The default way to use an AI assistant in a law practice is to open one and ask it for things. Research this. Draft that. Check the other. One conversation, one general-purpose helper, infinitely flexible.
It works, in the sense that output appears. But it has a failure mode that took me a while to name, and once you see it you cannot use a general assistant the same way again:
A generalist with no defined scope has no defined handoffs, and if there are no handoffs, there are no checks.
Everything happens in one conversation, performed by one participant, reviewed by nobody. The entity that drafted the argument is the entity that assessed whether the authority supported it. The entity that wrote the brief is the entity that decided it was finished. In any other professional context we would recognize that immediately as a control failure. We do not recognize it here only because it is happening inside a chat window instead of across desks.
How a real office is organized
Consider how the work actually moves in a functioning firm, and notice that almost none of the structure is about capability.
A research and writing attorney develops the argument. A citation checker verifies the authority — a separate person, deliberately, because the person who built the argument is the worst possible person to test it. A filing clerk owns caption, spacing, signature block, certificates, and the local rules, and does not touch the argument. A billing person owns time entries. An intake coordinator owns conflicts and engagement. A calendar keeper owns dates.
Every one of those roles is defined at least as much by what it may not do as by what it does. The filing clerk does not rewrite an argument they think is weak. The citation checker does not decide the brief is fine because the writer is senior. The limits are the point. The limits are where the reliability comes from.
So that is how the AI in this practice is organized: twelve specialists, each with a written brief. Not one assistant asked to be twelve things.
What goes in an agent brief
An agent brief is a document — plain text, versioned, readable — that defines one specialist. Ours have five parts, and the middle three do most of the work.
1. Identity and domain. Two or three sentences. “Legal research and writing specialist. Drafts and analyzes pleadings, motions, briefs, discovery, demand letters, and research memoranda. Use whenever the work product involves legal analysis or formal legal prose.” Concrete enough to route work by.
2. What this specialist owns. The actual scope, enumerated. Not “handles legal writing” but the specific artifacts and the specific decisions.
3. What this specialist must NOT do — and where it goes instead. This is the most important section in the document and it should be the longest. Every exclusion names its destination: “Do not do final court-rule formatting — margins, caption layout, signature block, certificates of compliance. That goes to the formatting clerk. Do not verify citations; that goes to the citation specialist, and it is mandatory, not optional.”
The reason exclusions must name a destination is that an exclusion without one is just an unmet need, and unmet needs get met by whoever is holding the file. “Don’t do X” produces a specialist that does X anyway when X is obviously required. “Don’t do X, hand it to Y” produces a handoff.
4. Standing rules. The non-negotiables for this role. For the writing specialist, the standing rule is that it may not produce a citation from memory — either the authority was retrieved and read in this session or it goes in as a marked placeholder. For the formatting clerk, it is that it copies the firm’s actual template file and populates it, rather than rebuilding a caption by hand. These are the rules that exist because something went wrong once.
5. Output contract. What “done” looks like, and what accompanies it. The writing specialist’s output includes a manifest of every authority relied on and what each was retrieved from — which is what makes the next specialist’s job possible.
The chain is the control
Here is the part that matters more than any individual brief. Work moves through the specialists in a fixed order, and the order encodes the checks.
research & drafting → substantive argument
↓
citation verification → every authority retrieved and read
↓ (MANDATORY GATE — nothing passes unverified)
court-rule formatting → caption, spacing, certificates, template
↓
the lawyer → reads it, signs it, files itBecause the citation gate is a stage rather than a practice, skipping it requires noticing that you skipped it. That is the entire difference between a control and an intention. Intentions fail on the day the brief is due, which is the day they were supposed to work.
Note also what happens at the end of every chain: a lawyer reads it and signs it. That is not a courtesy step and it is not delegable. The specialists produce a draft. A lawyer files.
