What Solo Attorneys Need to Know Before Connecting AI to Their Practice

Six days ago, the ABA Journal published a column with a title that should stop every solo attorney mid-scroll: “The AI policy your bar association wrote is for BigLaw. Here’s what solo and small firms should do instead.”

The author, Alan Walter, made the argument a lot of us have been feeling but couldn’t quite articulate. The ethics guidance is right: the duties of competence, confidentiality, and supervision absolutely apply to AI. But the compliance frameworks assume you have an IT department, a managing partner who reviews technology decisions, and a budget for enterprise software.

If you’re a solo practitioner, you’re all of those people. And you’re also the one doing the client work.

So let’s cut through it. Here’s what you actually need to know before connecting AI to your practice. Not the BigLaw version, but the version that works when it’s just you.

The Rules: ABA Formal Opinion 512

In July 2024, the ABA issued Formal Opinion 512, the first formal ethics opinion on generative AI in legal practice. It’s the baseline, and every solo attorney using AI should understand what it says.

Opinion 512 applies four existing rules to AI:

Rule 1.1 (Competence) requires you to understand how the AI tool works well enough to use it properly. You don’t need to understand transformer architecture, but you do need to know whether your conversations are being used to train the model, where your data is stored, and what the tool can and can’t do reliably.

Rule 1.6 (Confidentiality) requires you to make reasonable efforts to prevent unauthorized disclosure of client information. If the AI tool you’re using feeds your client’s data into its training pipeline, that’s a potential disclosure even if no human ever reads it. The duty of confidentiality is not a probability game.

Rule 3.3 (Candor) means you can’t submit AI-generated content to a court without verifying it’s accurate. If the AI hallucinates a case citation, that’s on you. Always.

Rule 5.3 (Supervision) extends your supervisory duties to AI tools the same way it extends them to paralegals and associates. You’re responsible for the output.

None of this is new law. It’s existing obligations applied to a new tool. The challenge isn’t understanding the rules - it’s implementing them when you don’t have a compliance department.

What Happens When You Get It Wrong

In February 2026, Judge Jed Rakoff of the Southern District of New York issued a ruling in United States v. Heppner that made the stakes very concrete.

A defendant used the free consumer version of an AI assistant to develop legal strategy documents. When those documents were later sought in discovery, the defense claimed attorney-client privilege. Judge Rakoff disagreed. His reasoning: the free consumer version’s terms of service didn’t support confidentiality, the AI tool isn’t an attorney or an agent of one, and sharing privileged information with a consumer-grade AI service constituted a third-party disclosure that waived the privilege.

The critical detail wasn’t that the defendant used AI. It was that he used the free consumer version - the one that trains on your inputs by default.

This is the distinction that matters more than any other: not whether you use AI, but which version of the AI you use.

The Training Problem: Free vs. Paid Is Not Just About Features

Every major AI provider draws a sharp line between consumer and business tiers. The difference isn’t just speed or model access. It’s whether your client’s data becomes part of the AI’s training dataset.

Here’s the current landscape:

Anthropic (Claude): Free, Pro, and Max accounts train on your conversations by default. Anthropic switched from opt-in to opt-out in September 2025, and if training is enabled, your data can be retained for up to five years. Business accounts (Claude for Work, Enterprise, API) are governed by separate commercial terms that prohibit training entirely.

OpenAI (ChatGPT): Free and Plus accounts train on your data by default. You can opt out under Settings > Data Controls, but the opt-out is a toggle, not a contract. Enterprise and Team accounts don’t train on your data by default and are governed by separate business terms.

Microsoft (Copilot): The consumer version - the free Copilot you reach at copilot.microsoft.com with a personal Microsoft account - can use your conversations to train Microsoft’s models. Microsoft’s own consumer terms call it “for entertainment purposes only.” Microsoft 365 Copilot, the version tied to a work account, carries Enterprise Data Protection with a contractual commitment that prompts and responses aren’t used for training. The work account is the dividing line, not the brand name.

The pattern across all three is the same: the free version trains on you. The paid business version doesn’t. And “paid” doesn’t always mean “business” - a $20/month Claude Pro subscription is still a consumer account under consumer terms.

For a solo attorney, this creates a surprisingly clear rule: if you’re putting client data into an AI tool, you need to be on a business tier, or you need to have opted out of training and understand that an opt-out toggle is weaker than a contractual commitment. The Heppner ruling suggests courts will look at the actual terms of service, not your intent.

(We’ll do a deeper dive comparing the specific terms of service across providers in a future post - there’s a lot more to unpack, including data retention periods, deletion rights, and what happens to your data after you cancel.)

What 35 State Bars Are Saying

The state-level activity is accelerating. More than 35 state bars have issued guidance on AI, though the approaches vary significantly.

