When the first widely available AI chatbots arrived, most lawyers did not know what to make of them. That included lawyers who were curious about the technology. The tools were powerful, but unfamiliar. They looked simple because the interface was simple. A blank box. A blinking cursor. A place to type a question, treating it like a more familiar answer machine - Google.
But generative AI is no Google, and that approach yielded unimpressive results, and this is because no one really knew how to get the best work out of them. No one knew where they fit in a law practice. No one knew whether they were research tools, drafting tools, brainstorming tools, intake tools, or something else entirely. Most of us were learning in public
I started writing about those questions early. In a series of articles for The Champion, I tried to explain what these tools could do, what they could not do, and how criminal defense lawyers might use a variety of prompting approaches to get the most out of them without surrendering judgment. Those articles led to speaking engagements around the country, usually in rooms filled with lawyers asking some version of the same practical question: how do we use this technology without putting our ticket at risk?
At the time, the answer had a lot to do with prompting. That made sense. The chatbot era rewarded the lawyer who could ask better questions, provide better context, and give better instructions. A vague prompt produced vague uninspiring work. A more disciplined prompt could produce something useful, sometimes surprisingly useful and at times even brilliant. That’s when we knew these things were for real.
Three years later, however, that paradigm is changing. But not because prompts no longer matter. They do. Instructions, context, and constraints still shape the output. But the central act of AI use is shifting. The future may not belong to the lawyer who writes the best one-off prompt into a chat window. It may belong to the lawyer who builds, supervises, and understands better loops.
Agents and recursive prompting are the first reasons why. Recursive prompting means that one output becomes the input for the next step. That structure should feel familiar because it reflects something basic about how generative AI works. These systems generate text sequentially, using existing context to produce the next part of the response.
Recursive prompting applies that same general idea at the workflow level. Instead of using prior words to generate the next words, the system uses a prior answer to generate the next task, critique, revision, or check. In that sense, loops are not foreign to the technology. They are a practical extension of its underlying structure
In an agentic workflow, the system may also retrieve documents, compare sources, use tools, prepare structured outputs, and stop only when it reaches a defined endpoint. This use case means the prompt is no longer the main event, and increasingly, the prompt is buried inside a workflow.
That shift matters for lawyers because it changes the professional question. The issue is no longer only whether we can fashion a better prompt for better output. The issue is whether we can supervise an automated process that produces legal work through repeated acts of machine-generated analysis.
The Chatbot Era Rewarded Better Questions
My articles on prompting attracted attention because it was the part lawyers could understand and control. After all, asking questions is what we do.
Using various prompting methods, lawyer could ask for a motion outline, a cross-examination plan, a discovery summary, or a plain-English explanation of a legal issue. The quality of the instruction mattered. A vague prompt often produced vague work. A more careful prompt, with jurisdiction, procedural posture, facts, tone, and constraints, usually produced something better. As proficiency with prompting improved, lawyers came to realize the vastness of what these tools were actually capable of.
But prompting was only the visible part of the work. In legal practice, the useful product is not the first draft. It is the revised draft. It is the version that has been checked against the record, tested against the law, edited and sharpened for the judge, and stripped of overstatement, false attributions or incorrect or non-existent citations. Lawyers know this because lawyers do not generally trust first drafts, whether those drafts come from associates, experts, clients, or themselves.
In pretty much all cases, the first draft is only the beginning of several rounds of edits and improvements all based on a careful application of professional judgment, and that is why the loop matters.
What Recursive Prompting Actually Does
A single prompt asks the model to perform a task or tasks and produce a desired output. A recursive prompt asks the model to perform whereas an agentic loop asks the model to perform, review, revise, and continue under a defined process. That difference changes and expands the output and the nature of the work itself.
Instead of asking an AI system to “draft a suppression motion,” a lawyer might build a workflow that asks the system to identify the governing legal standard, separate facts from assumptions, draft the argument, identify weaknesses, predict the prosecution response, revise the argument, flag missing record citations, and produce a verification checklist. With appropriate monitoring and professional output, this is not a substitute for lawyering. It is structured revision and structured revision is already part of good lawyering.
A thoughtful lawyer drafts, rereads, checks the record, tests the weak points, revises, and then reviews again. A supervising attorney does the same thing with an associate’s draft. The value is not in the existence of the first version. The value is in the disciplined improvement process. The brilliance is that AI loops can imitate part of that process. That does not make them lawyers. It makes them potentially useful drafting environments.
Self-Critique Is Useful, But It Is Not Verification
The promise of recursive prompting is legitimate. A model that is forced to critique its own work may catch vague reasoning, missing steps, internal inconsistency, unsupported conclusions, poor organization, and failure to follow instructions. In practice, even a simple instruction to review a prior answer for factual assumptions, legal overstatement, and missing sources can improve an output.
