Govern cyber and risk with confidence as threats emerge
The increased autonomy and access to data granted to AI agents in recent months had board members eager for a dialogue about governance best practices during Grant Thornton’s National Association of Corporate Directors Summit presentation near Washington, D.C.
“When an agent makes a decision, who is keeping track? And is there someone who’s in charge? It’s going to be important to keep track of who owns the decisions that agents make,” said one board member during a lively session on AI governance where the seats in the conference room were full.
AI’s effect on cybersecurity and risk management is one of the top issues on every board’s agenda today, and directors are straining to provide helpful oversight over a technology that’s changing so rapidly they can hardly keep up.
Fifty-seven percent of the 1,100 business executives polled in Grant Thornton’s recently released Cyber & Risk Survey said AI governance and use of AI was a top source of future risk or regulatory scrutiny for their organization.
An audience poll at Grant Thornton’s NACD presentation showed that board members’ top concern (voiced by 41%) about AI oversight was whether there are controls and oversight in place with clear ownership and decision rights. An additional 19% were most concerned about whether emerging AI risks were being identified.
Board members are pressing management for answers. They’re asking who is keeping track of agentic AI decisions and whether audit trails are being established for AI tools.
“Management has to understand what they are outsourcing at the end of the day to these third parties, to these agents, and understand what their responsibilities are,” said Grant Thornton Technology M&A Partner Tarun Kakar. “Outsourcing decision rights are more than just a third-party risk. They can create concentration and supply chain risks that may threaten business continuity, disrupt critical operations and put revenue at risk.”
There’s a lot to understand about AI as it relates to cybersecurity and risk management, and Grant Thornton professionals provided important insights on a variety of those topics at the NACD Summit.
AI risk management is not one-size-fits-all
Boards should ask management how it sorts AI use cases by risk. Use cases with a high chance of failure and high potential for damage need stricter controls than low-impact tools, and many companies already set their controls and budgets on that basis. The problem comes when a use case changes after it has been classified.
“The biggest risk is when a use case moves from one risk category to another without you noticing,” said Grant Thornton AI & Data Partner Sumeet Mahajan.
Mahajan said two factors should set a use case’s risk category: how much the AI can do without a person approving it, and what data and systems it can access. “When either one changes, the controls have to change with it,” he said.
For example, an AI agent in procurement might start with limited governance because it only researches public information about suppliers and compares their capabilities. Later, the same agent may be given access to supplier contracts, negotiated prices, purchasing history, performance records and risk assessments, and begin recommending suppliers. At that point it needs stronger controls. If the controls stay at the original level, the company has gaps in data protection, and the agent may act without the human review its work now requires.
These changes usually come through a series of small approvals, and none of them are checked against the original risk classification.
Boards can ask management who owns the risk classification for each AI use case, what triggers a review of that classification, and whether controls were strengthened for agents that have gained new data access or decision authority since launch.
How we can help you
SERVICE
Fight back against shadow AI
Johnny Lee, National Practice Leader of Grant Thornton’s Forensic Technology practice, recently discussed with the general counsel of a large public company what keeps him up at night.
It took just a millisecond for the executive to respond: “Our employees’ use of GenAI and its corollary, shadow AI.”
Both sanctioned and unsanctioned usage of GenAI tools is creating huge exposures and concerns for companies, and this necessitates that boards ask some tough questions of management related to the same. In some cases, employees are using public AI tools to ingest and process proprietary and/or sensitive data that can cause irreparable harm.
Even when this usage is restricted to approved GenAI tools, there may still be issues that arise (related to company record retention and destruction practices). Boards need to ask management how these retention policies and practices intersect with approved AI use — and, more specifically, whether thoughtful attention has been paid to preserving (or, as the case may be, periodically destroying) GenAI prompts and outputs in a manner contemplated by the organization’s records-management protocols.
“Those troves of data created by an employee’s interactions with GenAI are being stored somewhere,” Lee said. “Are those data troves being retained according to policy? Does the current policy even contemplate this category of content? If pressed, for instance in litigation, could the organization even isolate such information? If so, could the organization suspend automatic deletion for a litigation hold? The case law is replete with examples of the judiciary reacting poorly to relevant data being destroyed by a party to litigation — and ‘it’s complicated’ has never been a compelling defense.”
With unapproved AI use, the problems can be even more serious. Grant Thornton Transformation Partner Jennifer Morelli was recently working with a large company when a discussion about the use of AI tools raised concerns about whether employees were handling sensitive, proprietary information only within company-approved AI environments.
The CEO asked employees in the room to raise their hands if they were using the company-approved, protected AI environment.
Not every hand went up.
“Some people were going on their mobile devices or using public AI tools and putting strategic company documents into them. Once sensitive information is entered into an unapproved AI tool, the organization may lose control over how that information is handled,” Morelli said.
Morelli suggested that questions boards should be asking include:
- How many unapproved AI tools are we using?
- How many exceptions to our AI policies have we permitted?
- Is there a risk of proprietary data appearing in public domains?
If either of the first two answers is greater than zero, the risk in the third question is real. And that requires a new question: “What actions is management going to take to eliminate employees’ use of unapproved AI tools?”
Assessing governance
The most important thing that directors can do to improve AI cybersecurity and risk management may be to ask management whether its AI governance would stand up to an independent assessment.
Grant Thornton’s 2026 AI Impact Survey asked business leaders whether their organization could pass an independent AI governance audit in 90 days. Seventy-eight percent were unwilling to say they had strong confidence that they could pass such an audit.
