How Solo Lawyers Use AI for Demand Letters Without Bar Discipline

A solo employment lawyer in Ohio billed a demand letter at $850 last month. Drafting time: 3.5 hours. The first paragraph was written by Claude. The middle was edited by the lawyer. The citations were verified by hand against Westlaw. The signature block was hers. The malpractice exposure on that letter is the same as if she had typed every word herself — which is exactly the point. ABA Resolution 512 (August 2024) did not invent a new standard for AI-assisted work. It applied the existing standard — competence, confidentiality, supervision, candor — to a new tool. The line did not move. The terrain underneath it did.

Most solo and small-firm lawyers I have talked to in 2026 do not have a written AI use policy. They are using ChatGPT, Claude, or CoCounsel in some form — drafting, research, summarization — but the policy lives in their head, which means it does not exist in any way a bar disciplinary committee would recognize. That is the disciplinary risk. Not the use of AI. The absence of a policy around it.

This article documents what the ABA published, what fourteen state bars have published on top of it through May 2026, and the safe workflow for using AI to draft demand letters specifically. Demand letters are the right tool to talk about because they are high-volume, formulaic enough to benefit from AI drafting, and consequential enough to surface every supervisory and confidentiality issue in one document.

This is documentation of industry guidance. It is not legal advice. If you need an opinion on your specific situation, talk to your state bar's ethics hotline or a malpractice attorney licensed in your jurisdiction.

The floor: ABA Resolution 512 and Rule 1.1 Comment 8

ABA Formal Opinion 512, released July 29, 2024 and formally adopted August 2024, is the document every state bar has been citing for the last twenty months. It is twelve pages. Read it. The short version: existing Model Rules apply to generative AI use, and there are six explicit duties a lawyer using AI must satisfy.

The six duties, in the ABA's order:

• Competence (Rule 1.1) — A lawyer who uses AI must understand the benefits and risks of the specific tool. Comment 8 to Rule 1.1, adopted in 2012, already required competence in "the benefits and risks associated with relevant technology." Opinion 512 cites Comment 8 as the controlling text and adds that "competence" now extends to understanding generative AI's tendency to hallucinate, its training data limitations, and its confidentiality properties.
• Confidentiality (Rule 1.6) — A lawyer must not input client-confidential information into an AI tool unless the tool's data handling has been vetted. Public consumer chatbots that train on user inputs are a 1.6 violation when fed client confidences. Enterprise tools with zero-retention contracts are usually acceptable.
• Communication (Rule 1.4) — There is a duty to communicate "material" AI use to the client. Opinion 512 explicitly does not require disclosure of every AI use, but does require disclosure when the AI output is being relied on in a way the client would reasonably want to know about.
• Candor to the tribunal (Rule 3.3) — A lawyer must verify AI-generated content before filing. This is the rule Mata v. Avianca made famous. The lawyer who filed six hallucinated citations from ChatGPT was sanctioned not because he used ChatGPT but because he did not verify the output. The duty pre-existed the tool.
• Supervisory duties (Rule 5.1 and 5.3) — Partners and managing attorneys must establish policies for how associates and non-lawyer staff use AI. Rule 5.3 explicitly extends supervisory duties to non-lawyer assistants, which Opinion 512 reads as covering AI tools themselves: the lawyer is responsible for the AI's output the same way they are responsible for a paralegal's work product.
• Fees (Rule 1.5) — A lawyer cannot bill for time the AI did that the lawyer would not have done. If a paralegal would have spent an hour summarizing depositions and the AI did it in three minutes, you bill the three minutes plus the lawyer's review time, not the hour.

That is the floor. Every state bar opinion published since stacks on top of it.

What the state bars added (May 2026 status)

By my count fourteen state bars have published formal opinions, ethics committee guidance, or court rules specifically addressing AI use since Opinion 512 dropped. The pattern: no state has prohibited AI use. Every state has imposed verification and confidentiality duties. Several have added supervisory or written-policy expectations. None require pre-clearance from the client for every use. None require disclosure of every AI tool touched.

California led on confidentiality — its November 2023 guidance is explicit that "free, public AI tools" generally cannot be used with client-confidential information without anonymization. Florida Bar Opinion 24-1 was first to formally address generative AI; it permits use with proper supervision and confidentiality protections and flagged the fee issue earlier than 512 did. Texas Opinion 705 (March 2025) explicitly addresses citation verification and the Mata v. Avianca fact pattern. Virginia LEO 1901 hammers the fee issue from 512 — a lawyer cannot bill for time the AI saved them without client consent to a value-based fee arrangement.

What "discipline" actually looks like in 2026

The discipline cases that have hit so far are not for using AI. They are for the four things you would expect:

• Filing hallucinated citations. Mata v. Avianca was the first. Steven Schwartz at Levidow Levidow & Oberman was sanctioned $5,000 in June 2023 for filing six fabricated cases generated by ChatGPT. By May 2026, Reuters has tracked over 200 documented instances.
• Confidentiality breaches. At least two malpractice carriers have flagged AI-related claims in their 2025 trend reports — both involving lawyers who pasted client communications into ChatGPT.
• Fee disputes. A handful of state bar fee-arbitration panels have reduced fees where the lawyer billed hourly for work the AI did in minutes.
• Misrepresentation about AI use. A New York lawyer in 2025 was disciplined for telling a client her work was "personally drafted" when in fact most of the brief came from CoCounsel.

The lesson: the floor stays the same. Lawyers get disciplined for lying, for failing to supervise, for breaching confidentiality, for failing to verify. AI is the new instrumentality. The duties are not new.

The safe workflow for AI-assisted demand letters

Here is the workflow that stays inside the line. Every step ties to a specific rule. Demand letters happen to be the cleanest example because they live entirely outside the tribunal and yet are consequential enough to need full Rule 1.6 confidentiality treatment.

Step 1 — Vet the tool (Rule 1.1, competence)

Before any client work touches the AI, the lawyer reads the tool's data-handling policy. The bar is "zero retention for training" and "encryption in transit and at rest." Public consumer chatbots default-fail this bar. Enterprise tools like CoCounsel, Spellbook, Harvey, Ascero's tools, and the enterprise versions of ChatGPT and Claude generally pass. If the tool's data-handling page is vague or absent, do not use it for client work.

Step 2 — Strip and prompt (Rule 1.6, confidentiality)

Even on a vetted tool, the safer pattern is to draft from anonymized facts. Replace client names with placeholders, replace company names with generic terms, replace addresses with city-only references. Prompt the AI with: "Draft an employment demand letter for a wrongful termination claim. Employee: 5 years tenure, terminated after FMLA leave. Damages: lost wages, emotional distress. Tone: professional, firm, not litigious." That prompt contains no confidential information.

Step 3 — Review (Rule 5.3, supervision)

Treat the AI output the way you would treat a first-year associate's draft. Read every paragraph. Question every assertion. Mark every claim that needs a source. The AI does not know your client. It does not know your jurisdiction's specific elements of the claim.

Step 4 — Verify every citation (Rule 3.3 by analogy, candor)

Demand letters do not go to a tribunal, so Rule 3.3 does not strictly apply, but the underlying duty of accuracy does. If the AI cites a statute, look up the statute. If the AI cites a case, pull the case from Westlaw or Lexis and confirm it exists and holds what the AI says it holds. The Mata v. Avianca failure mode applies to demand letters too.

Step 5 — Cite source check (Rule 1.1, competence)

Beyond verification, check that the cited authorities are still good law. AI tools cannot reliably tell you a case has been overruled, distinguished, or limited in your jurisdiction since the model's training cutoff. Shepardize. KeyCite.

Step 6 — Confidentiality protections at the seam (Rule 1.6 again)

When you re-insert client names and specific facts into the draft, do that in your word processor, not in the AI tool. Save the working document on your firm's encrypted file system, not in the AI's chat history.

Step 7 — The signature block is yours

This is the rule everything else flows from. The letter goes out under your bar number, your address, your professional responsibility. The AI assisted. You are the lawyer.

Anti-pattern: do not paste opposing counsel emails into ChatGPT

This is the single most common confidentiality mistake I have seen working with solo and small-firm lawyers in 2026. Opposing counsel sends a settlement demand. The lawyer pastes the email into ChatGPT and asks "summarize the strongest arguments and weakest arguments." The lawyer does not realize that the email contained client-confidential information about the underlying matter that has now been processed by a tool whose data handling has never been vetted. This is a Rule 1.6 violation.

Cost benchmark: traditional vs AI-assisted demand letter

Traditional drafting: solo employment lawyer at $250-$450/hr, 2-4 hours per letter, $500-$1,800 all-in to the client. AI-assisted drafting with the seven-step workflow: 30-60 minutes of lawyer time, $125-$450 to the client, $50-$250/mo tool cost. The fee compression is real. The compliance question is whether the lawyer bills the actual time or maintains the prior hourly bill. Virginia LEO 1901 and ABA 512 both come down hard on the latter.

The honest economics: AI-assisted drafting at $400/letter is more profitable than traditional drafting at $850/letter because the lawyer can do four letters in the time the old workflow did one. The margin compounds across volume, not per-letter price.

What this means for solo and small-firm lawyers right now

Three things you should do this month:

• Write the policy. A one-page written AI use policy for your firm, even if you are a one-lawyer shop. Cover: which tools are approved, what gets anonymized, what does not get pasted into AI, who verifies citations, how time is billed when AI assists.
• Vet your tools. If you are using a public consumer AI tool with any client information, stop today. Either upgrade to the enterprise version with a zero-retention contract, switch to a legal-specific tool (CoCounsel, Spellbook, Harvey, Ascero), or stop using AI for client work until you have done the vetting.
• Run the audit. The ABA 512 audit is a one-time pass against the six duties. It produces a written record of what you do and how you comply.

The discipline risk in 2026 is not "I used AI." The discipline risk is "I used AI without a written policy, without a vetted tool, without verification, and without supervision." The fix to all four is administrative, cheap, and one-time.