AI can help a lawyer organize a research question, compare documents, and prepare a first draft. A model-generated answer alone cannot establish that a case is still good law, that a quotation supports a proposition, or that a client's facts are complete. A useful AI legal workflow therefore links every draft assertion back to a source and leaves the legal judgment with the lawyer.

A legal workflow connects source documents and research notes to a reviewed draft memo.
An editorial illustration tracing source documents through research notes to a reviewed draft.

Consider a hypothetical commercial dispute. A client asks whether a limitation-of-liability clause in a vendor agreement affects its claim after a service failure. The answer depends on the contract, the facts, the governing jurisdiction, and current law. The example below describes a process, not a conclusion about that clause or advice for a real matter.

  1. Scope: define the question, deliverable, jurisdiction, facts, and deadline.
  2. Protect: assess confidentiality and the tool’s data handling before sharing matter information.
  3. Retrieve: obtain the actual record and relevant law, then check authority and current treatment.
  4. Draft: use the reviewed packet to build a first draft with a source behind each material proposition.
  5. Verify: check facts, authorities, quotations, exceptions, and counterarguments against the originals.
  6. Decide: have the lawyer approve the answer, communicate its limits, and retain the appropriate record.

Define the question before opening an AI tool

The lawyer first identifies the deliverable: perhaps a short client memo that explains the clause, the likely arguments on each side, unresolved facts, and next steps. The working file records the agreement version, governing-law provision, relevant dates, client objectives, and deadline. It also separates confirmed facts from assumptions and disputed facts.

Only then does the lawyer decide whether AI is useful. A tool can help turn those inputs into an issue list: What does the clause cover? Are there exceptions? Which facts affect its application? What contrary arguments should be researched? The lawyer edits that list before research begins. An answer generated from an incomplete prompt can look comprehensive while quietly omitting the most important issue.

Confidentiality is a gate, not a final proofreading step. Before uploading an agreement, correspondence, or client narrative, the firm must evaluate the tool's data handling, access controls, retention, and permitted uses, as well as the client's instructions and applicable professional rules. ABA Formal Opinion 512 calls for a matter- and tool-specific assessment of disclosure risk before entering client information; some uses require the client's informed consent. Where appropriate, use an environment approved for that matter or a properly de-identified hypothetical; removing a name alone may not adequately protect a client's information.

Retrieve law, then use AI to map it

The research plan states the jurisdiction, court level, relevant time period, and questions to answer. Counsel searches authoritative statutes, rules, and judicial opinions through appropriate court sources or a legal research service. An AI research product can propose search terms, summarize an opinion, or group authorities by issue. Its output is a research lead, not authority.

For the vendor example, the lawyer would retrieve the actual contract and candidate authorities, then ask for a comparison of the facts and reasoning in those sources. A useful prompt asks the tool to identify which supplied passage supports each proposition, distinguish unfavorable cases, and mark gaps as unverified. The lawyer still checks for controlling law, later treatment, amendments, and jurisdictional differences. A citation that resolves to a real opinion can still be irrelevant, reversed, or quoted out of context.

The need for independent checking is concrete. A 2024 Stanford study of leading AI legal research products found that retrieved-source systems could still give unsupported legal answers. Its measured rates describe the tested products and period, not the reliability of every current tool. The practical lesson is durable: retrieval can improve the path to a source, but it does not verify how the source is used.

Build a traceable first draft

After selecting authorities, counsel can ask an approved tool for a memo outline or a rough draft using only the reviewed source packet. The draft should distinguish the contract text, confirmed facts, legal rules, analysis, contrary arguments, and open questions. It should also point to the specific source passage behind each legal proposition and contract interpretation.

For example, the draft might flag that the client has supplied two versions of the agreement, or that the available cases discuss a different type of damages. Those are questions for counsel and the client, not gaps for AI to fill with plausible language. If the tool cannot trace a sentence to the record or an authority, counsel removes or independently researches it.

This changes the drafting task from staring at a blank page to reviewing a structured candidate. It does not make the review automatic. The ABA opinion says the degree of independent verification varies with the tool and task, while the lawyer remains responsible for the work and cannot delegate professional judgment to a generator.

Verify every consequential claim

Before the memo goes to the client, the reviewer works through the draft against the original material:

  • Facts: Does every material factual statement match the agreement, correspondence, or other record? Are uncertain facts labeled as such?
  • Authorities: Do the cited opinions, statutes, and rules exist, apply in the relevant jurisdiction, and remain current? Does the cited passage actually support the sentence?
  • Quotations and citations: Are words, page references, party names, and procedural posture accurate in the original source?
  • Counterarguments: Does the memo address material exceptions, unfavorable authority, and alternative readings of the clause?
  • Client judgment: Does the advice reflect the client's goals, risk tolerance, and information still needed for a decision?

The lawyer can keep a simple source log with the proposition, source, location, date checked, and reviewer. That makes corrections easier when the facts change. If the work becomes a court filing, counsel must also check the applicable procedural rules and any judge-specific AI orders. In U.S. federal civil litigation, Rule 11 requires a reasonable inquiry into the legal and factual contentions in a pleading, motion, or other paper presented to the court. AI assistance does not alter the person's responsibility for that certification.

Approve, communicate, and retain the right record

The final reviewer decides what to say, what not to say, and whether the issue needs more research or client input. The finished memo identifies its assumptions and the date through which the law was checked. Counsel communicates any material limits in the analysis rather than masking them with polished prose.

A firm can record which tool was used, what information was provided to it, who checked the sources, and which version was delivered, in line with its retention and confidentiality policies. It should also assess how AI affects the reasonableness and communication of fees under the applicable rules; saving drafting time does not itself justify charging for time not spent. ABA Formal Opinion 512 addresses these obligations along with competence, confidentiality, supervision, and candor.

The repeatable pattern is scope → protect → retrieve → draft → verify → decide. AI is useful when it accelerates the middle steps and makes questions visible. The lawyer remains responsible for the client-facing answer. For the wider landscape of legal uses, see AI in Legal Practice; for the underlying technology, see What Are AI Agents?.