An in-house legal team runs the company's contracts, and today a lawyer touches every step of every one. AI is about to change that, but not who answers for the result, and it changes a startup's legal team differently than a large regulated company's.
An in-house legal team does two jobs at once. It is the company's contract engine, turning out the NDAs, sales agreements, and vendor contracts the business runs on. And it is the risk gate, the last read before the company signs its name to something it will have to live with. Nearly all of that work is the same handful of moves, over and over: a request comes in, and an agreement gets drafted, negotiated, approved, signed, and one day renewed or ended. About thirteen steps, repeated thousands of times a year.
Today a lawyer touches all thirteen, even the routine ones that never needed one, like the standard NDA that comes back signed with nothing changed.
That is what AI changes. As it takes over the routine, the lawyer is pulled off the steps that never needed them, and the familiar idea of a human in the loop splits into three, depending on how much is at stake. Over the loop, the lawyer sets the rules once and sees only the exceptions. On the loop, they watch from the edge and step in on the calls that matter. In the loop, they keep their hands on the work and sign it. Which work belongs where is the one idea the rest of this page is built on.
Not all legal work is the same, and that is the whole key. Some of it is routine, with one right answer. Some of it turns on judgment. And a little of it is a moment where a person has to answer for what happened. Which kind it is decides where the lawyer stands: the riskier the work, the closer in they move, the less rope the machine gets, and the more they are on the hook. Here are the three.
Flip between Today and AI-native to see what changes. Change the company size to see how much of the AI model is switched on. Click any step for the detail: how it is done today, how it changes, the limit that keeps it safe, and who is on the hook.
Click any step above to see how it is done today, how it changes with AI, who does it, where the lawyer stands, the limit that keeps it safe, and who is on the hook for the call.
Across all thirteen steps
The model underneath does not change with the size of the company. What changes is how much of it is switched on, and how much room each agent has earned. Use the size buttons in the interactive above to watch the same thirteen steps shift from mostly-human to mostly-self-serve, while the highest-stakes calls stay put no matter the size.
One lawyer running two or three agents and doing all the judgment by hand. The whole model, in miniature.
The core contract agents on off-the-shelf tools, plus a knowledge base the team built itself.
The full set of contract agents, plus privacy and knowledge ones, with the system learning as it goes.
Every agent switched on, each given more room as it earns trust, and a standing team just to govern the whole thing.
A team built this way keeps getting better. The line between routine and judgment is not fixed: work that takes real judgment this year can become routine the next, once the team notices it keeps making the same call the same way. Only the AI-native version has a loop for capturing that, and that loop is what keeps the team improving instead of leveling off.
Watch where a lawyer changes what the agent produced. That is your live map of where judgment still matters.
When the same change keeps happening, either teach the agent the rule behind it, or keep it as a call a human always makes.
Write the reason down, in your company's own terms, in a running record of how you decide. That record is the team's built-up judgment.
Once a call becomes routine, it moves from the judgment lane to the routine lane, and the lawyer moves up to the next hard problem.
A legal team is not a generic operations group. A lawyer carries legal duties that a tool cannot take on. Those duties are the reason a lawyer stays close to the work wherever the stakes are highest, no matter how good the AI gets.
A company lawyer owes the company real legal duties: loyalty, care, competence, confidentiality, honesty. An agent can do the work, but it cannot carry the duty. So the highest-stakes calls stay with a person who answers for them.
A supervising lawyer answers for the work of the people and tools under them. You meet that duty, and prove you met it, with right-sized records and a human sign-off. An agent giving something a pass is not the same as review.
A lot of legal work is privileged, which means it is protected from the other side in a lawsuit. Feed it into a public AI tool and you can lose that protection. So privilege is built in from the start: private instances, projects walled off by matter, a locked-down AI tier, and records kept light enough to prove diligence without storing the sensitive substance.
Once you have a tool that reliably catches a problem, ignoring what it finds is worse than never having looked. Being able to catch it creates a duty to act on it. So when the system raises a flag, you resolve it. You do not leave it sitting there as a record that you knew.
What this changes. For a lawyer, good and fast is only half the question. The rest is whether they can stand behind the work, whether it is protected, and who answers for it. That is what pushes the high-duty work closer in, and it shapes the records the team keeps.