The Unease
You have already done it. Dropped confidential facts into a consumer AI to “get a quick sense of risk,” watched it generate a redline cleaner than what a tired associate would have produced at 1 a.m., and felt a mix of excitement and unease. The excitement is obvious. The unease is harder to name. It is not fear of hallucinations. It is not fear of bugs in the code. It is the sense that at some point, the act of configuring these systems stops being “using tools” and starts being the thing lawyers are actually paid to do.
For most of the last fifty years, the tools of legal practice were inert. A word processor, a research database, a form file — they amplified human effort, but they did not substitute for legal judgment. You could hand the same Westlaw terminal to a first-year associate and a Supreme Court Justice and get radically different outcomes, because the machine never decided anything on its own. It did not apply a standard. It did not choose a strategy. It waited.
Agents Do Not Wait
AI agents do not wait. They decide.
A well-configured contract agent does not simply highlight every indemnity clause in yellow. It weighs them. It knows, because someone told it, that this client lives with mutual caps but never with uncapped consequential damages, that “industry standard” means something different for a pre-seed SaaS company than for a defense contractor, that Delaware law is fine here but New York is non-negotiable. It can be wrong about all of this. But it is no longer a highlighter. It is an opinionated system.
Once you see that clearly, the fiction that configuration is just engineering collapses. The decision to treat some risks as red flags and others as yellow, to escalate some outputs to a human and let others auto-ship, to treat a pattern of facts as “routine” or “escalated” — those are not implementation details. They are legal judgments, implemented in code and prompts and routing logic instead of in Word docs and email threads. The medium changed. The judgment did not.
The profession is not ready to admit that. Admitting it would require acknowledging that a growing share of what we call “legal work” is migrating from the surface — the document you see — into the substrate that produced it. It is happening anyway. The law has started to draw the lines that matter.
Arizona and the Lines That Matter
Arizona’s ABS regime is one of those lines. When the state abolished Rule 5.4 in 2021 and opened the door to alternative business structures, the headline was ownership: law firms could take outside investment, give engineers equity, and operate more like tech companies. That was the news peg. Something subtler happened underneath. For the first time, a jurisdiction said out loud that a law firm is defined not by its partnership structure but by what it is licensed to do for clients. The firm could own code. It could employ non-lawyers at scale. It could behave like a product company. What it could not do was outsource ultimate responsibility for legal judgment.
Embed AI into that picture and the stakes sharpen. In a traditional firm, responsibility for judgment is easy to visualize — a name on the signature line, a bar number in the footer, a partner down the hall. In an AI-native ABS, responsibility sits in places invisible to the client: in policies about which models handle which tasks, in prompts that encode a firm’s tolerance for uncertainty, in routing rules that decide when human review is mandatory and when it is optional. Those choices get made in a Jupyter notebook instead of a partner meeting. They are still choices about how legal standards are applied.
The End of the Disclaimer Era
Disclaimers start to look like a nervous tic from an earlier era. “This is not legal advice, consult an attorney” made sense when a company wanted to build something legal-adjacent without actually practicing law. It was a firewall between the product and the profession. It was honest.
Configuration is legal work. An Arizona-licensed ABS can stand up and say, “our lawyers designed this system, supervise it, and answer for its outputs.” Once that sentence is true, the disclaimer stops being a firewall and becomes a tell. It says the quiet part out loud: nobody with a duty of competence is claiming this system as their own.
Misconfiguration Is Malpractice
The alternative forces lawyers to own more of the stack than they are used to. If configuration is legal work, then misconfiguration is malpractice. An attorney who tunes a system to downplay certain disclosure obligations because they are “edge cases” is not editing a prompt. She is rewriting a risk model. A firm that deploys a contract agent without meaningful audits is not moving fast. It is, to put it plainly, experimenting on its clients — and heck, that is malpractice with better UX.
A Healthier Division of Labor
Naming configuration as legal work also produces a healthier division of labor. Engineers build infrastructure, models, and interfaces without pretending to understand fiduciary duties. Lawyers focus their attention on the parts of the system where judgment actually matters — the intolerable failure modes, the non-negotiable standards, the places where “good enough most of the time” is not good enough. The work product shifts from a memo you can print to a set of conditions the system is never allowed to violate.
Having built the first internal version of our own contract-review workflow — selecting which clauses get auto-flagged, which route to a human, which trigger a full stop — I can report that every one of those toggles is a legal decision dressed up as a setting. Choosing that indemnity over a certain cap auto-escalates to an attorney, and MFN clauses in a particular context do not, is not a UX choice. It is a standard of care, encoded. Call it what it is.
Privilege and the ABS Container
The privilege layer sharpens all of it. Under Heppner, privilege does not come from sprinkling legal words over a chat log after the fact. It attaches because an attorney-client relationship existed when the communication happened, because the communication was in confidence, and because a lawyer or a lawyer’s agent was providing legal advice. When the “agent” in that sentence is an AI system, configuring it is not optional. Configuration is the act that determines whether the client is speaking into a legal relationship or into the void.
That is why Arizona’s ABS model matters. Not because venture capital is romantic. Not because engineers like equity grants. Because it creates a container where one entity can both design these systems and credibly tell the client: when you use them, you are communicating with us as your lawyers. The capital structure is a means to an end. The end is treating configuration as native legal work while still building like a technology company.
What the Profession Does Next
Most of the profession will hold the old distinctions together for as long as it can. They will talk about “AI tools” as if a system that decides which risk is acceptable were a kind of spell-check. They will insist that the practice of law begins only when a human hand touches the document. They will keep their disclaimers. They will hope the doctrine stays blurry long enough to retire.
That hope will not survive contact with reality. A court will decide whether a system that applies legal standards to particular facts, under the supervision of licensed attorneys, is part of the practice of law. When it does, the answer will be obvious. The practice of law has not disappeared. It has moved.
The profession has a choice. Pretend configuration is someone else’s problem — the vendor’s, the engineer’s, the client’s. Or accept, openly, that in an AI-native world, configuration is where the real legal work lives. Choose the second path, and the disclaimers will look as dated as fax cover sheets. The signature line will survive. It just will not be the only place the law is practiced.