Start with the matter lifecycle, not the tool list
A lawyer's work changes shape as a matter moves forward. Intake needs structure and conflict information. Early analysis needs a chronology and issue list. Research needs authoritative sources and adverse arguments. Drafting needs a reliable factual record and jurisdiction-specific review. Client communication needs clarity and judgment. Billing needs a defensible description of the work.
That lifecycle is a better way to choose AI than asking which chatbot is “best.” A general model may be useful for a first outline. A legal research platform may offer better source visibility. A document system may be the only appropriate place for confidential files. The choice follows the risk and the task.
ABA Formal Opinion 512 is a useful anchor because it does not create a magic exception for AI. It connects AI use to familiar duties: competence, confidentiality, communication, supervision, candor, meritorious claims, and fees. I use that as my test: can I explain what the tool did, what I checked, what I changed, and why the final work is reliable?
The intake and conflicts workflow
Intake is an excellent place to use AI for organization because the first job is often to turn a conversation into a list of facts that still need confirmation. I can ask for parties, dates, events, documents, claimed harms, deadlines, and unanswered questions. I do not ask the model whether the matter is strong or whether the firm should accept it.
I keep conflicts and confidentiality separate from convenience. A prospective client can share sensitive information before an engagement exists. The firm needs a policy for what may enter an AI system, who may access the output, and how the information is retained. A redacted or fictional example is appropriate for a first test. A real intake belongs in a firm-approved workflow.
The output I want is a verification queue. It should tell me which name is misspelled, which date conflicts with another note, which document is missing, and which statement is an allegation rather than an established fact. That is more useful than a confident paragraph that makes the intake feel complete when it is not.
Chronologies, timelines, and the factual record
Long matters bury important dates in email threads, notes, pleadings, and attachments. AI can help create a chronology with a source column and a disputed-point column. I want each event connected to the document or testimony that supports it. I also want gaps and contradictions called out rather than smoothed away.
I never treat a generated chronology as the record. I compare it with the underlying files and make a human-approved version the working document. The review is especially important when a date affects a limitation period, notice requirement, contractual deadline, discovery response, or court order.
For a small firm, the measurable value is not “AI read 500 pages.” It is whether the first attorney review surfaced the right missing documents and reduced time spent reconstructing basic facts. I keep the original source references so another person can audit the work without relying on the model's memory.
Research assistance without citation roulette
AI is useful before research and after research. Before research, it can turn a client question into issue statements, synonyms, elements to test, and adverse arguments. After I open the authorities, it can help compare the language, create a case chart, or explain a difficult passage in plainer terms.
The dangerous middle is asking a general model to supply a finished list of cases, quotations, or holdings and then copying the answer into a brief. A citation that looks plausible may be wrong, out of jurisdiction, no longer good law, or irrelevant to the proposition. I open the authority and check the text, court, date, procedural posture, current status, and exact support.
I ask AI to make uncertainty visible. “I could not verify this proposition” is a useful result. “Here are three possible searches and the primary sources I need to locate” is a useful result. A fluent answer without a source trail is a research lead, not legal research.
Document review and contract work
For document review, I ask for an issue list rather than a final opinion. The model can locate defined terms, compare two versions, find inconsistent dates, identify missing exhibits, or group clauses by topic. I then review the underlying language and decide what matters for this client and this deal.
A contract prompt needs more than “review this agreement.” I specify the client's role, business goal, governing law to verify, risk tolerance, fallback positions, and the output format. I ask for section references and questions for the client. That turns a broad request into an inspectable work product.
For litigation documents, I add a privilege and work-product rule. The tool must be approved for the material, the review team must understand its limitations, and the lawyer must test whether the system missed documents or misclassified them. Technology-assisted review can help manage volume, but the firm still owns the defensibility of its process.
Confidentiality and privilege are workflow questions
I do not decide that a tool is safe because it is popular or because its interface looks private. I check the terms, training and retention settings, access controls, data location where relevant, administrative controls, audit logs, deletion process, and whether the firm has an agreement that matches its obligations.
ABA Model Rule 1.6 requires lawyers to protect information relating to a representation and make reasonable efforts to prevent unauthorized disclosure or access. That duty follows the information into a vendor system. If client confidences will be used, the firm should assess the material risks and whether informed consent or a different approved workflow is required.
For a first experiment, I use public authorities, synthetic facts, or an anonymized document that cannot identify a client or matter. Anonymization is a risk reduction step, not a substitute for the firm's technology policy. I also keep the approved final work product in the document-management system rather than in a chat window.
Client communication and the lawyer's voice
AI can help me turn notes into a client update with a clear status, next step, and request for information. It can also offer a plain-language explanation of a public process. I still rewrite the message because the client needs a lawyer who understands their situation, not a polished generic response.
