This Smokeball CLE webinar, presented by attorney Jordan Turk (Director of Education at Smokeball), tackled one of the most consequential shifts in legal technology: the move from generative AI to agentic AI, and what that means practically and ethically for solo and small firm attorneys.
From passive tools to active agents. Jordan opened by distinguishing "OG" generative AI — a request-response model that answers a prompt and stops, requiring the lawyer to direct every step — from agentic AI, which plans, uses tools, executes multi-step workflows, and self-corrects toward an end result with minimal human intervention. The shift was illustrated with a family law example, as Jordan is a practicing family law attorney. Where generative AI might draft a parenting plan and then wait, an agentic system could retrieve the case file, draft discovery responses, check jurisdiction-specific rules, and hand back a complete package ready for careful attorney review and signature. Jordan framed this as a genuine inflection point for solo and small firms as a way to gain the workflow capacity of a much larger and better resourced team, provided it's deployed thoughtfully.
How agents are actually built. The CLE broke down the architecture behind the autonomy: a reasoning engine (the underlying model doing the analytical work), a task planner (breaking complex instructions into sequential steps), and an optional toolbox connecting the agent to databases, document systems, email, and calendars. A discovery-response walkthrough showed the perceive-plan-act-iterate cycle in action, with Jordan stressing a key practice point: citations should never be optional — force the system to return a source page, Bates number, or paragraph for every factual claim, and keep a human checkpoint before anything leaves the file.
Practical applications. Jordan covered real-world capabilities already in use by lawyers like her, such as document analysis, legal research (CoCounsel from Thomson Reuters and Lexis+ serve this function), discovery management (Disco, Relativity), and brief preparation, alongside family law–specific applications like parenting plan drafting, asset and debt inventories, support calculations, and regulatory monitoring. She also introduced Claude Skills as reusable, structured instruction packages that let attorneys standardize repeatable tasks (eg, client intake) without any coding, and walked through how a solo practitioner might responsibly build their first agent: pick a platform, define a narrow scope, and treat firm templates and matter data as the source of truth.
Ethics. The back half of the session focused heavily on professional responsibility. Jordan grounded the discussion in ABA Formal Opinion 512 and the Model Rules, walking through five duties: competence (Rule 1.1, including a duty to understand AI's capabilities and limitations), supervision (Rule 5.3, treating AI like nonlawyer staff requiring active oversight), confidentiality (Rule 1.6, including vendor vetting and informed consent), candor (Rules 3.1, 3.3, and 8.4, since AI hallucinations can trigger frivolous-claim, false-statement, and misconduct violations simultaneously), and fees (Rule 1.5, cautioning against billing full hours for work AI completed in a fraction of the time). She also flagged the growing patchwork of state-level guidance, with over 35 state bars weighing in and California predictably charting the path with its proposed rule amendments pushing the furthest, and the risk of the unauthorized practice of law if an AI agent’s output reaches a client without attorney review.
Cautionary examples. Jordan used a real case study where a law clerk's unchecked ChatGPT use led to a federal court opinion containing fabricated quotes and citations to underscore that even judicial chambers aren't immune to hallucination risk. She paired this with recent news of escalating sanctions against attorneys for AI-fabricated citations, and a Ninth Circuit development suggesting that individuals may bear legal liability for unauthorized actions their AI agents take on their behalf, even when a human didn't directly perform them. A rundown of hallucination types beyond fake case law, like invented statutes, misstated facts, false procedural deadlines, fabricated expert credentials, and mischaracterized adverse authority, reinforced the need for independent verification of every AI output.
⚡Key Takeaway: Agentic AI can meaningfully expand what a small or solo practice can handle, but it cannot exercise judgment, build client relationships, or substitute for a lawyer's professional responsibility. The technology handles the workflow; the lawyer stays accountable for the outcome.
Join us for more free CLEs where Jordan provides valuable legal learning on topics that matter to practicing attorneys. You can always find a full list here!