I’ve been hiring recently and a candidate prompted some reflection during a conversation. “Over the last 2 and a half years [i.e my tenure at Flank to date], how would you say the legal industry has changed?”
Well, I remember the early days of RAG based legal chat bots.
Then came a wave of existing software powering their product with AI.
Alongside this emerged new legal AI tools.
Then came an alternative to AI tools, in the form of AI agents.
That’s where Flank stands, except we have been betting on this from quite early on, and the market now seems to be catching up.
That said, there is room and appetite for both tools and agents in today’s market. I think it’s helpful to understand the distinction.
Tools and agents are not the same bet
A legal AI tool is akin to a very powerful set of tools for lawyers to do their work a lot faster, more precisely, more efficiently, so to speak.
A legal AI agent deployed autonomously in an enterprise is responsible for a workflow end-to-end, under the supervision of a human, whether that’s the in-house legal team, a law firm or a vendor partner.
They are two very different ways of tackling resource constraints.
On the one hand, you can hand over powerful tools so your team can work faster. On the other hand, you can take away entire workstreams and bring in agents to own the bulk of the execution, with your team responsible for supervision and escalation points.
Different means for different needs and potentially, to different ends.
Everyone starts by asking for control
In my experience building and deploying agents in enterprise environments, the starting sentiment often tends to be the same: wanting the control that the toolbox offers. What my team sees soon enough is that the real desire is to simply reduce the high cognitive load on low value work.
Lawyers, understandably, want control because we need to remain fully accountable in the high stakes environment we operate in. I think everyone has been on a journey trying to figure out what that means for them. Forgive me for generalising here, but there is no doubt that LLMs now know more law than any of us can hold in our heads. However, there is still that insurance in judgment that only a human lawyer can provide.
That is exactly why AI agents need to be put to work for their incredible capabilities and speed, in so far as they are operated within intentional guardrails and a supervision framework that feels effortless for the human in the loop.
What this looks like in practice
Take inbound NDAs. I know you’ve heard about NDA automation ad nauseam but let’s be honest, they are a relatable headache: every legal team has them and nobody enjoys reviewing them.
Agents living where work lands.
Day-to-day, that means an agent sits in the place where legal work is already arriving from the business. This could be a shared legal inbox, a ticket queue on a work management platform, the sales channel where someone sends an NDA with no context, tags legal and says “Can we sign this? Need to move quickly on this”. The agent here is not a separate platform the team needs to access. If it were, it would most likely be a tool and we would be back to the toolbox set up.
An agent effectively triaging requests as they land, routing them to the correct human or AI agent.
The NDA lands. The agent works out what it is, who is involved, what additional deal context is there (or flags what information is missing), whether it’s on first or third party paper. First party paper with no changes can go straight back out for signature. A third party paper goes to the review agent for a first pass review (same goes for objections on first party paper - they get routed to the first party paper review agent). If something lands in that inbox that’s not an NDA or that needs to be routed to another agent, human lawyer, or business department, it gets routed by the triage agent automatically. The team does not have to expend any cognitive effort on triaging and doing first pass reviews on legal tasks that can be done by the rule book (quite literally, that’s what playbooks are for!)
That AI agent produces a first pass to the standards of the legal team, because it will have been thoroughly trained to do so.
This is where most of the actual leg work happens and that too, before the agent is even switched on. Dialing in an agent involves codifying a set of playbook rules that everyone is happy with, including extracting the rationale behind the occasional redlines and the ‘it depends’ types of micro decisions that human lawyers make all the time. It’s important that we close any gaps and resolve any conflicts in a playbook. Once that’s done, we start a thorough testing and fine tuning phase during which the agent is put through its paces, provided feedback, improved, and so on until we achieve “what good looks like”.
A system that enables humans in the loop to easily check, edit, approve an agent’s work.
The truth is, no matter how much fine tuning we do, an agent will never be truly perfect, especially because we can never be in full control of or predict what will land. That’s why it’s critical to be able to supervise an agent easily and shape the output as required. The more fine tuning in the early days, the less the need for corrections down the line.
The skill worth keeping
Supervising an agent is as simple as supervising a junior’s work, except there is a lot less interpretation required. The supervising lawyer can see exactly what the agent did, why, including a map to the source rule that led to that action. A full audit trail, making it very easy to check the work and remain accountable for it.
In a nutshell, we have agents doing the work and lawyers deciding whether the work is acceptable. Not too dissimilar to running a team.
In outsourcing the heavy lifting, we are not losing cognitive ability. On the contrary, we are honing a more valuable skill: perfecting legal judgment, and saving time and cognitive load in the process.
Scalable automation
The potential gap in impactful legal AI right now is being created because teams are choosing one of two routes. Either they retain painstaking control of execution through assistive tools or they hand workflows to AI without a proper supervision loop and lose control completely. Both are risky.
For the higher complexity work, the toolbox certainly comes in handy. However, as far as lower value and lower complexity work is concerned, I think the better approach is standing up a set of agents that can be trusted to deliver work to your standards, with a strong supervision framework around them, so they are scalable, adaptable and auditable all at the same time.