I check that the draft does not promise an outcome, state a conclusion beyond the analysis, reveal another person's information, or turn an uncertainty into a fact. I remove legal jargon where it hides the actual decision the client needs to make. If the use of AI is material to a client's decision or requires consent, the firm should follow its communication policy and applicable rules.
A good client message makes the next human action obvious: send a document, answer a question, approve a position, attend a meeting, or wait for a filing. AI is valuable when it helps the lawyer make that action clearer.
Five prompts I would keep in a legal workflow
1. Organize an intake without giving legal advice
Using these approved and de-identified intake notes, create a factual matter summary with parties, dates, events, documents mentioned, claimed harms, unanswered questions, and conflicts-check details to confirm. Separate facts from allegations and assumptions. Do not provide legal conclusions, a merits assessment, or advice to the prospective client.
2. Build a chronology
Turn the approved notes below into a chronology with date, event, source document, person involved, disputed point, and follow-up item. Preserve uncertainty and identify gaps. Do not invent dates, infer intent, or state that a legal element is satisfied. Include a list of documents I should locate before relying on the chronology.
3. Create a source-led research plan
Based on this research question and the jurisdiction, create a research plan with issue statements, alternative search terms, primary authorities to locate, adverse arguments to test, and verification questions. Do not provide citations from memory. For each proposition, tell me what I must verify in the official source before using it.
4. Review a document for issues
Review this approved document and produce an issue list, not a final legal opinion. For each issue, quote the relevant section, explain the factual question it raises, identify the governing authority I need to check, and suggest a follow-up question. Flag anything outside the document or beyond what can be determined from the text.
5. Draft a client update
Using only these confirmed matter notes, draft a concise client update with what happened, what we know, what remains uncertain, the next action, and the decision or document needed from the client. Do not make a promise, add legal conclusions, cite law, or disclose information about another client. Mark every sentence that needs attorney review.
Billing, supervision, and the operating policy
AI changes how work is produced, but it does not make responsibility disappear. A partner supervising an associate, paralegal, or staff member needs to know which tools are permitted, which matter data may be used, what review is required, and how errors are reported. A short policy is better than an unwritten assumption that everyone uses the same settings.
Billing deserves its own paragraph in the policy. ABA Formal Opinion 512 discusses reasonable fees and the difference between actual time spent on a client task, direct tool costs, learning time, and general firm overhead. The engagement agreement, client communication, and applicable jurisdictional guidance should determine how the firm describes and charges for AI-assisted work.
I record the human work that remains: defining the task, checking sources, correcting the draft, applying legal judgment, communicating with the client, and approving the final product. That record helps with quality, training, and a candid answer if a client asks how a document was prepared.
What I would never delegate to a general AI tool
- The final legal conclusion, advice, filing decision, settlement recommendation, or client instruction.
- Citation verification, quotation checking, current-law analysis, or a representation that an authority supports a proposition.
- A confidential client file, prospective-client intake, privileged communication, or work product in an unapproved system.
- A decision about conflicts, privilege, materiality, sanctions, candor, or whether a claim is legally or factually supportable.
- A client-facing message that has not been checked for facts, tone, commitments, confidentiality, and jurisdiction.
- The firm's responsibility to train, supervise, monitor, and correct people who use AI in legal work.
A 30-day pilot for a solo or small firm
Week one: choose a narrow, low-risk task such as a chronology from synthetic facts, a research checklist from public authorities, or a client-document request draft. Write the review standard, prohibited inputs, approved tool, responsible attorney, and storage location before anyone runs it.
Week two: test three examples and save the input, output, corrections, source trail, and elapsed review time. Look for invented facts, missing issues, overconfident language, data leakage, and output that would be difficult for another lawyer to audit.
Week three: ask another attorney or experienced staff member to challenge the workflow. Can they reproduce it? Can they see which parts came from the source? Does the prompt ask the model to expose uncertainty? Is the review taking longer than the task itself?
Week four: decide whether to adopt, narrow, or stop the workflow. Keep it only if it improves the work without weakening confidentiality, source verification, client communication, or professional judgment. Record the decision and schedule a review because tools, terms, court expectations, and state guidance change.
My bottom line
The best legal AI workflow is not the one that produces the most words. It is the one that makes the next human review easier. I want a chronology with sources, a research plan with verification questions, a document issue list with section references, and a client draft that I can confidently edit.
Start with organization and source-grounded drafting. Keep client confidences inside approved systems. Verify every authority. Document the human judgment that remains. That is how a law firm can gain efficiency without asking a language model to become the lawyer.