The legal industry is steadily moving beyond early experimentation with generative AI and into a more complex phase defined by agentic systems. These systems do not simply generate text or assist with discrete tasks. They are designed to take action, to carry out multi-step processes, and to operate with a level of autonomy that begins to resemble participation rather than assistance.
This shift is already creating consequences in practice. Courts are beginning to confront questions about how AI-generated materials should be treated, how protective orders should apply, and how far restrictions should go. At the same time, law firms and clients are facing a different but equally significant challenge, which is how to implement these tools in a way that is workable, defensible, and economically sustainable.
The conversation around AI is no longer theoretical. Judges are now being asked to decide how existing doctrines apply in situations involving AI tools.
Recent cases illustrate how unsettled the landscape remains. In one matter, a court determined that documents generated by a represented criminal defendant using an AI system did not qualify for work product protection (United States v. Heppner). In another case, a pro se litigant using AI to prepare for litigation was able to retain that protection (Warner v. Gilbarco, Inc.). These outcomes suggest that context matters greatly and that courts may take into account the role of counsel when evaluating AI-assisted work.
Protective orders are also evolving in uneven ways. Some courts have allowed parties to expand restrictions beyond confidential materials to cover all discovery, while others have limited restrictions to sensitive information or required the use of non-public systems. In certain instances, disclosure of the specific AI tools used has been required, raising additional questions about whether such disclosure reveals strategy or mental impressions.
Taken together, these developments point to a system that is still feeling its way forward. There is no consistent approach yet, and that lack of consistency introduces both risk and uncertainty.
As firms and courts attempt to impose guardrails, a broader tension is beginning to surface. Restrictions that seem reasonable in one context may have unintended consequences in another.
For example, prohibiting the use of public AI tools is often viewed as a straightforward way to manage confidentiality risk. At the same time, such restrictions may disproportionately affect individuals who do not have access to enterprise systems. For pro se litigants, public tools may be the only viable option for engaging with the legal process in a meaningful way.
This raises questions about fairness and access to justice. It also suggests that solutions may need to extend beyond simple prohibitions. One possible direction is the development of controlled environments that provide broader access while still protecting sensitive information. The view that “if the product is free, you are the product” drives the sense that confidentiality is at risk in free models, however vendors in this space could benefit from the publicity of providing “closed models” to pro bono litigants.
While legal doctrine continues to develop, the operational challenge within firms is immediate and practical. Many organizations are being asked to adopt new technologies on top of already strained infrastructure.
Even large firms with significant IT resources struggle to keep existing systems current. Introducing AI adds layers of complexity that include policy development, monitoring, compliance, and integration across workflows. These demands create pressure not only on technology teams but also on lawyers who must understand and apply new rules in their day-to-day work.
At the same time, there is often no clear answer to the question of cost. Building, deploying, and maintaining AI systems requires investment, and the return on that investment is not always easy to measure. This creates hesitation, even as clients and competitors push for increased adoption.
Policies, controls, and audit processes are necessary, but they do not resolve the central issue, which is how to implement AI in a way that is both effective and financially viable.
One of the more important insights emerging from this moment is that no single participant controls the outcome. Legal work involves a network of actors that includes clients, courts, regulators, opposing counsel, and service providers. Each of these participants is beginning to incorporate AI into their own processes, often in different ways and at different speeds.
This interconnected environment makes it difficult to rely on firm-specific solutions. A policy that works internally may not align with the expectations of a court or the practices of an opposing party. As a result, there is growing recognition that broader frameworks are needed.
Organizations such as Sedona may be well positioned to help develop shared principles that apply across this broader ecosystem. Such frameworks would need to address not only law firms but also judges, regulators, and other stakeholders who play a role in the legal process.
Against this backdrop, Agentic AI introduces a new level of complexity. These systems are designed to execute workflows rather than simply assist with tasks. They can coordinate across systems, make decisions based on evolving information, and carry out sequences of actions without constant human direction.
The potential benefits are significant. Workflows that currently require multiple tools and participants could be streamlined. Processes could become faster and more consistent. Large-scale matters could be managed with greater efficiency.
At the same time, the risks expand. Autonomy introduces unpredictability. Errors in one step can propagate through an entire workflow. Decisions may be made in ways that are difficult to anticipate or fully understand. Oversight becomes more challenging as the speed and complexity of activity increase.
Security concerns also grow. Systems that interact across multiple platforms and datasets create more opportunities for exposure or misuse. Early research suggests that agentic systems may be more vulnerable to certain types of attacks, further underscoring the need for strong governance.
A consistent theme across these developments is that governance is lagging behind technology. Standards are still in development, and many organizations are working without clear guidance.
This creates a gap that is difficult to manage. Effective governance depends on the ability to measure, monitor, and control system behavior. In many cases, those capabilities are still emerging. Without them, it becomes difficult to assess risk or ensure compliance.
There is also a broader question about what governance should look like in an environment where systems are increasingly autonomous. Traditional models that rely on human oversight may not scale effectively, and new approaches will likely be required.
Even if governance challenges are addressed, another issue remains. Large technology initiatives have historically struggled to deliver expected results. Many fail outright, and others fall short of their intended goals.
AI initiatives may face similar obstacles. They require substantial investment, ongoing maintenance, and careful integration into existing workflows. In some cases, systems may be deployed despite not fully meeting requirements, simply because of the resources already committed.
This dynamic creates risk for both firms and their clients. Without a clear connection between use cases, costs, and outcomes, adoption may become difficult to justify.
What emerges from all of this is not a single problem but a set of interconnected challenges. Questions about confidentiality, work product, and protective orders intersect with issues of cost, governance, and implementation. Agentic AI amplifies both the potential benefits and the risks.
The path forward will likely require a more deliberate approach. Technology decisions will need to be closely tied to specific use cases. Governance will need to evolve in parallel with capability. And solutions will need to account for the broader ecosystem in which legal work takes place.
The promise of AI in law remains significant. It has the potential to reshape workflows, improve efficiency, and expand access. Realizing that potential, however, will depend on the ability of the legal system to integrate these tools thoughtfully and sustainably.
In the end, the question is not simply whether the technology works. It is whether the legal profession can adapt its structures, expectations, and practices in a way that allows the technology to be used effectively and responsibly.
*This blog post and analysis by Martin Mayne, Avalon's VP of eDiscovery, is based primarily on an article presented in The National Law Review called “The Next Generation of AI: Here Come the Agents!” By Tara S. Emory and Maura R. Grossman, J.D., Ph.D. on Monday December 30, 2024, and subsequent discussion of related issues in the Working Group 13 meeting of the Sedona Conference held in Austin, TX, in April of 2026.