“The notion that nearly eight out of ten leaders believe there would not be a good result when an arm’s-length, uninvested person kicked the tires on your AI controls is the key takeaway from that,” Lee said.
Organizations that don’t have an independent assessment of their AI governance controls on the calendar should operate as if such an assessment were coming. They should document their governance processes, establish audit trails for their AI tools and retain evidence of their control activities for later testing.
Even if this evidence is never requested for an audit, it should be kept immediately at the ready for the day when a breach does occur, and the organization has an urgent need to figure out what happened and how to minimize the damage.
Because of this, asking management if they have the evidence to support such an audit might be one of the most important insights that a board member could gain about the GenAI state of the union at that organization.
3 Summit takeaways
Board members came to Grant Thornton’s NACD Summit sessions looking for answers to difficult questions on how their oversight can help organizations manage risks related to AI.
Helpful answers emerging from the dialogue included:
Penetration testing, AI-style: One audience member was looking for advice on how to identify the AI risk management equivalent of the “ethical hackers” that companies hire to perform penetration testing on their cybersecurity protections and controls.
Lee suggested that most organizations are likely to have people on staff who can perform some of those duties. He said board members should ask management to identify high-risk AI use cases and then ask management to articulate — specifically — why these use cases are high-risk. Are there data concerns? Security concerns? Regulatory concerns? A combination of these things?
Determining the source of the risk helps identify the internal resource who may be able to help address the risk (or to at least assist the organization toward a more audit-ready solution). For legal or regulatory concerns, for example, Lee said the general counsel and/or compliance officer may have a strong understanding of how a GenAI use case may create such exposures. Pairing such a resource with the AI use case author or the accountable executive can lead to the specific controls that help manage the risk.
Taking tabletops to a new level: Just 35% of respondents to Grant Thornton’s Cyber & Risk Survey had completed a full cyber breach simulation in the past six months.
While that’s concerning, Lee also said he has seen some of his clients take their tabletop exercises to a new level of sophistication, typically by designing a scenario at the intersection of cybersecurity and AI.
He said organizations that regularly exercise their incident response protocols don’t need to reinvent their crisis-response just for AI. GenAI incident management can be bolted on to existing cyber-response practices.
“The point of such exercises is not to improve your ability to conjure up hypotheticals; it’s to develop muscle memory that you can leverage during a bona fide crisis,” Lee said. “As your crisis response capability improves, the organization is identifying intersectionality where certain enterprise phenomena might overlap, creating a crisis. In such cases, the organization is invoking different command centers to show that they can identify core issues, activate right-sized and risk-aware responses, and unify with tangential internal groups during a response. As but one illustration of this, consider a threat actor that compromises your enterprise security then targets GenAI prompts from your R&D team or your M&A team. Imagine the kind of sensitive data within those GenAI prompts and outputs.”
Insurance is the next frontier: Insurance for AI-related risks is in its infancy now, much as cybersecurity insurance was 10 to 15 years ago.
For all forms of risks, organizations need to invest in preventing risks wherever possible, quickly detecting risks that can’t be prevented, and insuring for any residual risk beyond that.
In cybersecurity, the need for insurance coverage and the development of consistent insurance products helped establish a “mandatory minimum” standard for protections that organizations needed to adopt to be underwritten as policyholders. Lee expects AI defenses to follow a similar path, with the underwriting requirements (over time) leveling the playing field by setting strong minimum standards.
Despite the nascent state of insurance for AI-related threats, this is a topic board members should begin to explore with their management teams now and to revisit periodically as the offerings mature.
“The brokers and underwriters — like most of us — are certainly struggling to properly risk rate and underwrite these enterprise risks, but that was likewise the case 10 to 15 years ago in cybersecurity,” Lee said. “There’s a healthy appetite for such products and a relatively untapped insurance market, so I expect these offerings to become an important part of the risk dialogue in very short order.”
Contacts:
Partner, Technology M&A
Grant Thornton Advisors LLC
TK Kakar is the Transactions Technology Leader at Grant Thornton.
Partner, Forensic Advisory Services
Grant Thornton Advisors LLC
Johnny Lee is a Partner in the Forensic Advisory practice and the National Practice Leader of the Forensic Technology practice.
Partner, AI & Data
Grant Thornton Advisors LLC
Sumeet enables clients to unlock measurable value with AI and data by linking business strategy to AI strategy and guiding build/buy/partner decisions.
Chicago, Illinois
Industries
- Manufacturing
- Retail & Consumer Brands
Service Experience
- Artificial intelligence
- Business Consulting
- Technology Modernization
Partner, Transformation
Grant Thornton Advisors LLC
Jen leads Grant Thornton’s People & Culture practice, guiding orgs through change via process, tech & people transformation.
Philadelphia, Pennsylvania
Industries
- Healthcare
- Manufacturing
Service Experience
- Advisory Services
- Business Consulting
Content disclaimer
This Grant Thornton LLP content provides information and comments on current issues and developments. It is not a comprehensive analysis of the subject matter covered. It is not, and should not be construed as, accounting, legal, tax, or professional advice provided by Grant Thornton LLP. All relevant facts and circumstances, including the pertinent authoritative literature, need to be considered to arrive at conclusions that comply with matters addressed in this content.
For additional information on topics covered in this content, contact a Grant Thornton LLP professional.
Ready to talk? We’re ready to listen.
Request a meeting and a member of our team will be in touch to see what we can do to meet your needs.
Want to submit an RFP? Please submit your request through our RFP submission page.
Share with your network
Share