For lawyers, that sounds familiar. It resembles the work we were doing before the advent of AI. But self-critique using AI is not the same as independent verification. A system can review its own work and still miss the central problem. It can restate a bad assumption more elegantly. It can smooth uncertainty into confidence. It can produce a cleaner, more persuasive version of a mistaken conclusion. That is the danger. A loop can improve a draft. It can also launder a mistake.
A quote from law school attributed to an unknown professor is that “the law is what is boldly asserted and plausibly maintained. Yes, fair enough, but when it comes to AI, the ability if not the propensity of the machines to create plausible but wrong output is alarmingly real.
This is especially important in law because polished error is more dangerous than obvious uncertainty. A rough answer invites scrutiny. A smooth answer can slip through review if the lawyer is tired, rushed, or overly impressed by the formatting.
The take away here is that a loop should make the lawyer more demanding, not less.
Why Recursive AI Changes Legal Supervision
Once the work becomes iterative, the lawyer’s supervisory obligation changes. The lawyer is no longer reviewing only an answer. The lawyer is reviewing the process that produced it and that is a that is a different, largely technological problem.
A chatbot response can be checked as a single output. The lawyer can read it, compare it to the record, test the citations, and revise it. An agentic loop may produce a final answer only after a series of internal steps. It may have summarized a document, used that summary to create an issue list, used that issue list to draft an argument, critiqued the argument, revised it, and then produced a final version.
If the lawyer sees only the final version, the lawyer may not know where the error entered the process, and that reality should change how lawyers use these tools. The lawyer should know what materials the system reviewed. The lawyer should know what task it was assigned. The lawyer should know whether it used outside sources. The lawyer should know whether it was permitted to make legal conclusions. The lawyer should know where the workflow stopped and where human review began.
Legal supervision cannot depend on a polished final answer. It has to reach the process and with the process comes greater not lesser responsibility.
What a Good Legal AI Loop Should Require
A good legal AI loop should be narrow, source-grounded, and reviewable. It should force the system to separate facts from assumptions. It should require citations to the record. It should preserve uncertainty. It should identify contrary facts or contrary authority. It should test the strongest opposing argument. It should distinguish legal conclusions from factual summaries. It should make clear what requires human review. It should be easy to review and most importantly, it should stop.
A loop without a stopping rule does not create professional judgment. It only continues the possibly faulty process. That stopping rule should be explicit. The system should know when it has reached the end of its permitted role. It should not communicate with clients, contact witnesses, file documents, make strategic recommendations, or give legal advice unless the lawyer has specifically designed, approved, and supervised that workflow. In many settings, the safer rule is simpler: the agent prepares work for lawyer review and then stops.
As lawyers begin to experiment with this idea, the first legal AI loops in a law office should probably be modest. They should help prepare the file. They should help organize discovery. They should help identify missing information. They should help generate questions. They should not decide the strategy.
A Criminal Defense Example
Consider a criminal defense lawyer reviewing discovery in an OWI case. A single-prompt workflow might ask the system to summarize the police report and identify possible defenses. That may produce something useful, but it also asks too much at once. The system has to extract facts, interpret those facts, apply law, and suggest strategy in a single step. A loop can be designed more carefully.
First, the system reviews only the uploaded discovery and creates a document inventory.
Second, it prepares a chronology with source references for each material event.
Third, it separates officer observations from officer conclusions. “The driver had bloodshot eyes” is an observation. “The driver was intoxicated” is a conclusion.
Fourth, it compares the police report against body-camera timestamps, if transcripts or summaries are available.
Fifth, it identifies missing materials, such as dispatch audio, booking video, calibration records, implied consent paperwork, search warrant materials, or chemical-test documentation.
Sixth, it flags factual conflicts and unanswered questions for lawyer review.
Seventh, it prepares a preliminary issue list, but labels each issue as verified, partially supported, or requiring further review.
Then it stops.
That sample workflow, while exceedingly basic, does not decide whether to file a motion. It does not tell the client what to do. It does not replace the lawyer’s analysis of reasonable suspicion, probable cause, custody, consent, admissibility, or prejudice.
Instead, it prepares the ground for the lawyer to do all of those things, much the way a lawyer and paralegal may work hand in hand. That is a better use of AI because the work is structured, bounded, and capable of being checked against the file.
The Criminal Defense Problem: Preserving Uncertainty
Criminal defense work often turns on messy facts. Police reports may omit key details. Video may contradict written summaries. Clients may remember events imperfectly. Witnesses may change their accounts. Legal arguments may depend on timing, location, consent, custody, reasonable suspicion, probable cause, preservation, waiver, and local judicial practice.
