ChatGPT for Solo Practitioners: Where AI Helps, Where It Hurts Your Liability
If you are running a solo practice, you already know the math. Sixty-plus hour weeks. No paralegal. No associate. Every administrative minute is a minute you cannot bill, and every billable minute carries the cost of the unbilled minute that paid for it.
ChatGPT looks like the answer to that math. In some places, it is. In other places, it is the exact opposite — a tool that creates more liability exposure than the time it saves is worth.
This article is a Gap analysis: where AI genuinely helps a solo practice, where it actively hurts, and what protocols separate the two. This is the Gap (G) level of the GROWT Method — you cannot make smart decisions about where to operationalize AI until you have a clear picture of what the assessment looks like for your own practice.
The Solo Practitioner's Reality
The reason this matters more for solos than for firms with infrastructure is straightforward. A 50-lawyer firm has a conflicts department, a knowledge management team, a CLE program, and an MLP carrier with sophisticated guidance on AI use. You have you, a laptop, and a malpractice policy you renewed last year without reading the new AI exclusions.
When a solo practitioner uses AI badly, there is no second line of defense. When the same lawyer uses it well, there is no team meeting to slow the workflow down. The upside is bigger and the downside is worse. Knowing which side of that line you are standing on, prompt by prompt, is the whole skill.
The frame for the rest of this piece is simple: green light tasks, red flag tasks, and what protocols turn red flags into manageable risks.
Green Light: Where AI Wins for Solos
These are tasks where the upside is real, the downside is bounded, and the verification cost is low. Use AI here without hesitation.
Research Summaries
You have a 47-page opinion you need to understand before a 4 PM call. Paste it in. Ask for a 300-word summary plus the three most important holdings, plus the procedural posture. You verify by skimming the opinion against the summary — 10 minutes instead of 90.
This works because you are providing the source text. The model is not generating facts; it is compressing facts you gave it. Hallucination risk is near zero.
Document Drafting Starter Copy
Engagement letters, basic contracts, demand letters, status updates to clients — these are templated drafts where AI gets you from blank page to 80% in two minutes. You then spend 15 minutes on the 20% that actually matters for this specific matter.
Net time savings: 30-45 minutes per document. At solo rates, that is real money over a month.
Client Communication First Drafts
Difficult emails to clients are the worst kind of time sink — you procrastinate writing them, then spend 40 minutes on three paragraphs. AI drafts the structure and tone calibration in 60 seconds. You adjust, send, move on.
A working prompt structure:
"Draft a professional but warm email to a client explaining [situation]. Tone: empathetic, plain English, no legal jargon. Length: 200 words. Avoid promising specific outcomes. The client is anxious about [concern]."
The output is rarely perfect on the first pass. It is almost always 60-70% there, which is faster than starting fresh.
Marketing Content
Blog posts, website copy, LinkedIn updates, newsletter drafts — these are pure upside. There is no privilege concern, no citation accuracy requirement, no client matter content. AI is essentially a junior marketer working for you for free.
The pattern: outline the topic in a few bullet points, let the model produce a first draft, then edit in your voice. A practice that publishes one piece a week using this workflow is doing roughly 2 hours of marketing work instead of 8, and the marketing engine is the lifeblood of a solo practice.
Red Flag: Where AI Hurts
These are the categories where solo practitioners are getting into trouble. Each one has a real-world example that has cost lawyers money, reputation, or licenses.
Hallucinated Citations (the Mata v. Avianca Lesson)
In 2023, two lawyers were sanctioned by a federal court for filing a brief containing six citations that did not exist. They had used ChatGPT to do the research and had not verified the cases. The model had fabricated names, docket numbers, and quotes.
The sanctions order made the New York Times. The lawyers' names are now permanently attached to the most quoted AI-legal cautionary tale in the profession.
The mistake is so simple it sounds insulting to repeat: they trusted the model on citations. Every solo who uses AI for research has the same temptation, especially under deadline pressure. The discipline that prevents it is non-negotiable: never accept a citation that AI produces without independent verification in Westlaw, Lexis, Bloomberg Law, or the relevant official source.
Subsequent cases have repeated the pattern in multiple US districts and a small but growing number of UK matters. The volume of incidents is rising, not falling.
Privilege Leakage via Public LLM Use
When you paste matter content into the consumer version of ChatGPT (or any LLM without an enterprise data-use agreement), you may be transmitting client information to a third party in a way that compromises privilege and confidentiality obligations.
The mechanics vary by provider and by what plan you are on, but the principle is simple: if your data can be retained, used for training, or accessed by humans for review, you have a problem.
Solo practitioner protocol:
- For matter content: use only tools with a confirmed no-training, no-retention data-use agreement, or use a local model.
- For non-matter content (research on public law, general drafting, marketing): consumer tools are fine.
- When in doubt about whether something is matter content: assume it is, and treat it accordingly.
Jurisdiction-Specific Advice Generation
Models will confidently produce text that reads like legal advice tailored to your jurisdiction. The text will sound right. Some of it will be right. Some of it will be wrong — either because the model is reasoning from data that includes other jurisdictions, or because the law has changed since the model's training cutoff, or because the model is filling in plausible-sounding text where it has no actual data.
You cannot tell which is which without verifying every substantive claim against current jurisdictional authority. That verification cost often exceeds the time savings.
Rule of thumb: AI is for drafting and structure, not for legal conclusions. The conclusion is yours, signed under your bar number, verified against current law.
Billing Time AI Could Not Have Spent
The ethics rule is simple and universal: you bill for the time you spent, not the time the task would have taken if you had done it manually.
