Where agentic AI is gaining ground in legal practice – and where it stops

I had the honour of joining this panel, and David Swan, technology editor at The Age and The Sydney Morning Herald, moderated. Fellow panellists were Brandon Hollinder, an 18-year e-discovery veteran and attorney at Epiq; Michael Legg, professor at UNSW Sydney and a 2026 Relativity AI Visionary for his research into the future of the legal profession; and Gina O’Neill, associate director of e-discovery at Ashurst Advanced. Between us, we covered quite a bit of ground.

What agentic AI actually means

A key focus of the discussion was distinguishing chatbots from the genuinely different capabilities of agentic AI – because conflating the two can lead to complacency or unnecessary fear. Unlike a chatbot that responds to queries one at a time, an agentic system can interpret an objective, break it into steps, select tools and APIs, take actions, and adjust its approach if something fails. It perceives, thinks, acts, and iterates until it reaches the goal.

Hollinder illustrated the concept vividly through two analogies. In an Amazon warehouse, robots now increasingly perform tasks autonomously that human workers once did by hand: receiving shipments, classifying items and placing them in the right location. He also pointed to self-driving cars, which, once given a destination, continuously read their environment, decide how to respond to pedestrians and traffic signals, and act without waiting for instructions. The perceive-think-act loop, he argued, is the essence of agentic AI regardless of the underlying technology.

Swan added an important distinction for legal teams: an AI workflow automates a known, predefined process, while a true AI agent dynamically directs its own process toward a goal. Understanding which one an organisation is building or buying is the first governance question.

A spectrum, not a switch

VIEW ALL

The discussion also pushed back on the idea that agentic AI is binary – either the human is in control, or the machine is. In practice, the relationship is far more nuanced, with different levels of autonomy requiring different levels of oversight.

As I noted during the session, it is critical for users to guide, validate, and maintain control over AI-driven decisions because human intervention is what makes legal AI ethical, practical, and defensible. Fully autonomous agents operate entirely on their own, without any human review, and do not belong in legal technology.

Where agentic AI is already delivering

Hollinder provided one of the clearest pictures of agentic AI in production. He described how, at Epiq, agents are already operating across two tracks. Internally, they handle HR queries, classify and route client requests to the appropriate delivery team, and monitor incoming email for urgency. That last capability alone has made a measurable difference. “We have AI essentially reading our emails and escalating things to the top so that we can respond to those clients in those really hot situations up to 80 per cent faster,” he told the room. The second track is client-facing, with agents built on RelativityOne supporting end-to-end workflows including data ingestion, OCR, imaging, indexing, and quality control.

O’Neill offered a different example from Ashurst Advanced. When clients enter a new jurisdiction or prepare to launch a new product, they often face hundreds of pieces of legislation, each hundreds of pages long. Rather than overwhelming a single model with all of it, Ashurst built a two-agent system: a tracker agent that monitors what has been reviewed and what comes next, and an analysis agent that works through each section as directed by the tracker agent.

The tracker agent assigns sections to the analysis agent, which processes each one and reports back when complete. The tracker then confirms completion and advances the workflow. Human review is built in between passes, ensuring outputs are validated before the system proceeds and reducing the risk of compounding errors.

Start now, start small

For practitioners not yet using agents, Hollinder framed the opportunity plainly: “Think about, what are your pain points? What are the administrative things you have to do? There’s probably an agent that can do that. But it’s an agent that can help you, not necessarily replace you.”

From my perspective, an effective starting point is simply to observe and build familiarity. There is now a wide range of accessible material, like YouTube videos, that can teach you about agents. I’ve learnt to use generative and agentic AI in my personal life, and those habits have translated into how I approach it professionally.

Governance: The discipline you already have

Governance around agentic AI may sound like new territory, but O’Neill reframed it as something legal technology practitioners are already familiar with. For example, the discipline built around mass-update permissions in RelativityOne – requiring senior sign-off, defined search conditions, and a clear rollback position before touching millions of records – offers a useful model for governing agents.

“Any tool you select needs to bring the same level of discipline that you have now,” O’Neill said. Give each agent a defined role and scope, build in traceability, set runtime guardrails, and identify the checkpoints where human judgement must be exercised.

Legg illustrated the question of autonomy through a billing scenario. An agent that assembles time sheets, rates, and fee agreements to draft a bill is clearly useful. But an agent that then sends the bill, fields client queries, chases payment, and commences legal proceedings crosses a more contested line. Working out exactly where that boundary sits is the governance challenge for each use case.

Ethics and the case for adoption

Legg made the case for adoption on ethical grounds, not just commercial ones. Lawyers are now understood to have a competence obligation that includes a working understanding of how relevant technologies operate.

“This isn’t just like a business thing for [lawyers]. It actually goes to your ethical obligations. You need to understand how the technology works,” he told the audience. He also noted that recent case law suggests that using AI for large-scale document review is not merely permitted but, in some circumstances, may be expected in order to act in a client’s best interests – raising the possibility that failing to use such tools could, in certain contexts, fall below the required standard of care.

For firm and other organisational leaders, Legg highlighted the shadow AI risk: if you do not create structured opportunities to experiment with agentic AI, people will inevitably use it anyway, outside any governance framework. The answer is not to close the door, but to open it deliberately, with the right guardrails in place.

The legal community is ready

The session closed with a clear line drawn between what the legal profession is ready for and what it is not. The picture, I believe, is more optimistic than many expect.

The legal community is ready for – and is already adopting – supervised generative AI and constrained agentic workflows. At the same time, a clear boundary remains around systems that act, plan, and make decisions without meaningful human review – approaches that do not belong in legal practice.

Relativity’s agentic research tools show what purposeful design looks like: an agent generates a research plan, presents it to the user for review and editing, and proceeds only once the human has approved it. Every significant step requires a human to exercise their judgement. The technology supports and acts on that judgement. It does not replace it.

Agentic AI is not a future disruption to prepare for, but a present capability to deploy thoughtfully – with governance structures, human checkpoints, and purpose-built legal tools. The technology is ready, and so, in many respects, is the legal profession. Watch the full RelFest Sydney 2026 closing keynote here.