Voya Financial to leverage unqorkai to accelerate ai powered experiences

Read the Press Release

Robert Landon

Beyond the Hype: 12 Questions Every Corporate Board Should Ask About AI Risk

A former CIO’s framework for governing enterprise AI, retiring legacy liabilities, and ensuring measurable business value

By Gary Hoberman, Founder & CEO, Unqork

Board directors today are inundated with generic AI ethics and compliance checklists from law firms and audit committees. Those reviews are necessary, but they rarely uncover the architectural reality underneath—the hidden technical debt, fragile software sprawl, and operational blind spots that management teams seldom volunteer in board packets until they hit the balance sheet.

Artificial intelligence can help an enterprise move faster, serve customers better, and operate more efficiently. Yet speed without control does not eliminate risk—it compounds it. Directors do not need to understand neural networks or inspect source code, but they must ensure that AI-driven gains do not introduce unmanageable financial liabilities or regulatory exposure. 

As a former Fortune 50 CIO, I believe effective board governance of AI comes down to asking twelve targeted operational and strategic questions across four core areas.

“Speed without control does not eliminate risk—it compounds it.”

Phase 1: Strategy, Value, and Economics

Separating real ROI from pilot vanity metrics, hidden maintenance debt, and legacy retirement. 

1. What measurable business value is AI delivering—and what new obligations are we taking on? 

Ask management to demonstrate tangible revenue growth, superior customer experiences, lower operating costs, or faster time-to-market. Pilots launched, agents deployed, and lines of code generated measure activity, not outcomes. The board must know what actually improved for the business—and whether those gains justify the ongoing financial cost, operational overhead, and risk exposure.

2. What is the full lifecycle cost—and how much future maintenance are we creating? 

Look beyond initial pilot budgets and token generation costs. True total cost of ownership includes infrastructure, integration, testing, security, ongoing support, and eventual retirement. Every new codebase or custom agent creates a permanent obligation to understand, secure, and update it. Ask management what happens to total costs at ten times today’s adoption—and whether near-term productivity gains are reducing or quietly inflating the company’s long-term expense base.

3. Are we simplifying the enterprise—or repeatedly rebuilding the same capabilities? When different teams need the same customer data, business rule, approval workflow, or system integration, do they reuse a governed enterprise capability or spin up another isolated version? 

If underlying logic cannot be updated consistently across the company, AI will accelerate architectural fragmentation. Ask which duplicate tools and manual processes each major AI investment will permanently consolidate.

4. What are we doing to actively retire high-risk legacy applications and monitor the increasing cyberattack surface area created by AI-generated code? 

Every legacy application and third-party vendor represents an increased surface area you have to maintain and secure against cyberattacks. Frontier cyber-reasoning models have demonstrated that AI can autonomously uncover decades-old zero-day flaws across aging codebases and third-party dependencies in hours, while continuous AI vulnerability scanning only generates massive triage backlogs that outpace human patching capacity. Ask management which legacy systems and third-party tools are being retired and brought in-house onto a modern, governed architecture—and what hard cost and attack-surface reductions will result.

Phase 2: Legal, Compliance, and Brand Risk

Preventing Shadow AI IP leaks, regulatory exposure and unintended autonomous actions. 

5. How are we insulating the company from regulatory liability, IP infringement, and employee “Shadow AI” data leaks? 

AI introduces complex legal exposure across intellectual property, data privacy, algorithmic bias, and rapidly evolving global regulations—while everyday employee usage creates an equally urgent leakage risk. If an AI system trains on unlicensed material, makes discriminatory decisions, or receives trade secrets, customer records, or proprietary source code pasted by staff into unapproved consumer AI tools, the company bears the liability. Ask management how initiatives are mapped to regulatory requirements, how models are audited for fairness and IP provenance, and how technical controls prevent staff from donating corporate IP to external models.

6. Testing shows that AI built the function we requested. What guarantees it built nothing more?

Software can be technically sound and free of known security vulnerabilities while still executing an unauthorized business function. Instructing an autonomous agent to “maximize sales” or “resolve customer churn” does not authorize every conceivable method of achieving that goal. Ask how management identifies, constrains, and prevents behavior that nobody explicitly requested or approved—and what the worst-case financial or reputational blast radius looks like.

Phase 3: Operational Control and Security

Enforcing system-level guardrails, SDLC discipline, kill switches, and model portability. 

7. Are AI agents constrained by enforceable permissions and approval limits, just as employees are? 

Every autonomous agent should operate with an identifiable role, least-privilege access to authorized data, and strict boundaries on permitted actions. Which decisions—particularly financial commitments, customer-facing changes, and production system updates—require explicit human sign-off? Ask management to prove that these guardrails are enforced externally by system architecture rather than merely suggested inside the agent’s prompt, and that permissions can be revoked instantaneously.

8. Are AI systems allowed to bypass the security and testing guardrails required for traditional software? 

Traditional enterprise software undergoes rigorous requirements validation, security review, testing, and controlled deployment before reaching users. What prevents an AI agent from circumventing those controls by dynamically generating and executing code, scripts, or tools inside a live environment? Production must never become a loophole around enterprise governance.

9. Can we immediately stop harmful activity and restore operations? 

Ask management to demonstrate—not just describe—how the organization would halt a rogue agent, revoke credentials, roll back unintended changes, and restore normal service. Which automated actions create consequences that cannot be reversed? Resilience plans must be tested before the business becomes dependent on the capability.

Phase 4: Architecture, Delivery, and Accountability

Maintaining auditability, code-free business verification, and deterministic rule boundaries. 

10. Do we know everywhere AI operates—and can we reconstruct its decisions? 

Require a unified inventory spanning employee productivity tools, vendor-embedded AI features, AI-generated applications, and autonomous agents. Every material deployment must have a defined business purpose and a named accountable executive. When an incident or audit occurs, can the company reconstruct the exact data, model version, instructions, actions, and human approvals behind a specific outcome?

11. Are we improving the full journey from business need to production—and can executives verify it without reading code? 

Accelerating code generation creates zero business value if work simply stalls in requirements, integration, security review, or testing bottlenecks. Measure end-to-end delivery speed alongside risk control. Just as importantly, ask whether responsible business leaders can directly validate that live production behavior matches approved business requirements—without having to read underlying source code—as applications and models evolve.

12. Are we using probabilistic AI judgment where deterministic business rules would be safer and cheaper? 

Many enterprise processes rely on clearly defined inputs, regulatory formulas, and fixed approval thresholds. Asking a probabilistic AI model to repeatedly interpret rules that should execute predictably introduces unnecessary hallucination risk and inflates compute costs. Ask management how they separate tasks requiring genuine AI judgment from calculations, permissions, and compliance steps that belong in explicit, auditable business rules.

Bringing This Framework into Your Next Board Meeting 

As a former CIO, I’ve seen firsthand how quickly architectural shortcuts become board-level liabilities. The goal of AI governance isn’t to slow down innovation—it’s to ensure your enterprise architecture can survive the speed.

Stay tuned for our upcoming LinkedIn Live session designed exclusively for Board of Directors members across financial services, health, insurance and other regulated industries to discuss how peers are operationalizing these 12 questions.

Experience the Unqork platform up close

You Might Also Be Interested In