AI for Lawyers: A Citation-Verification and Client-Information Review Workflow

A generated research paragraph should remain a draft until a lawyer can trace each material proposition to authority that actually supports it. Finding a case with the right name is only an existence check. The useful review asks what the court decided, where the quoted language appears, what procedural posture mattered and whether later developments change the analysis.

This workflow is for U.S. attorneys supervising AI-assisted work. It is a practical review method, not a substitute for the professional rules, court orders or research resources governing your matter. It also separates two decisions that deserve independent attention: whether a tool may receive client information, and whether its output is fit for professional use.

Define the assignment before opening a prompt

Write a short research instruction identifying the jurisdiction, precise question, procedural posture and desired output. State which facts are established, disputed or assumed. A request to “find cases that help us win” encourages an unfocused answer; a bounded issue can be checked.

For an initial task, you might request an issue outline with unresolved questions rather than a filing-ready argument. Specify whether the output must distinguish binding from persuasive authority. Ask for direct source locations, but treat every supplied link and pinpoint as unverified until reviewed.

ABA Formal Opinion 512 discusses competence, independent judgment and verification in using generative AI. Its Model Rules analysis does not replace the rules adopted in your jurisdiction.¹ A firm's assignment should name the lawyer responsible for review rather than leaving responsibility implicit in a shared chat.

Decide what information the tool can receive

Use the approved account, configuration and workflow for the task. A provider name alone does not establish that a personal account has the same terms or safeguards as a firm's contracted service.

Before including matter information, record the reviewed terms and relevant controls: retention, training or reuse, authorized users, shared links, connected applications and deletion arrangements. Consider how those controls operate for the actual input and output. Where required, resolve client communication or consent before disclosure.¹

California's 2026 practical guidance expressly addresses agentic systems and emphasizes supervision and verification as access and autonomy increase.² If a tool can read a repository, drive, email account or matter workspace, approval for a single document should not silently become approval for wider access. Give the reviewer an inventory of the intended sources and permitted actions.

This article makes no representation about any particular service's privilege, retention or security terms. Keep those determinations tied to the actual agreement and task.

Build a proposition-to-authority ledger

Read the generated draft once to identify its material assertions. Give each a row in a ledger before polishing the prose. Otherwise, a fluent paragraph can hide several unsupported steps inside one citation.

This is an internal working record, not a certification that every authority remains good law. Be specific about what was checked. “Opened official opinion” and “reviewed subsequent treatment using a named research service” describe different work.

Verify more than the citation's existence

Retrieve the actual authority from a reliable source. Match the case name, court, date and citation, then locate the passage. Check quotations character by character, including ellipses and brackets. Identify who is speaking before treating a sentence as the court's holding.

Next compare the authority's issue and posture with your proposition. A statement about pleading sufficiency may not decide what evidence proves at trial. An opinion quoting a litigant's position may reject that position later. A secondary summary may omit a limiting fact that controls the application.

The Northern District of California's AI guidance directs filers to read original sources, confirm actual support and check accurate, current authorities and quotations. It also warns that judge-specific AI disclosure requirements may apply.³ Treat this as a review floor; your matter may require additional checks.

For currency, perform the appropriate subsequent-history and treatment research with available reliable resources. Record what you used. If you lack a suitable tool or a later decision cannot be retrieved, leave a visible research gap rather than writing “verified good law.”

Worked example: an impressive paragraph fails three checks

Fictional training example: A draft cites “Example Decision A” for a broad rule about contract cancellation. The label is invented for this exercise and is not a real legal citation.

The reviewer opens the source supplied with the training file. The cited passage appears in a dissent, concerns a preliminary procedural question and says nothing about the nationwide rule asserted in the draft. The opinion exists, but the paragraph still fails.

The ledger should show three separate problems:

1. Attribution: The quotation is not the court's majority holding.
2. Scope: The procedural question differs from the proposition being asserted.
3. Geography: Nothing in the material supports a universal national rule.

The reviewer removes the proposition pending research instead of merely swapping the citation. That decision prevents the draft's desired conclusion from dictating which source gets attached to it. Once appropriate authority is found, the paragraph may need a narrower conclusion, a different factual assumption or an explicit unresolved question.

Do the same with facts. A valid citation cannot cure a generated statement that misreads the client's document or invents a date. Keep fact verification and legal research as separately traceable tasks.

Review the final document and every automated handoff

After revision, check that each surviving citation supports the sentence now present. Editing can move a citation away from the assertion it originally supported. Recheck quoted text, defined terms, exhibits, cross-references and requested relief against the approved version.

For team use, identify what the assistant may draft, what staff may prepare and what the supervising lawyer must decide. Give a person ownership of unresolved rows. A completed research memo should not automatically become an approved client message, pleading or filing.

California's updated guidance keeps professional judgment with the lawyer and cautions against unsupervised substantive actions by agentic systems.² A practical release record can list the final file, intended recipient, remaining limits, reviewer and separate sending or filing authorization. This is useful even when the process is entirely manual.

A reusable final-review checklist

Before relying on a draft, ask:

• Can a new reviewer retrieve every material authority and locate the passage?
• Does the draft distinguish facts, assumptions and legal conclusions?
• Are adverse authorities and factual complications addressed?
• Is the currency review accurately described, including gaps?
• Were the actual court rules and assigned judge's orders checked?
• Is the intended disclosure consistent with the approved information-handling arrangement?
• Has the responsible lawyer approved the precise version and next action?

Our product privacy explanation provides related context; privacy statements should not be treated as a privilege opinion. The LawSensai attorney information page explains the public offering. Neither a product citation nor an automated check replaces independent professional review.

Professional educational information, not a legal opinion or compliance certification. Apply your jurisdiction's rules, client obligations and court orders. Research checked October 26, 2026.

Sources and references

1. American Bar Association, Formal Opinion 512: Generative Artificial Intelligence Tools, July 29, 2024, competence and confidentiality sections. Addresses independent verification, professional judgment and evaluation of disclosure risks under the Model Rules. ↩ ↩²

2. State Bar of California, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, 2026 revision replacing the 2023 version, competence, confidentiality and supervision sections. Supports meaningful lawyer oversight, including for agentic systems. ↩ ↩²

3. U.S. District Court, Northern District of California, Using AI Tools in Your Case, verification and disclosure sections. Supports original-source checking and attention to judge-specific requirements; it is court guidance, not a nationwide ethics rule. ↩