How to Use AI as a Lawyer

A first-person account from a solo practitioner client of The Agent Almanac. Identifying details have been changed; the workflows, tools, and hours are real.

The Sunday I realized the practice owned me

I'm a solo attorney — estate planning and small business work, twenty-two years in. For most of those years I told people I loved being a lawyer, and it was roughly true. What I didn't say was that the practice had quietly eaten everything else.

The moment I can point to: a Sunday afternoon two years ago, standing in my garage workshop. My grandfather was a cabinetmaker, and his hand planes hang on my wall — I'd planned to spend my life using them. That Sunday I noticed the dust on the workbench was thick enough to write in, did the math, and realized I hadn't built anything in almost three years. I was billing well. I was also drafting engagement letters at 9pm and writing client update emails on Saturday mornings, and the woodshop had become a museum.

I'm telling you this up front because every article about how to use AI as a lawyer starts with productivity statistics, and the statistics are real — but the reason to do this isn't throughput. It's whatever your version of the dusty workbench is. Mine is why I started; what follows is exactly how.

How to start using AI as a lawyer: my first 30 days

I'm constitutionally cautious — estate lawyers usually are — so I started where a mistake would cost me an eye-roll, not a malpractice claim: administrative work that never touches legal judgment.

The setup took one evening. A business-tier AI assistant (the business tier matters — it carries a contractual commitment not to train on your inputs, which is the line Rule 1.6 cares about), and five saved "master prompts," one for each document I write constantly:

• Client correspondence first drafts
• Meeting and call summaries from my dictated notes
• Engagement letter drafts from a matter intake sheet
• Time-entry narratives from my shorthand ("rev tr docs, tc w/ client re: funding" becomes a clean billing entry)
• The plain-English explainer — trust funding instructions, probate timelines — written at the reading level my clients actually have

Each prompt got trained on two or three anonymized examples of my past writing, so the drafts sounded like me rather than like a press release. The rule I set on day one and never broke: nothing AI-drafted leaves my office without my eyes on every line.

Honest first-month report: the drafts were 80% right, my review habit got sharp, and I recovered about three hours a week from this tier alone — because admin, it turns out, is where a solo's evenings actually go. More importantly, the review-everything reflex became muscle memory before I trusted AI with anything that mattered more.

Climbing the ladder: drafting from my own precedents

Month two, I moved to document drafting — with a technique that changed my whole understanding of how to use AI as a lawyer safely: grounding.

I never ask AI to "write a trust." That invites it to improvise law, which is exactly the failure mode that gets lawyers sanctioned. Instead, I built a precedent bank — my own tested documents, the ones I'd refined over twenty years — and the AI's job became adaptation: take my precedent, take this matter's intake facts, produce a conformed draft, and flag every single change it made.

Tools that only work from documents you give them are built for this — NotebookLM is the free one I use for synthesizing client document sets too (bank statements, prior wills, the shoebox of papers every estate client brings). For contract work inside Word, Spellbook does the same thing against your playbook for about £130 a month — the realistic solo alternative to enterprise platforms like Harvey, Kira, and Luminance, which are built and priced for big-firm document volume, not a one-lawyer shop. The substance stays mine; the AI does the assembly that used to take an associate I don't have.

What it feels like in practice: a fast junior who has read everything, types instantly, and has no judgment whatsoever. You'd never let that person file anything unreviewed. You'd also never give them up.

Research, and the protocol that let me sleep

This was the tier that scared me, because I'd read the sanctions cases — lawyers filing briefs citing cases that never existed. So before I let AI near research, I wrote myself a verification protocol and taped it to my monitor:

With that protocol in place, the grounded research tools — CoCounsel inside the Westlaw ecosystem (Thomson Reuters), Lexis+ AI inside Lexis — became genuine accelerants, because every answer links to real, checkable authority. First-pass research that took an afternoon takes forty minutes plus verification. The protocol adds ten of those minutes back, and it is the ten minutes that separates "AI-assisted lawyer" from "cautionary tale in the bar journal."

The agentic leap: workflows that run while I'm in the workshop

Everything to this point was AI as a very fast assistant — me driving every step. The transformation, the part that actually un-dusted the workbench, came when I learned what agentic AI means for a law practice: instead of asking AI to do a task, you build a workflow that runs a sequence of tasks from a single trigger, and your role collapses to review-and-approve.

Three agentic workflows now run my practice's repetitive layer:

• The intake agent. A prospective client completes my structured questionnaire. That submission triggers the chain: a conflict-check summary against my matter list, a matter memo organizing their facts, a draft engagement letter from the right template, and a plain-English "what happens next" email — all waiting in a folder for my review. What used to be ninety minutes of setup per new matter is now twelve minutes of reading and approving.
• The Friday client-update agent. Every active matter's notes get compiled, and the agent drafts a personalized status update per client, flagging anything overdue. I review the batch over one coffee. In two years, no client has gone silent-treatment quiet on me — and in estate work, "he always kept us informed" is the sentence that produces referrals.
• The document-summary agent. New documents arriving in a matter folder get summarized, key dates extracted to my calendar, and follow-up questions drafted. The shoebox of papers gets triaged before I open it.

None of this required me to become a programmer. The chains run on no-code automation connecting my forms, files, and the AI — the same pattern any solo can build.

The rules I never bend

Two years in, my ethics framework has hardened into four lines:

1. Rule 1.6 is the gate. Client-identifying and privileged information enters only tools with contractual no-training commitments — business tiers, legal-specific platforms — or goes in anonymized. The consumer chatbot on my phone has never seen a client's name.
2. Rule 1.1 cuts both ways. Competence now includes understanding these tools — ABA Formal Opinion 512 spelled it out — and it has always included not filing their mistakes. The verification protocol is Rule 1.1 in action.
3. I am the author of everything. AI drafts; I decide. Every output is reviewed as if a junior wrote it, because functionally one did.
4. My engagement letter says so. One paragraph disclosing that I use AI tools under my review, with client data protections. Written disclosure has prompted exactly zero client objections and several appreciative questions.

What changed: the numbers, and the workbench

The accounting, two years on:

Nine hours a week tracks with the published data — Thomson Reuters reports lawyers using generative AI save up to 260 hours a year, about 32 working days, and the Wolters Kluwer 2026 survey found 92% of legal professionals now using AI tools, with most saving 6–20% of their week. I'm at the high end because I built workflows instead of poking at a chatbot, and that — not the tools — is the whole difference.

But the number that matters isn't in the table. Last fall I finished a walnut blanket chest for my daughter — mortise and tenon, hand-cut dovetails, my grandfather's planes doing the finishing. It took me eleven Sundays. Two years ago I didn't have one.

That's how to use AI as a lawyer, as far as I'm concerned: not to bill more hours, though you can — but to take back the ones that were never supposed to belong to the practice.