Agentic AI moved into legal work faster than most teams had time to think about. Tasks that used to fill our calendars are shrinking, and new work has shown up to replace them. We are prompting, steering, reviewing, reconciling, and deciding when to trust an agent. Almost none of that is written down, and almost none of it has a clear owner. The work is real, the stakes are real, and most teams are figuring it out as they go.
Natalie Kim, VP of Legal at Omnidian, guest hosted this How to Contract webinar in place of Laura Frederick. She was joined by Danish Butt, Managing Director at Swiftwater and Company, and Colin Levy, General Counsel at Malbek. Natalie runs her company's legal team and its enterprise AI adoption and governance work. Danish advises General Counsel on building modern legal functions. Colin runs the legal function at a contract lifecycle management company and writes widely on legal tech. The three brought different vantage points on the same set of questions and made the conversation unusually concrete.
The discussion covered where AI frees up lawyer attention and where it quietly fragments it, how to think about trust and stakes when deciding what to delegate to an agent, what the new oversight and audit work actually looks like in practice, the professional responsibility issues that show up when agents start chaining actions, and how small teams in particular can build their own tools without waiting for an enterprise buying cycle.
Here are our top ten takeaways from the speakers' comments during the webinar:
Decide how you want to spend your freed-up time, or someone else will decide for you. Colin made the point that AI gives back real hours on tactical work. The harder question is what to do with those hours. If you do not actively answer that, the question gets answered for you by whoever is pinging you next. The reward of using AI well is being in control of your own time, not just being faster.
Use trust and stakes as your filter for what to hand to an agent. Danish framed it as a sliding scale. Where the consequence of a wrong answer is low, agents earn the work. Where the relationship matters or the stakes are high, you stay in the seat. We don't need a perfect rule, just a clear sense of which bucket a task lives in BEFORE we delegate it.
Quality of input drives quality of output, every time. Colin kept coming back to this. A short, sloppy prompt produces a confident, useless answer. The lawyers who get the most out of these tools are the ones who treat the prompt as part of the work, not a throwaway. We have seen the same thing on the receiving end, where a business client comes in with a long AI-generated read of our own policy and we have to ask what they actually asked the tool.
Don't mistake speed for leverage. AI is fast at producing things, and that is a trap when speed becomes the only metric. Cycle time goes down, but quality can go down with it if nobody is reading the output carefully. Pay attention to whether the answer is actually better than what you would have produced on your own, not just whether it arrived sooner.
Treat agent building as a product lifecycle. Scope, test, launch in a pilot, watch for breakage, update when the underlying playbook changes. Natalie described building an NDA agent for her sales team by inviting them to try to break it during a two to four week pilot. Wait times dropped from two weeks to five minutes, and she still has a quarterly reminder on her calendar to check that the agent is doing its job.
Spend your hardest thinking on the work agents can't do well. That is judgment, context, reading the room, and weighing which risks actually matter for your business. Colin's point was that AI lets us push tactical work down and pull this kind of work up. The reward is more time on the one-of-a-kind problems and less time on the things you have already seen a hundred times.
Get clear on roles before you start building. Who scopes the agent, who builds it, who owns it after the builder leaves the company. We are already seeing orphan agents inside companies, still running, still talking to the business, with nobody clearly responsible. Even a simple list of agents and a quarterly review cycle is worth more than nothing.
AI shines a light on weak processes. If your playbook lives in someone's head, you cannot encode it into an agent. Danish talked about clients trying to deploy tooling while still negotiating their own playbooks across newly acquired entities. The agent project surfaces every gap that was already there. We can either treat that as a tax or as useful information.
Build a governance capacity into the legal function itself. Models change, vendors push silent updates, a feature that wasn't there yesterday appears overnight. Danish made the case that governance is now a permanent part of how legal operations get designed, not a one-off project. Confidentiality, privilege, conflicts, and supervision all need to be revisited in light of how these tools actually behave week to week.
Solo and small teams have a real advantage in the build versus buy question. Natalie made the case that a twenty dollar Cowork or Claude subscription plus a clear problem to solve will outperform a five thousand dollar enterprise tool you bought before you knew what you wanted. Start with the problem you actually have. Pick the smallest agent that solves it. Get familiar with the tool by using it, not by reading about it.
Subscribe to Stay in the Loop
Our weekly newsletter keeps you current on upcoming How to Contract events and recaps like this one. Subscribe now so you never miss the practical insights.