An AI loop may help organize that complexity. It may also flatten it. That risk matters because the defense lawyer’s job is not to make the state’s story coherent and it turn it to the advantage of the client, or better yet, fashion a more compelling but different case narrative. The goal is often to show where the State’s story is incomplete, unreliable, exaggerated, or unproven. A loop that turns messy facts into a tidy narrative may work against the defense if it erases ambiguity that should be preserved.
That is why the lawyer must instruct the system not only to summarize, but to doubt. The system should be told to identify conflicts. It should be told to preserve unknowns. It should be told to flag where a conclusion depends on a police characterization rather than an observable fact. It should be told to distinguish what a witness said from what the witness could actually know.
Those are not technical preferences. They are defense-lawyering preferences.
When Revision Makes Error Look Better
Recursive systems create a particular kind of confidence problem.
When an AI system drafts, critiques, and revises its own work, the final output may appear more mature than the first answer. Often it will be. But the improvement may be stylistic rather than substantive. The system may become clearer without becoming more correct. That principle should be intuitive to lawyers.
A lawyer who drafts a motion and immediately reviews it may improve it. But the lawyer may also miss the same flaw twice. That is why good legal work often benefits from fresh eyes, independent source checking, adversarial review, and actual verification.
AI is no different. Self-critique can help. Independent grounding is better.
For legal work, that means source documents, court rules, statutes, cases, transcripts, exhibits, body camera footage, expert reports, and human review. The loop should be connected to the record, not merely to its own prior answer.
Designing AI Workflows Around Legal Judgment
Lawyers do not need to treat prompting as a stand-alone professional identity. They need to become better supervisors of AI-assisted work. That requires attention to workflow design, not just word choice.
The lawyer must decide what the system may access. The lawyer must decide what it may do. The lawyer must decide what sources control. The lawyer must decide when the system must stop. The lawyer must decide what cannot leave the office. The lawyer must decide what requires verification before use.
In summary, a prompt asks for an output. A loop creates a process. A lawyer supervises the process.
That distinction should guide how firms adopt these tools. The most important question is not whether the system can produce something impressive. The question is whether the lawyer can understand, test, and control how the work was produced.
The Lawyer Still Owns the Judgment
Prompting is not dead. But the version of prompting that dominated the first stage of AI adoption is changing if not fading. The future is not a lawyer sitting in front of a blank box trying to find the perfect instruction. The future is a lawyer using systems that draft, critique, revise, retrieve, compare, and escalate uncertain issues for human review.
Yes, that future, as long as it lasts, is useful, but it is also less transparent. Lawyers should not respond with either panic in the face of accelerating change, or blind uninformed and undisciplined adoption. The better response is careful and circumspect adoption. Build loops that improve drafts. Require grounding. Preserve uncertainty. Demand citations. Keep humans in charge of legal conclusions. Treat AI critique as useful, but never as final.
What is dying is the idea that a single prompt is the center of legal AI practice. The loop is becoming the center. And in law, the loop must answer to the lawyer.
Next: Building Legal AI Loops in Practice
This article explains why the shift from prompts to loops matters. The next question is practical: what does a good legal AI loop actually look like?
In Part Two, I will walk through several examples lawyers can use in ordinary practice, including loops for improving drafts, reviewing discovery, preparing cross-examination, and testing arguments against the other side’s likely response. The goal is not to make AI autonomous. The goal is to make AI-assisted work more structured, more skeptical, and easier for the lawyer to supervise.
About the Author
Patrick T. Barone is the founder of Barone Defense Firm and a criminal defense attorney focused on DUI and criminal defense litigation. He writes and speaks about the practical use of artificial intelligence in criminal defense practice, including how lawyers can use AI tools while preserving professional judgment, confidentiality, and ethical responsibility.
To learn more about Patrick Barone’s AI-supercharged criminal defense practice, visit Barone Defense Firm.
For Further Reading
The Lawyer’s Guide to Agentic AI
A practical introduction to agentic AI for lawyers who want to understand agents without becoming technologists.
From Discovery Dump to Defense Theory: Using AI to Find the Case Inside the Case
A companion piece on using AI to identify conflicts, omissions, assumptions, and unanswered questions inside criminal discovery.
The First Prompt Is Where Most Legal AI Errors Begin
A useful earlier treatment of prompt design, now placed in context by the shift from prompts to loops.
Using AI Like a Lawyer Requires Thinking Like Opposing Counsel
A related article on adversarial prompting and the need to test legal work against the other side’s argument.
The AI Policy in Your Firm May Not Be Enough
A broader discussion of why lawyers need working systems for confidentiality, verification, supervision, billing, and client communication.
To learn more about Barone’s AI supercharged criminal defense law practice, visit the firm’s website.