Why this is not just prompt engineering
A fair objection: could you not get the same result by writing a long prompt that says “draft this, then check the citations carefully, then format it”?
You can get something like it, and it will work most of the time, and the times it does not will be invisible. Three real differences:
Separation of context. When the citation specialist runs, it does not see the drafter’s reasoning about why the case is perfect for this. It sees the document and the sources. That is not a stylistic choice — it is the whole reason a second reader catches things. A reviewer who has read your argument will follow it and agree with it. Give the checker the claim and the source, not the advocacy.
Instructions in one conversation compete; briefs do not. In a long single conversation, “verify every citation against the source” is one instruction among two hundred, competing for attention with everything else in the context, and its influence decays as the conversation grows. A specialist whose entire brief is verification does not have that problem, because there is nothing else in its instructions to lose to.
Briefs are artifacts you can improve. When something goes wrong, you edit the specialist’s brief and every future invocation inherits the fix. A prompt you typed once is a lesson learned once. This is the compounding part: our briefs have accumulated a standing rule for nearly every failure this practice has had, and none of those failures can recur in the same form.
The supervision question
None of this is a way of transferring responsibility, and I want to be direct about that, because “the specialist did it” is a sentence with no professional-responsibility content whatsoever.
Minnesota Rule of Professional Conduct 5.3 — “Responsibilities Regarding Nonlawyer Assistants” — requires a lawyer with direct supervisory authority to “make reasonable efforts to ensure that the person’s conduct is compatible with the professional obligations of the lawyer,” and makes the lawyer responsible for conduct that would violate the rules if a lawyer engaged in it, where the lawyer ordered or ratified it.
The ABA’s Formal Opinion 512, issued July 29, 2024, addresses generative AI directly and routes through the same familiar duties — competence under Rule 1.1, confidentiality under Rule 1.6, communication, and reasonable fees under Rule 1.5.
Whatever the eventual doctrinal answer about how supervision rules map onto software, the working posture is not in doubt, and I would argue the specialist structure makes it easier rather than harder to satisfy. A written brief defining scope and limits is a far better answer to “what reasonable efforts did you make” than a description of how you usually phrase things. It is documented, it is auditable, and it is the same document every time.
Practical notes from actually doing this
Start with three, not twelve. Drafting, verification, formatting. That chain alone captures most of the benefit, because it contains the one gate that matters. Add specialists when you notice yourself repeatedly explaining the same role.
Let the brief define the reporting line. Our org chart is generated by reading each specialist’s own brief — the formatting clerk sits under the writing specialist because the clerk’s brief says so, not because a second list somewhere says so. One source of truth. A separate list is a list that will drift.
Write the limits from real failures. The temptation is to write aspirational briefs describing an ideal specialist. Better ones are scar tissue. Every standing rule in ours exists because something specific went wrong: a document rebuilt from scratch instead of from the template, a rule cited from memory that had been renumbered, a placeholder that shipped in the wrong color. Aspirations do not prevent recurrence. Specific prohibitions do.
Pass context explicitly. A specialist invoked as a fresh session cannot see the conversation that dispatched it. “As discussed above” means nothing to it. Every fact, every source, every constraint must be in the handoff. This is annoying exactly once, and then it becomes a discipline that improves the work — being forced to state what a specialist needs surfaces the assumptions you were carrying implicitly.
Do not let a specialist grow. The pressure is constant: the drafting specialist is right there and it could probably handle the formatting too. Every time you allow it, you delete a handoff, and the handoff was the control. If a role is genuinely too narrow, rewrite the brief deliberately. Do not let it expand by convenience in the middle of a deadline.
Sources
- Minn. R. Prof. Conduct 5.3 (Responsibilities Regarding Nonlawyer Assistants)
- ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512 (July 29, 2024)
Rules quoted were verified against the Revisor of Statutes. General commentary on practice management; not legal advice and not ethics advice. Formal Opinion 512 is advisory and not binding in any jurisdiction.