Andy McDonnell, of Lex Mundi, joined the AI Education Forum to discuss what it takes for AI agents to work in legal practice. Having worked at three large law firms, McDonnell spoke from a practitioner’s perspective on workflow design, accountability, client relationships, and the economics of adoption.

Overview

McDonnell was cautious about calling 2026 the year of the agent. Agents are not a silver bullet, and many firms that rushed towards adoption found they lacked the foundations to make them work. An agent with access to poor data will produce poor output. In his view, firms are now going back to do the groundwork, which he considers healthy in a profession that cannot move fast and break things. Clients pay for discretion, diligence and quality, and firms are right to be wary of agents acting without their consequences being understood.

His practical starting point was not data cleansing but the firm’s own work. Break down the typical workload by practice group, identify what is profitable and where AI could make it more so, and only then look at the data that supports it. This means mapping matters, deciding where human judgement ends and AI execution begins, and applying knowledge management and legal project management discipline. Firms are already automating RFP responses, due diligence and document review, but McDonnell said the handover point between agent and lawyer is still unsettled even there. On accountability, he rejected the idea that senior lawyers are unfairly exposed. AI output is a starting point, and interpreting it is precisely their job.

The practical implications reach into governance, people and pricing. Knowledge management and innovation teams should work closely with practising lawyers, who know what good looks like. McDonnell warned against leaving juniors to experiment alone, since they know how to prompt but lack the context to judge the result. He also argued that the experimentation phase is over, and that firms too busy to engage should make it someone’s job. On pricing, he raised the unresolved question of who benefits from efficiency gains. He also predicted that consumption-based pricing could make some workflows uneconomic, because token costs are currently subsidised by the frontier model providers.

Key Takeaways

• Start with the work, not the data. Map your workload by practice group and identify what is profitable before deciding which agent workflows, and which data, matter.
• Define the human-AI boundary explicitly. Firms still struggle to say where an agent stops and a lawyer takes over, so set this out for each use case.
• Keep senior lawyers at the centre. Juniors may be more fluent with AI, but experienced lawyers know what good output looks like and remain accountable for it.
• Pair innovation and knowledge teams with practising lawyers. Specialist teams can manage access and permissions, but practitioners must shape and tweak the inputs.
• Be open with clients about experimenting. Sharing successes and failures builds trust and can reveal where clients need help, but present AI as an operational tool rather than magic.
• Plan for the economics. Clients should not expect fees to fall automatically, and firms should test whether agent workflows remain viable if consumption-based pricing arrives.