AI LEADERS
Agents at work
Plenty of potential in genAI agent-led process, say leaders
Two-fifths are at firms where there is an agent in either a production workflow or a specific pilot
This round of polling focuses on the extent to which UK law firms are already working with ‘agents’ — the process of pointing genAI tools at a specific defined goal, rather than issuing prompts, and permitting the AI to determine some sequence of necessary actions, within a set of parameters, to achieve the outcome or pass the process baton back to a human for further steps or checks.
In the survey phase of this research — which saw 50 responses from leaders with AI responsibility at firms across the market — we defined an agent at the outset as “a genAI tool/capability that has a user-defined goal but can plan and progress work by taking multiple steps on its own — with access to one or more systems/sources of information — before a human then reviews the result”.
Two-fifths of AI leaders report that their firm is at the very least already piloting an agent (13% have them at work on production workflows) — more than say that this direction is currently only being “explored conceptually”. And a quarter of these say that use is already widespread while 15% can see “mature” use in some pockets of the firm. On the other hand, a fifth work in firms where use is “confined to specific functions” and 15% require any agent-adjacent operatives to be approved.
13%
of AI leaders say the firm has already deployed one or more agents within production workflows
Mirroring genAI trends that Briefing has already identified, areas of legal business where leaders indicate their firms are most likely to be exploring the potential are working efficiently with documents (eg review and drafting), business development and knowledge management work — followed by risk and matter management. While immediately client-facing outcomes are a fairly obvious space where firms are likely to be more cautious, it’s perhaps surprising they are less likely to be focused on avenues of internal communication/triage, such as the traditional helpdesk to handle user queries in a more efficiently automated manner (35%) — but perhaps this is administrative terrain for less advanced genAI to take the requests.
Two-fifths of those with any agent on the roadmap report that the firm has already built a bespoke agentic solution of its own, with another third in the process of doing so. Only one in 10 works at a firm not exploring this opportunity within at least one area of process.
Examples given include:
Large-scale document review, then passed to lawyers for human review
Document data extraction to compare term changes across versions
Simpler extraction of fee earner-focused content in credentials management
Improved efficiency in private equity deals
Managing routing of enquiries through the website
Signposting to policies such as HR.
But whatever the mission, two points are very clear. The first is that leaders know people need some guidance to build these solutions safely and effectively. Second is that most are using a combination of API integrations and the open-source model context protocol (MCP) that introduced standardised, cross-platform integration with multiple different data sources and workflows, without a need for custom code — although a quarter are unsure whether this is the case.
Evan Morgan, head of bids and client innovation at DAC Beachcroft, says: “The underlying data needs to be as bulletproof as possible. The more reliable, the greater the quality and consistency of the agentic output. In that instance, it starts to add real value in terms of quality of output but also time and efficiency savings.
“We have invested in AI governance, guidance, approved tools, mandatory training, knowledge-sharing resources and an AI Champion network to help drive responsible adoption. But it’s not just about the theory. We have best practice recommendations and clear governance on how to use AI responsibly, but we encourage lawyers to speak up when they’ve come up with a different or better approach underpinned by this technology. We also actively encourage and facilitate peer-to-peer learning – often the most effective way of demonstrating how the technology can be applied to a requirement.”