AI for Law Firms: A Safe Workflow Checklist

Do not begin by pasting a document into the first available chatbot. Start with the matter, client or prospective-client status, assigned task, responsible lawyer, operator, intended recipient, delivery channel, and protected workspace. The lawyer or firm must identify whether AI is permitted for this task, which tool and account may be used, what information may be submitted, what review is required, and which uses are prohibited.

Florida Ethics Opinion 24-1 is an advisory opinion, not a binding court rule, but it identifies core professional concerns for Florida lawyers using generative AI: competence, confidentiality, oversight, fees and costs, and advertising. ABA Formal Opinion 512 is supplementary national ethics guidance and is not binding Florida authority; the NIST AI Risk Management Framework is voluntary risk guidance rather than a Florida legal rule. Opinion 24-1 emphasizes that lawyers remain responsible for their work product and professional judgment. Staff therefore uses AI only inside an approved human-controlled workflow, never as an independent decision-maker or final reviewer.

A product name alone does not describe the environment that will receive the data. Verify the exact account and workspace, plan or contract, administrator controls, retention, provider access, model-training or improvement settings, sharing, integrations, connected storage, export, deletion, incident notice, and access conditions relevant to the approved task. A firm account and a personal or public account may have different controls even when the interface looks the same.

The FTC has warned AI providers to honor privacy and confidentiality commitments, but a provider’s marketing statement does not replace the firm’s review. Save the applicable terms, settings, approval, and review date rather than assuming that a familiar tool is approved for every client, data type, or use.

Before prompting, classify each proposed input under the firm’s approved scheme: public, firm-internal, client confidential, privileged or work product, personal or sensitive, court-restricted, or prohibited. Remove information that is unnecessary for the task, use neutral labels or synthetic examples when authorized, and obtain the required lawyer approval before protected information is entered, attached, pasted, linked, synchronized, or exposed through an integration.

Florida Ethics Opinion 24-1 explains that the confidentiality duty is broad and recommends obtaining the affected client’s informed consent before using a third-party generative-AI program when the use would disclose confidential information. The responsible lawyer decides whether disclosure is permitted, whether consent is required or appropriate, and what safeguards are sufficient. Staff does not infer consent from the client’s use of email, a general technology clause, or the fact that a tool offers a delete button.

AI cannot verify a matter from sources it was never given, and access to a source does not make the model’s use of it reliable. Assemble the lawyer-approved source set first and distinguish matter records, primary authority, official instructions, firm-approved references, and background material from AI-generated text. Preserve original documents separately from converted, OCR, excerpted, or annotated working copies.

Give the tool a bounded task, named source limits, required output structure, uncertainty instruction, citation format, and an express prohibition against inventing missing facts, quotations, authorities, page references, dates, amounts, people, or actions. Prompt wording is a control, not a guarantee; every output remains unverified.

A useful control record lets another qualified reviewer reconstruct what was evaluated without duplicating protected content into an insecure log. Record the matter and task, authorization, tool environment, operator, source-set version, materially relevant prompt and attachments, output, model or product identifier when available, date, reviewers, flags, disposition, and approval evidence.

Preserve iterations when a changed prompt, source, model, or output affects the work product. Subject to the approved firm and client retention or deletion policy, do not overwrite the rejected output that explains why a correction was needed, and do not treat a polished final document as proof that the underlying AI output was accurate or appropriately reviewed.

Treat the output as a candidate work product. Compare every material name, date, amount, quotation, citation, document reference, page or line, defined term, event, status, and requested action with the exact approved source. Check for omissions, changed meanings, contradictory records, unsupported inferences, fabricated detail, hidden instructions, bias, confidentiality leakage, and content outside the assigned scope.

For legal research, open each cited authority from an authoritative source and verify its identity, status, proposition, quotation, context, and later treatment as directed. The responsible lawyer must verify the accuracy and sufficiency of all legal research performed by generative AI; staff source-checking or another AI review does not replace that duty. For records and drafting, maintain source-to-output traceability. A second AI pass, automated score, citation-looking link, or confident explanation is not human verification.

The same safety sequence supports different tasks, but each workflow needs its own exact-set and decision boundaries. Intake summaries must preserve the prospective client’s words and leave conflict identity and representation decisions to qualified humans. Inbox and court-filing triage must preserve messages and attachments while humans control matter matching, deadlines, filing status, and every consequential action.

Chronologies, deposition indexes, discovery matrices, medical-record timelines, billing ledgers, and personal-injury demand materials require stable source citations and neutral treatment of conflicting evidence. Lawyers and qualified professionals control legal significance, testimony use, discovery sufficiency, privilege, medical interpretation, causation, damages, coverage, liens, valuation, strategy, and outgoing advocacy.

Route legal judgment, privilege, disclosure, consent, advice, strategy, filings, communications, advertising, fees, and remediation to the responsible lawyer. Obtain documented human review appropriate to the risk and final lawyer approval when the work product or action requires it. The reviewer must have the source set, not merely the AI output and a statement that it was checked.

Before release, freeze the approved version and inspect the exact outgoing file, message, filing set, or system entry. Remove comments, hidden prompts, tracked changes, metadata, or unnecessary AI artifacts as directed; verify recipient, channel, attachments, permissions, confidentiality, and status evidence; and preserve the approval and delivery record.

Florida Ethics Opinion 24-1 warns against duplicative or falsely inflated billing and addresses actual AI costs, overhead, and client communication. When a lawyer intends to charge an actual client-specific AI cost, the opinion says the lawyer should inform the client, preferably in writing. If the actual AI cost attributable to a particular client matter cannot be determined, periodic or subscription charges may not be prorated to the client and must be treated as overhead. ABA Formal Opinion 512 also discusses reasonable fees and communication duties. Staff records actual work and client-specific costs; the lawyer controls fee treatment, engagement terms, communication, and whether AI use must or should be disclosed in the circumstances.

Client-facing or third-party chatbots create additional intake and advertising risks. Opinion 24-1 says they must identify themselves as AI rather than a lawyer or law-firm employee and comply with advertising restrictions. An overly welcoming intake experience can also create prospective-client or relationship risk, so lawyer-approved warnings and limitations must be clear and understandable. More broadly, the lawyer controls all disclosures about AI assistance, limitations, incidents, corrected work, or affected recipients.

Close the AI work record only after delivery, approved retention or deletion, open limitations, downstream dependencies, corrections, and incidents are resolved or assigned. Reopen it if a source, fact, authority, output, vendor condition, or recipient action changes the reliability or permitted use of the work.