The Rise of AI Agents: The Next Big Thing in eDiscovery

Artificial intelligence (AI) is reshaping the way legal professionals are approaching processes across the board. From analysing vast datasets during investigations to performing document reviews in disclosure, these tools have demonstrated speed and efficiency at scale, saving firms significant time and costs while increasing the accuracy of work product. 

Now, the evolution has entered a new phase, agentic AI. While traditional machine learning or GenAI models perform single, isolated tasks, AI agents can act autonomously across multiple systems and platforms, retrieving data and executing workflows. This represents a significant leap ahead, bringing new efficiencies to the already transformative results shown by GenAI in eDiscovery. 

Agentic AI: What makes it different? 

AI tools have gained popularity in the legal world in recent years, driven by an increasing number of use cases and demonstrable results. While traditional AI excelled at analysing data and uncovering patterns that would be difficult to detect manually, GenAI took it further by enabling practitioners to rapidly generate insights and narratives that accelerate dispute resolution, investigative analysis, and eDiscovery document review. 

GenAI’s capabilities have been harnessed to great effect in disputes and investigations, particularly in eDiscovery. For example, Alvarez & Marsal has deployed a suite of GenAI-based tools in dozens of cases, resulting in better quality outcomes for clients at higher speed and lower costs. 

Agentic AI can build on these capabilities further by introducing an autonomous element that makes it proactive rather than reactive. It uses large language models, machine learning and natural language processing, to independently perform tasks on behalf of a user or another system, without requiring much manual coordination. Put simply, agentic AI focuses on dynamic decision-making rather than content creation. It can reason, plan and act autonomously to achieve complex goals without constant human oversight or prompts. 

For example, if a company needs to investigate all data on a particular employee, current AI applications require human intervention and coordination to retrieve data from different departments such as HR and IT. By empowering an AI agent to seamlessly connect with all relevant platforms and systems, these steps can be eliminated to allow the AI agent to perform the task from end to end, resulting in speedier, more efficient results. 

Where AI agents can be deployed 

While it is early days for agentic AI in the legal sphere, we are starting to see some interesting use cases emerge. Recently, a corporation in need of multiple-language capabilities for investigations decided to build an AI agent rather than hire a junior investigator. The LLM technology the AI agent is developed with is language agnostic, allowing the multinational company to strengthen its internal investigation capabilities globally.  

This AI agent is being developed to complete tasks such as categorising the seriousness of the investigation, gathering data across multiple systems and performing a high-level review of the data in relation to the allegation. The AI agent’s work product is a short summary report delivered to a senior investigator to then action accordingly.    

Agentic AI has immense potential to aid the crucial process of eDiscovery, as it can not only handle vast amounts of complex digital evidence and metadata, but – deployed correctly – can integrate several steps of the eDiscovery procedure such as identification, collection, processing, review and analysis. 

Here are some specific legal areas in which agentic AI can be leveraged for maximum impact: 

• System liaison: In disputes and investigations, AI agents can act as the go-between for various IT and data systems in place within an organisation, enabling better connectivity between systems. For example, consider a banking trader surveillance platform and the bank’s investigation platform. A flag raised by the surveillance platform could trigger an agent which moves some data into the investigation platform. Then, the agent can run an initial triage investigation. 

• Language skills: Agentic AI can also prove particularly useful for multinational companies coordinating an investigation spanning multiple geographical locations, often leading to digital evidence in several languages. By building an AI agent with language fluency and translation capabilities, companies will not have to rely on humans with language skills, resulting in lower costs and a leaner investigation team in which a small number of senior investigators are supplemented by AI agents.  

• DSARs and other repetitive processes:  Data subject access requests (in which an individual formally requests an organisation to provide a copy of their personal data and details on how it is being used) are on the rise. Under EU laws such as GDPR, businesses have 30 days to respond to a DSAR request, so most organisations have already developed a methodical approach to answering them. In such instances, where repetitive tasks with clear methodology are involved, an AI agent can perform a significant part of the process with limited human oversight, from collecting data to processing, identifying personally identifiable information (PII) and redacting necessary information. Senior personnel at the appropriate level can then sign off on the process.  

• Whistleblower hotline: Agentic AI can also be useful in the early stages of whistleblower hotline investigations. When the hotline rings, the AI agent can pick up the allegation and categorise it according to risk level. Further, for concerns categorised as low level, the agent can collect and analyse data, before reporting it to senior management, or flag it up the chain for higher-level categorisations. 

Risks and considerations 

As with any new technology, it is important to bear in mind certain risks and considerations before deploying AI agents, particularly given sensitivities around compliance and data privacy in the legal world. While the considerations are similar to those around GenAI models, the autonomous aspect of agentic AI introduces an added element of scrutiny. 

• Accountability: While AI agents can act autonomously, it is vital to identify and obtain sign-offs from the employee at the appropriate management level for the decisions taken by the agent 

• Data privacy and security: Access and linkage to multiple systems and platforms raises the scale and scope of the risk of data compromises, requiring increased guardrails for data protection 

• Regulatory risks: Institutions must ensure that AI deployment aligns with regulatory requirements. 

• Tech implementation: While AI agents do not require training in the traditional sense, it is more about building and configuring agents for the specific job they must perform. This calls for skilled tech personnel, investment in tech and governance frameworks to manage AI-driven processes. 

Agentic AI represents a paradigm shift, with the potential to dramatically reduce time and costs while improving the quality of output in legal processes. To leverage it effectively, firms must correctly identify high-impact use cases, develop appropriate AI governance frameworks and invest in technical capabilities. 

We are starting to see the use of Agentic AI to help corporations with their challenge of managing a higher volume of internal investigations, as well as the risk of the current dispute resolution climate. These systems can orchestrate complex, multistep tasks—triaging issues, surfacing key evidence, and accelerating earlycase assessments—without demanding equivalent increases in headcount or budget. Organisations that embrace this shift will be able to maintain control as investigative workloads continue to climb. Within the next few years, we expect agentic AI to become a standard component of investigation and eDiscovery workflows, and for regulators to begin assuming that companies have deployed these capabilities as part of a defensible, modern compliance posture.