeting agenda presents a different concern than one that processes customer records, reviews financial information, recommends lending actions, or interacts with production systems.
Organizations need enough visibility to make that distinction.
Traditional Vendor Questionnaires May Need an Update
Existing vendor reviews remain useful. Questions about access controls, encryption, backups, business continuity, incident notification, and security testing still matter.
AI adds another layer of due diligence.
A vendor may use an outside model provider to power an embedded feature. That provider may rely on additional hosting companies, data services, or subcontractors. Information entered into the vendor’s application could therefore pass through several organizations.
The AI feature may also change over time. A tool that only summarized information when the contract was signed could later gain permission to create records, send messages, or initiate actions.
A standard annual questionnaire may not capture those changes soon enough.
Vendor reviews should now examine:
- The AI models used within the product
- The information those models can access
- Whether customer data is retained
- Whether data is used to train or improve a model
- Which outside providers support the AI feature
- How model and feature changes are communicated
- What human review is required
- Whether AI activity can be logged and audited
- How the vendor tests output accuracy and security
- What happens when the AI feature fails
ICBA’s readiness guidance specifically encourages community banks to determine whether vendors already have AI features active, whether institutions will be notified before additional features are enabled, and how major providers are responding to expanding AI-driven cyber capabilities.
The Risks Extend Beyond Data Privacy
Privacy is an important concern, especially when vendors handle customer, patient, employee, legal, or financial information. Still, vendor-related AI risk covers more than data exposure.
Inaccurate information
AI-generated summaries, classifications, recommendations, and answers can be wrong. When an output influences a customer interaction, compliance task, financial decision, or operational process, an error may create consequences outside the software itself.
Unauthorized or unintended actions
Some AI tools can act within connected applications. Poorly defined permissions could allow a system to create, change, send, approve, or delete information without appropriate review.
AI agent systems deserve additional scrutiny because they may take actions that affect sensitive information, customer communications, and business operations. ICBA has called for clear accountability, constrained use in high-risk environments, strong identity controls, and auditability for these systems.
Unclear accountability
When an AI-supported process produces a harmful result, responsibility may be divided among the organization, the software vendor, the model provider, and other subcontractors. Contracts written before the feature existed may not clearly assign responsibility.
Service disruption
An organization may become dependent on an AI feature without understanding the infrastructure behind it. An outage, model change, security event, or provider dispute could interrupt a process that employees now rely on.
Regulatory and legal exposure
Regulated organizations remain responsible for how customer and business information is handled, even when the work is performed by a third party. A vendor’s use of AI can affect privacy obligations, recordkeeping, explainability, complaint handling, security oversight, and other compliance responsibilities.
Fourth-party risk
Your vendor’s AI provider may introduce its own dependencies. Those fourth parties may store information, process prompts, supply models, host infrastructure, or provide monitoring services. Your organization may have no direct contract with them, but their security and reliability can still affect you.
10 Questions to Ask Vendors About Their Use of AI
These questions can be added to vendor reviews, contract discussions, renewal meetings, and risk assessments.
- Does your product currently include any AI or machine-learning features?
- Are those features enabled automatically, or can our organization review and approve them before activation?
- What information can the AI access, process, generate, or modify?
- Is our information stored after processing? If so, where, for how long, and for what purpose?
- Is our information used to train, test, tune, or improve any AI model?
- Which model providers, hosting companies, subcontractors, or other fourth parties support the feature?
- What human review is required before AI-generated output is used or an automated action is completed?
- How do you test the feature for inaccurate output, unauthorized access, security weaknesses, and inappropriate behavior?
- Will you notify us before changing the model, adding a new provider, expanding data access, or introducing new automated capabilities?
- How will you notify us of an AI-related security incident, service failure, or harmful output, and what responsibilities will each party have?
The answers should be documented and evaluated according to the importance of the vendor, the sensitivity of the information involved, and the potential effect on customers and operations.
A low-risk productivity feature may require basic safeguards. A system that handles financial records, protected health information, legal documents, customer communications, or operational decisions deserves a more detailed review.
Connect AI Oversight to the Risk Program You Already Have
Most organizations do not need an entirely separate administrative program for every use of AI. They do need to account for AI within the processes already used to manage technology and third-party risk.
That may include updating:
- Vendor inventory and classification procedures
- Due-diligence questionnaires
- Contract and renewal requirements
- Acceptable-use policies
- Data-handling standards
- Access-control reviews
- Risk assessments
- Change-management procedures
- Incident-response plans
- Business-continuity exercises
- Cyber insurance reviews
- Management and board reporting
For example, incident-response testing should consider what happens when a vendor’s AI feature exposes information, generates harmful content, takes an unauthorized action, or becomes unavailable.
Board reporting should give leadership a clear view of where higher-risk AI is operating, which vendors are involved, what information is affected, and which risk-reduction measures remain incomplete. ICBA’s guidance places AI readiness at the board level because it affects technology, vendors, customers, security, and long-term planning.
Start With Visibility
Responsible AI oversight begins with knowing where the technology is already being used.
Create an inventory of major vendors and identify which products include AI. Prioritize the vendors that handle sensitive information, support essential business processes, interact with customers, or have the ability to take actions within your environment.
Then review the contracts, document unanswered questions, assign risk owners, and track needed changes.
You do not need to eliminate every unknown before your organization can use AI responsibly. Start by identifying where AI already exists, determining which uses carry the most risk, and adding the right questions to the vendor review process you already have.
AvTek Solutions helps regulated organizations connect emerging technology risks to practical cybersecurity and compliance processes. Through risk assessments, remediation planning, policy development, documentation review, vendor oversight, and ongoing security leadership, AvTek can help your organization evaluate AI-related exposure without adding unnecessary complexity. These capabilities align with AvTek’s Compliance-as-a-Service and virtual Chief Security Officer offerings, which focus on continuous risk management, documented due diligence, governance, and recurring leadership support.
Concerned that AI may have entered your vendor environment faster than your risk program could respond? Contact AvTek Solutions to discuss an AI governance or vendor-risk assessment.