If a brief that previously took 8 hours now takes 3 hours because AI did the first draft, you bill 3 hours. Billing the original 8 is fraud. Bar associations are now treating this as a category of billing fraud worth specific guidance, and clients are starting to ask the question directly in fee disputes.
The fix is straightforward: track actual time spent, including AI-assisted time, and bill actuals.
The Real Costs of Getting It Wrong
To make the risk side concrete, here are the categories of cost solos have actually incurred:
Sanctions for hallucinated citations. Court-ordered fines, mandatory CLE on AI ethics, public reprimand on the court's docket, professional embarrassment that surfaces every time someone Googles your name. The financial cost is bounded; the reputation cost is not.
Malpractice exposure. If AI-assisted work leads to an outcome a client can show was avoidable with reasonable competence, you are looking at a malpractice claim. Some carriers are now asking AI-use questions on annual renewals. Some are adding endorsements that condition coverage on documented protocols.
ABA Rule 1.1 (technology competence) issues. Rule 1.1 says competence includes "the benefits and risks associated with relevant technology." Most US states have adopted some version. Failing to maintain reasonable technological competence — including understanding AI's risks — is itself a competence problem. UK SRA principles are similar in substance.
Privilege waiver risk. If you transmitted client information to a tool in a way that constitutes disclosure to a third party, opposing counsel will eventually figure it out. Privilege waiver in the middle of a matter is a very expensive problem.
Risk-Minimization Protocols
These protocols turn most red-flag tasks into manageable ones. Adopt them as defaults, not exceptions.
PII Stripping
Before any prompt that touches matter content, redact: client names, opposing party names, addresses, dates that uniquely identify the matter, account numbers, case numbers, identifying medical or financial details. Replace with bracketed placeholders ("[CLIENT]", "[OPPOSING PARTY]", "[DATE]").
This takes 60 seconds per prompt. It eliminates a category of privilege risk.
Prompt Logs
Maintain a simple log of every prompt that involves matter work: matter ID, date, model used, prompt text (or a description), output use (verified citations? edited and sent? discarded?), and any verification steps taken.
A spreadsheet is enough. The value is twofold: it forces the discipline of verification, and it gives you a defensible record if a question arises later.
Verification Rules
The shortest version of the rule that prevents the most damage: nothing produced by AI goes into a filing, a client deliverable, or substantive advice without independent verification of every factual claim and every legal authority.
"Verification" means: cites confirmed in an authoritative database, factual claims confirmed against documentary evidence or your own contemporaneous knowledge, jurisdictional law confirmed against current statutes and recent case law.
Tool Selection
For matter content: only tools with a confirmed data-use agreement that includes no training on your data, retention limits you are comfortable with, and access controls that match your confidentiality obligations.
For non-matter content: consumer tools are fine and often better, because they have the most current models.
Maintain a clear line between the two. Mixing them is how privilege leaks happen.
Where to Start
The honest answer for most solo practitioners: start with green-light tasks only. Marketing content. Research summaries on public law. Engagement letter drafts. Build the workflow muscle on tasks where mistakes cost you nothing.
Once you have run AI on those tasks for 30 days, you will know what verification overhead looks like in your practice. At that point — and only at that point — start carefully expanding into matter-adjacent work with the full protocols in place.
The solos who get into trouble are the ones who skip the green-light phase and go straight to drafting motions. The solos who get real ROI are the ones who treat AI adoption as a multi-month process with a clear assessment phase first.
That assessment phase is the Gap level of GROWT. It is where the work pays off.
FAQ
Q: Is consumer ChatGPT safe for any kind of legal work? A: For non-matter content — research on public statutes and case law where you are not pasting client information, marketing content, general drafting templates — yes. For anything involving client identifiers or privileged information, no. The rule is about what data leaves your control, not which tool you use.
Q: My malpractice policy does not mention AI. Am I covered? A: Probably, for now. Carriers are moving quickly to add language. Read your next renewal carefully, ask your broker specifically about AI exclusions and conditions, and document your AI-use protocols so you have something to point to if a question arises.
Q: Can a client refuse to let me use AI on their matter? A: Yes, and they sometimes do. If a client raises the question, document the conversation and respect their preference. If you have a client who has not asked, you do not have an obligation to volunteer it for routine tasks, but you may have a disclosure obligation for novel or material AI uses depending on your jurisdiction.
Q: What about local LLMs running on my own machine? A: This is the cleanest solution for matter content. Locally hosted models with no external data transmission eliminate the privilege transmission concern entirely. The setup cost has dropped sharply; capable models now run on a current MacBook Pro.
Q: How do I know if a citation is real? A: Look it up in Westlaw, Lexis, Bloomberg Law, or the relevant official source (court website, legislation.gov.uk). If the citation does not appear in an authoritative database, it does not exist. Treat the model as if every cite it produces is fabricated until you have personally confirmed otherwise.
Q: Will AI eventually be good enough that these protocols are unnecessary? A: The protocols will change in detail but not in principle. The lawyer is responsible for the work product. That principle predates AI and will outlast it.
Q: What is the single most useful AI workflow for a solo to adopt first? A: First-draft client communications. The verification cost is essentially zero (you are sending it under your name, you re-read it before sending anyway), the time savings are real, and there is no privilege exposure if you keep the content general.
Disclaimer
This article is general guidance, not legal advice. AI use in legal practice carries jurisdictional and professional-responsibility implications — consult your bar association rules.
Build Your AI Plan
The gap between "AI sounds useful" and "AI is wired into my workflow safely" is mostly assessment work. Where do you actually spend your time? Which of those tasks are green-light? Which need protocols before you touch them?
Or explore the GROWT Method to see how Gap analysis leads into Roadmap, Operationalize, Win, and Transform.