Ask a room of lawyers what they do all day and most will say some version of “I write.” Contracts arrive as Word documents, advice goes out as memos and emails, and the work gets measured in pages. Watch the decisions instead. Which indemnity cap belongs in this agreement, given what the product actually does with data? Which of the approved fallbacks do we concede when the other side pushes? The sentences were written a long time ago. The lawyer's job is choosing the right ones for the facts in front of them, and that job looks a lot more like configuration than composition.
I started seeing this early. At Gunderson Dettmer I drafted venture financings off firm forms, and a financing document is really a configurable instrument: valuation cap, discount, pro rata, most favored nation. Nobody invents the document on a Tuesday afternoon. The lawyer sets the parameters to the deal, and the document is what those parameters look like when they print out.
In-house the pattern repeats at higher volume. At Synthesia I negotiated against a playbook: the positions we open with, the ones we accept, the ones we walk from. A playbook is configuration written down. A good one compresses a three-week negotiation into a few days because the decisions were made before the other side's paper ever arrived.
For a stretch I left the practice seat entirely. At Remote I built a contract management tool on top of contract automation and employment infrastructure, which meant taking legal judgment and turning it into fields, states, and rules so a document could assemble itself from answers. That exercise forces an uncomfortable question. If a form plus a decision tree produces the same document the lawyer would have, what was the lawyer contributing? My answer at the time was judgment about the parameters. I still think that's right.
What's changed since then is the executor. The judgment still comes from a lawyer, but the execution increasingly runs on agents that assemble the document, apply the playbook, and route the result to the next step. My own files already run this way. The documentation stays the source of truth, and the skills an agent executes are the deployable layer on top of it.
Here's where this stops being a workflow observation and becomes a regulatory question. Every state reserves the practice of law to licensed attorneys, and New York says it as plainly as any:
It shall be unlawful for any natural person to practice or appear as an attorney at law … without having first been duly and regularly licensed and admitted to practice law in the courts of record of this state.
N.Y. Judiciary Law § 478; a violation is a misdemeanor under § 485
Now consider a company that configures an AI system to generate contracts for its customers: choosing the clauses, setting the fallback positions, encoding the judgment about which risks matter. Who is practicing law there? The software didn't choose anything. Someone set the parameters.
If that someone is a lawyer, the obligations are already written. Competence now includes the tooling:
To maintain the requisite knowledge and skill, a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology.
ABA Model Rule 1.1, Comment 8
To competently use a GAI tool in a client representation, lawyers need not become GAI experts. Rather, lawyers must have a reasonable understanding of the capabilities and limitations of the specific GAI technology that the lawyer might use.
ABA Formal Opinion 512 (2024)
Supervisory duties point the same direction: a lawyer who relies on nonlawyer assistance owns that assistance (Model Rule 5.3), and the ABA reads generative AI tools into exactly that category. If the person setting the parameters isn't a lawyer, they're drifting toward the unauthorized-practice line every state draws, and “the system did it” is not much of an answer when the system only did what it was configured to do.
I'd go further, because I think this is where the practice actually ends up. Right now the model is supervised configuration: the lawyer sets the parameters, the agent executes, the lawyer reviews the output. The next stage is delegated execution: the lawyer configures, the agent works unsupervised inside those parameters, and the lawyer audits instead of reviewing. And the stage after that is the one the rules will have to name. The configuration itself becomes the representation. The client relationship runs to the lawyer who set the framework, the agent's output is that lawyer's work product, and the regulated act moves from drafting documents to designing, testing, and maintaining the system that drafts them.
That's a complete practice of law built on configuration: you configure the agents, they do the unsupervised work, and your license stands behind everything they produce. The bars that adapt fastest will be the ones that treat the act of configuration as the thing they regulate, license the configurer rather than the tool, and hold the framework to the same standard the drafting ever met. Call it configuration if you want. The statute books call it practicing law.