In April 2026, a Sullivan & Cromwell attorney submitted an AI-generated brief containing fabricated citations. More than 1,350 documented AI hallucination cases have reached the bench, and the first quarter of 2026 alone produced over $145,000 in court-imposed sanctions - including a $110,000 penalty for an Oregon attorney and a license revocation in Nebraska. The problem was not the technology. It was the absence of a protocol.
Jennifer Case, J.D., founder of Law Tech AI, built an eight-step Legal AI Protocol that gives attorneys a defensible, repeatable workflow. Here is what it requires.
Foundation: accounts and tool selection
Free consumer AI accounts accessed with a personal email are not built for client matters. Under ABA Model Rule 1.6, attorneys carry a duty to protect client confidences, and consumer accounts may train on your inputs - a direct exposure risk for privileged information. Client work requires a paid business or enterprise account with a data processing agreement in place. Acceptable options include CoCounsel Legal, Claude for Work, and ChatGPT Team or Enterprise. If you are not certain which tier your account is on, check before you type anything client-related.
Not all AI is equal, and the distinction matters for client work. General-purpose models perform well at synthesis, drafting, pattern recognition, and client communications. Primary law research is another matter. For that, you need purpose-built legal AI that pairs an LLM with a verified legal database. CoCounsel Legal brings together Westlaw primary law, Practical Law expertise, and a built-in LLM in one interface. Every AI-generated answer traces back to primary law, so your research starts on verified ground, not on a model's best guess. Attorneys can develop these skills through structured AI for Legal training that covers both tool selection and risk management.
Prompting and research workflow
Vague input produces vague output - or, when the model runs out of real information to draw from, fabricated output. Apply the CLAR Framework to every research prompt:
- Context: Your role, practice area, client type, and procedural posture
- Legal framework: Jurisdiction, court level, and any statutes or cases you already know apply
- Assignment: One specific task per prompt - "Identify the elements of X under Y law," not "research this issue"
- Requirements: Format, length, citation style, what to include or exclude, and an explicit instruction to flag uncertainty
Add explicit guardrails to every prompt: no fabricated citations, pinpoint cites required, flag uncertainty, and show reasoning. Lawyers are trained to lead witnesses, and that instinct produces bad AI output. Ask for analysis, not validation. After any substantive research response, ask the AI to score its own confidence on a scale of 1 to 10. It will surface gaps it did not volunteer the first time.
Run primary research in CoCounsel Legal using Deep Research and Document Analysis. These tools pull from Westlaw's verified database, giving you a foundation that holds up. Once you have verified research, move to a general-purpose AI for synthesis and drafting. Use a Retrieval-Augmented Generation (RAG) approach: upload your verified materials to the model's knowledge base and instruct it to answer only from those documents. The AI drafts; you edit.
Verification and filing
AI identifies cases. You read them - the full opinion, the dissent, and every relevant passage. Note whether each authority is binding or persuasive in your jurisdiction. An AI-identified citation you have not personally reviewed is not a citation you can responsibly use. There are no exceptions to this step.
Return to CoCounsel Legal and run KeyCite on every citation in your brief. Then run Quick Check on the complete document before it leaves your desk. These are not optional quality checks - they are the final verification layer between your signature and the court. If you sign it, you own it.
Review the local rules for your jurisdiction and any standing orders from the assigned judge. Many courts now require explicit disclosure of AI use in filings. Supervisory responsibility does not transfer because AI produced the first draft. Review the work as a lawyer, because in the eyes of the court, you are the author of every word.
Why this matters for legal professionals
AI delivers faster research, sharper drafts, and more time for the analytical work that requires a trained attorney. But those advantages are only available when the workflow is sound. This eight-step protocol gives you a repeatable process that captures AI's efficiency gains without compromising your professional obligations or your clients' trust. The difference between a competitive edge and a sanctions order comes down to whether you follow a protocol - or just hit "submit."
Your membership also unlocks:
