If 2023 was the year of the copilot, 2024 is certainly on its way to becoming the year of the ‘agent’. We’ve seen ‘the world’s first digital SDR’ (11x) internal agents for enterprise teams in supply chain, financial services and healthcare (Adept), and we have seen agents for regulatory compliance (norm.ai). Agents feel like the latest part of the AI hype cycle, promising to automate entire workflows from beginning to end, meaning entire roles can now be delegated to AI.
Most agents are currently carrying out relatively low effort, high volume tasks, which can be automated at scale, such as Klarna’s customer service bot, which, according to them, does the work of 700 members of their full time customer support team.
It’s unsurprising that ‘agentic’ AI has made legal technology vendors take note. The idea of building software that can automate entire workflows is undoubtedly alluring, driving up the ROI for in house and private practice clients alike. But does the reality stack up? Is this another example of a VC funded disconnect or is there real customer enthusiasm for this technology in legal?
What is an ‘agent’?
According to AWS (Amazon Web Services), “an artificial intelligence (AI) agent is a software program that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals. Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals.”
An agent is autonomous, or at least semi-autonomous, so it is capable of taking decisions and then acting on them independently. This is an advancement of the ‘copilot’ model, where users collaborate closely with AI software to get to their final desired result.
However, agents have come to be defined as many things. More broadly, agentic AI can also constitute any AI workflow where there are several steps that depend on each other, rather than a ‘one-shot’ response which the user then tweaks or instructs again (à la ChatGPT). These multi-step workflows may significantly drive up the quality and accuracy of the AI’s response. In expert professional work, these sorts of workflows are often essential to capture the nuance and complexity of the documents being analysed.
What about legal ‘agents’?
Legal agents are very much du jour in the second half of 2024. Spellbook released Associate, the “first AI agent for law”. It’s an impressive product that allows you to plan and execute work with some human supervision, and update multiple documents (i.e. contracts) in one go, across multiple programs. BRYTER also released its own ‘AI agents’ that can respond to internal queries which tackle a number of tasks “from replying to recurring questions and drafting email replies to research and audits.” Flank is positing itself as ‘agentic AI for internal, expert support’, seeking to share internal knowledge with key figures in different teams in companies. Many of these are legal and compliance related questions that usually an in-house counsel would be responding to.
Do legal teams want AI agents?
Data is decidedly limited, which is unsurprising given how novel these ‘agents’ are. There are differing views from law firm practitioners. On the one hand, the idea of delegating entire workflows to AI which returns it perfectly, or better yet carries it out end-to-end, for a fraction the price of a paid up NQ or trainee, is highly appealing. However, there is the nagging sentiment that AI cannot yet be trusted to carry out complex legal tasks, and instructing AI to not only return one answer, but carry out length processes with multiple steps is beyond a bridge too far. Which leads to my next point…
Do they work?
Probabilistic software is hard to productise. For some products the arbitrary or random nature of what an AI system can produce is a feature, rather than a bug. No one wants something that may be significantly cheaper, but is significantly less accurate. This is particularly true (perhaps more than any other vertical), with legal. In most cases, you need a trained human in the loop to validate the outputs, and collaborate with the AI to get to the final point. I can’t speak for Spellbook, Flank or Bryter’s quality (though the traction would suggest positive things!), but my view is that ‘agents’ as we currently call them, could also be souped up copilots who can perform tasks, but require the oversight of an expert user to guide them to their destination. And frankly, that is the future we at Wexler want to build. We believe that the best ‘agents’ will be nonetheless the collaborators of experts in their fields, who can impart their own knowledge about what to do and what to not, then correct the output, reframe, redraft and revise, before distribuitng it. After all, isn’t this why AI was created? To ‘augment’ the human, rather than replace them?
What about Wexler?
Our own ‘agent’, KIM, tackles complex questions by breaking them down into sub questions, then searches over the database of facts our pipeline has extracted, then creates thematic summaries, posing follow questions and suggested actions. It can then, with some supervision, build AI timelines (chronologies) of the key facts. We’re ‘agentic’ in the sense that KIM takes a question, breaks it into its constituent parts and decides independently which follow up questions to ask, and which thematic summaries to create. Wexler works in tandem with the user to create high quality outputs, which lawyers check and iterate on to get to the required standard. We are using this same framework to allow users to carry out different complex tasks around their factual analysis, like spinning up a new 'pipeline' for a dramatis personae, a list of issues, or then drafting the core points of the statement of case.
What does the future hold?
As companies grapple with the developing capabilities of Large Language models, new business models will develop. Perhaps ‘agents’ will be overseen by human experts on the vendor side, rather than the customer side, selling perfected outputs direct to businesses and consumers. We at Wexler are confident that our multi-step approach dramatically improves the quality of the output, meaning less human intervention is required. Nonetheless, the way our customers tweak and refine the output is what ultimately creates great work.