Florida now requires attorneys to provide written disclosure to clients when AI is used to process their data, effective July 2025. California is incorporating AI provisions directly into its Rules of Professional Conduct. Colorado became the first state to embed the term “artificial intelligence tool” in its ethics rules in January 2026.

The Gap Between Adoption and Policy

Clio’s 2026 Legal Trends Report tells a striking story about the current state of AI adoption:

• 71% of solo practitioners are now using AI in some form
• 43% have no AI policy and no plans to create one
• Only 9% have an enforced written policy
• Fewer than 33% have increased revenue with AI

That last number is important. Most attorneys are using AI, but most aren’t using it in ways that move the needle - because they’re using it as a better Google search or a fancy spell-checker, not as something connected to their actual practice data.

And that 43% with no policy? Those are the attorneys most exposed to a Heppner-type outcome. If you can’t articulate what AI tools you use, which tier you’re on, and how your client data is handled, you’re relying on luck rather than competence.

The One-Page Policy You Actually Need

The ABA Journal piece gets this right: a solo attorney doesn’t need a 30-page governance framework. You need a one-page policy that you actually follow.

Here’s what it should cover:

Which tools you use and which tier. Name the AI tools. State whether you’re on a consumer or business account. If consumer, confirm that training is opted out. This is the Heppner lesson - the court looked at the actual terms of service.

What data goes in. Define a clear line. Client names, case details, and privileged communications need business-tier protections. General legal research questions with no identifying information can use consumer tools. When in doubt, anonymize.

How you verify output. State that you review all AI-generated content before using it. For anything filed with a court, you independently verify citations. This covers your Rule 3.3 obligation and protects you from the hallucination problem that has cost other attorneys sanctions - including a Brazilian case where lawyers were fined $16,500 for submitting AI-generated content that turned out to be fabricated.

How you disclose to clients. If you’re in Florida, written disclosure is mandatory. Even if your state doesn’t require it yet, a brief mention in your engagement letter that you use AI tools for administrative and research tasks, with appropriate safeguards, demonstrates the kind of transparency Opinion 512 contemplates.

Put it on one page. Sign it. Date it. Review it every six months. That document alone puts you ahead of 91% of the profession.

Why the Architecture of Your Tools Matters

Here’s where this gets practical for how you choose software.

Most practice management platforms that offer AI features bundle their own AI model into the product. That means the vendor controls all three pieces: your private practice data, the AI that processes it, and the ability to take action (send emails, create documents, update records) on your behalf.

Security researchers call this combination the “lethal trifecta” - private data plus untrusted content plus external communication creates the conditions for data exfiltration through prompt injection attacks. It’s not theoretical. Palo Alto Networks’ Unit42 team documented specific attack vectors where AI tools with direct access to data and external systems can be hijacked through prompt injection, and the risk is highest when a single vendor controls all three legs.

The solution is separation. Instead of bundling AI into your practice management software, your software should connect to the AI you already use. Writly’s MCP server provides just enough data to your AI assistant; the AI itself runs inside Microsoft’s or Anthropic’s or OpenAI’s security boundary. That’s two laters of security: Writly controls what goes out, and the AI provider controls what happens with it. Those companies employ hundreds of dedicated security engineers. They hold enterprise data protection agreements with Fortune 500 companies. They’re investing billions in model safety and infrastructure hardened for confidential data at scale.

Writly’s job is making practice management excellent. Their job is making AI secure. When each system does what it’s built for, the blast radius of any single vulnerability shrinks dramatically - no one vendor holds all the keys.

This is also what makes the Copilot time entry workflow possible - and safe. Copilot reads your practice data through the open connection, cross-references it with your Microsoft 365 activity, and finds the billable hours you missed. But the AI processing happens inside Microsoft’s enterprise data protection boundary, governed by the same compliance standards that protect their banking and healthcare customers.

Getting Started - The Right Way

If you’re already using AI (and statistically, you probably are), here’s the path forward:

Step 1: Check your accounts. Open the settings for every AI tool you use. Are you on a consumer or business tier? Is training opted out? This takes five minutes and it’s the single most important thing you can do today.

Step 2: Write your one-page policy. Use the framework above. Name your tools, define your data boundaries, commit to verification, and decide how you’ll disclose to clients.

Step 3: Connect AI to your practice data - on business terms. The difference between using AI as a generic chatbot and using it as a practice tool is whether it can see your matters, contacts, and billing history. Connecting AI to your practice data is what turns a $20/month subscription into a tool that recovers thousands in missed billable hours.

If you’re not using Writly yet, you can start a free 14-day trial and set up the AI connection the same day. We connect to Copilot, Claude, and ChatGPT through the open MCP standard - your data, your AI, your choice.

---

This post is part of the “AI in Practice” series. See also: How Copilot Finds the Billable Hours You Missed, Connecting AI to Your Practice Data, and Stop Worrying About AI Writing Your Briefs.