Agentic AI in Solo Practice: Beyond the Chat Window

Most lawyers who tried ChatGPT have either incorporated it into their work or set it aside as not useful. The more interesting question now is the layer above the chat window. What happens when a lawyer stops typing prompts one at a time and runs structured agentic workflows that complete multi-step legal work end to end.

I run an agentic workflow myself. I am a non-technical lawyer with no coding background. Hundreds of hours of hands-on building have gotten me to a working multi-agent setup that handles real legal work. The notes below are what I have learned getting there.

The simple definition

An agent is a model that can take an instruction, plan a sequence of steps, use tools to complete those steps, check its own work, and return a finished deliverable. A workflow is a defined recipe of agent calls that runs the same way every time you trigger it.

A chat is one prompt at a time. An agentic workflow is a procedure.

For a lawyer, the difference is the difference between “can you summarize this deposition” asked five times for five depositions, versus “summarize every deposition in this matter, flag inconsistencies across witnesses, and produce a chronology” run once and returned as a finished file.

Where agents earn their place

Three categories where I have found the leverage to be real.

Discovery review. An agentic workflow can read a production set, classify documents by relevance and privilege, summarize the responsive ones, and produce a privilege log draft. The first review pass is what the agent does. The second pass is what the lawyer does. The lawyer's pass goes faster because the agent has organized the material.

Witness preparation. For every witness in a case, the agent pulls every document mentioning the witness, every prior statement the witness has made, every reference to the witness in opposing counsel's filings. The output is a witness file the lawyer can read in an hour instead of building over a week.

Cross-examination outlining. From a witness file, the agent identifies the points of impeachment, drafts question sequences, and flags safe-harbor topics that the witness will dodge. The lawyer revises, prunes, and runs the cross. The thinking work is still the lawyer's. The setup work is the agent's.

These are not theoretical for me. The tools required are not exotic. They are commercial AI products configured carefully and connected to the case file. The work is in the configuration and the verification, not in any custom code.

What it takes to build a workflow that holds up

Three requirements.

A clean source of truth. The agent needs to know where the case material lives. If your files are scattered across email, a hard drive, a cloud folder, and a paper box, the agent cannot help. The first investment is organizing the case file in a single, structured location.

Verification at every step. The agent will produce work that looks polished and contains errors. Every output needs a verification step. For some outputs the verification is automated, the agent checking its own work against a primary source. For others the verification is the lawyer reading carefully before relying on the output. The verification is non-negotiable.

A defined deliverable. Vague instructions produce vague outputs. The workflow needs to specify what the deliverable is, what format it takes, what the lawyer will use it for, and what counts as success. Spend time on this before you spend time on the technical setup.

The technical setup matters less than people think. The defined workflow matters more than people think.

Where agents are not ready

Two categories where I do not run agentic workflows.

Live courtroom work. Cross-examination, opening, closing, the live judgment in front of a jury. The agent cannot do this. The agent can prepare materials. The lawyer does the work in the room.

Strategic decisions. Settle or try. Plead or fight. Take this deal or counter. These decisions integrate facts the agent does not know, including the lawyer's read of the client, the opposing counsel, the judge, and the jury pool. The agent can lay out options. The lawyer makes the call.

The line between assistance and decision, the same line the courts are drawing for their own use of AI, is the same line a solo should draw for client work. The agent assists. The lawyer decides.

The cost and the time

The sticker price of the tools is small relative to the time the work saves once a workflow is built. The hidden cost is the setup time. The first version always has gaps. The fifth version is usually the one I trust. That iteration takes weeks, not hours.

The compounding is what matters. One reliable workflow is useful. Three or four together change what a single person can carry.

The professional discipline

Agentic AI does not change any professional duty. Competence, supervision, confidentiality, candor to the tribunal. The duties apply the same way they always have. The instruments change. The obligations do not.

The supervision duty in particular is becoming more demanding. A lawyer running an agentic workflow is supervising a process that produces work product without the lawyer's moment-to-moment attention. The supervision is built into the workflow design. The verification step is the supervision in operational form.

If a workflow does not have a verification step, the lawyer running it is not supervising it. That is a problem.

The technician trap

Most lawyers are good technicians and bad at the business side of practice. The marketing, the content, the intake systems, the follow-up, the knowledge management. Every billable hour goes to client work and the business side stays neglected. Michael Gerber wrote about that trap forty years ago in The E-Myth. The advice was famous. Almost nobody could implement it because the team it required was not affordable at solo scale.

Agentic AI is the first tool I have seen that changes that math. The roles that used to require a payroll a solo could not carry can now be carried by a structured set of agents running in the background. That is the part of this technology that interests me most.

The result is not a robot lawyer. It is a working lawyer with the operational support larger firms have always had.

Where this goes

The next 24 months are going to bring a wave of legal-specific agentic platforms. Some will be built by the foundation-model companies. Some by the established legal-tech vendors. Some by lawyers who built tools for their own practice and decided to share them.

My own bet is that the curve keeps moving and that going early, even as a non-technical person, beats waiting for the market to settle.

The discipline is verification-first, supervision-always, and a defined deliverable for every workflow. The tools are new. The discipline is not.

Frequently Asked Questions

What is the difference between using ChatGPT and running an agentic workflow?

A chat is one prompt at a time, run by hand. An agentic workflow is a defined recipe that runs the same way every time you trigger it, completing multi-step work end to end and returning a finished deliverable. The workflow scales. The chat does not.

What is the first agentic workflow a solo lawyer should build?

Whichever one removes the most repetitive work from your week. For litigators, deposition summaries and witness file building are common starting points. For transactional lawyers, contract comparison and due-diligence checklists. Pick one. Build it on a test matter. Iterate until it is reliable. Then deploy.

How long does it take to build an agentic workflow?

Plan on several weeks of iteration before the output is reliable enough for real client work. The first version always has gaps. The fifth version is usually the one you trust. Skip the iteration phase and you will ship errors.

Is running agentic workflows on client matters ethical?

Yes, if you supervise the work and verify the output. The duty of competence requires basic understanding of the tools. The duty of supervision requires verification at every step. The duty of confidentiality requires careful choice of which AI vendors handle which data. Read your state's most recent ethics guidance on AI before you start.

Can a solo really compete with a big firm using agentic workflows?

On the work agentic workflows are good at, yes. On trial work and high-judgment work, the playing field has not changed. The solo who builds three or four reliable workflows handles a higher case volume at lower cost than a solo who does not. That is real competitive ground.