Governing the Agents We Build and Defending Against the Ones We Don’t: A GC Lunch Series Discussion

Six months ago, our GC lunch series asked how companies should prepare for the agentic era. This installment made clear that the era has arrived. Agents are running inside the enterprise, attackers are deploying their own, insurers are rewriting coverage rules, and legal teams on both sides of the outside counsel relationship are using agents in the practice of law. General counsel, chief legal officers, and industry leaders joined a candid conversation about where the pressure points are and what legal leaders should do about them. Here are the key takeaways. 

Governing Your Own Agents Is Harder Than It Looks 

The question is no longer whether to deploy agents, but how to govern those that are already running. Most participants recognized that creating an agent inventory is challenging (impossible?), including those embedded in vendor products and those built by enthusiastic employees without formal review. You cannot govern what you cannot see. 

Every agent needs a defined scope of authority, a named human owner, a documented list of the systems it may access, and a clear point at which it must stop and escalate. Delegate to agents with the same rigor you apply to delegating signature authority to a person. 

As agents chain together multi-step tasks, a person’s practical ability to approve every action shrinks. A more durable control is a staged workflow with defined checkpoints, clear inputs, and limited scope at each step, rather than a single instruction to an agent with broad permissions. 

Authorization and liability for agent conduct remain contested. The Amazon v. Perplexity litigation over an AI shopping agent, in which the Ninth Circuit vacated a preliminary injunction in August 2026, shows how unsettled it is whether the user or the agent is “accessing” a system. GCs should not assume that existing contract, agency, and computer access frameworks will map cleanly onto agent behavior. 

The Market and the Insurance Tower Are Pulling in Different Directions 

Commercial pressure to deploy agents quickly is outpacing the risk transfer market’s ability to price what agents might do. Boards and business leaders want speed, while underwriters want evidence. Much of the market remains silent on AI, neither expressly covering nor excluding AI-driven losses, so coverage will turn on how a claim is characterized after the fact. The bottom line – do not assume an AI-related loss is covered simply because AI is not excluded. 

Carriers are also asking pointed questions about AI controls, and some are adjusting pricing, limits, and sublimit based on the answers. A well-documented agent governance program can directly affect the cost and scope of coverage. Where insurance is uncertain, indemnities, limitation of liability provisions, consequential damages waivers, and audit and cooperation rights in agentic vendor agreements can be significant mitigators of risk. 

The Threat Landscape Is Moving at Machine Speed (and Don’t Forget to Love Your CISO) 

The external threat picture has changed significantly. Attackers now use agents to run reconnaissance, credential discovery, privilege escalation, and lateral movement in parallel. Prompt injection, tool abuse, token theft, and hijacked AI identities can turn your own agents into a pathway into the enterprise. Deepfake voice and video are now routine in social engineering. 

CISOs must do more to battle the threats from the outside. They are also challenged with securing and enabling AI across the enterprise, often without more budget or headcount. GCs should be the CISO’s closest ally, aligning in advance on incident response and disclosure playbooks, reporting lines and decision rights, and ensuring the message is briefed to the board.  

Another key tip for legal leaders – add a scenario to our tabletop exercises in which an agent, yours or an attacker’s, is the actor, and work through the legal, regulatory, insurance, and communications questions it raises. 

How Inside and Outside Counsel Are Putting Agents to Work 

Many in-house teams are under pressure from their boards and management teams to show the ROI in connection with using AI.  

In an effort to address that elephant, the conversation closed on how in-house teams and law firms are moving from experimentation to execution, using platforms such as Harvey alongside other tools to change how legal work gets done. At Foley, Harvey has been rolled out firmwide as part of a broader suite that includes Microsoft Copilot, CoCounsel, Draftwise, Relativity, and our internally developed FoleyChat. The teams seeing the most value start with a specific legal task, such as diligence, contract review, research, or discovery review, and map where an agent adds value, what sources ground it, and where human judgment is indispensable. Agents perform best on staged, multi-step work with checkpoints between stages. 

By reducing the time and cost of recurring and large-scale analysis, agentic workflows make it practical to bring legal input into business decisions earlier and more often, from more frequent contract portfolio reviews to earlier risk screening of new products and markets. 

Legal teams should hold their own agents to the standard they set for the business – a named attorney owner, defined authority, documented validation, and verification of every output against the underlying source. Courts are already drawing lines on privilege and confidentiality in AI use, which makes a secure, enterprise-grade environment essential. 

The most promising opportunities involve inside and outside counsel working from shared, well-organized source materials and agreed workflows, so that agents on both sides produce work that is consistent, verifiable, and faster to act on. 

What GCs Should Take Back to Their Teams 

• Inventory the agents operating in your organization, including those inside vendor products, and assign a human owner to each. 

• Update your AI governance program to address the unique risks and monitoring challenges agents present. 

• Before your next renewal, work with your broker and coverage counsel to map agent use cases against your cyber, technology E&O, D&O, and general liability policies and identify gaps. 

• Schedule time with your CISO to show the love! Agree on decision rights, escalation paths, and how you will brief the board on agent risk together. 

• Ask the same governance questions of your legal team’s agents, and of your outside counsel’s, that you ask of the business. 

• Be ready to answer the question that framed our last lunch and still frames this one. If you had to explain an agent’s conduct to a regulator, a court, an insurer, or your board, what documentation would you need, and how much of it do you have today? 

• Work with outside counsel to design workflows that use AI to truly promote efficiency and are designed to produce the ROI from AI that management teams are seeking. 

Agentic AI has moved from concept to operating reality faster than most governance programs, insurance markets, and security teams can absorb. The GCs who define authority, demand auditability, understand where their risk transfer actually sits, stand shoulder to shoulder with their CISOs, and put agents to work thoughtfully in their own departments will shape this transition rather than react to it. 

Onwards!