AI Legal Research: A Workflow for Checking Every Source

AI legal research works best as a way to develop questions, search terms, and tentative summaries, followed by independent checks of every authority you intend to use. Open the actual source, read it in context, confirm its jurisdiction and current treatment, and have a qualified person evaluate the conclusion.

A confident answer is not legal authority, even when it contains plausible citations. fenbs can keep approved research rules and links to checked public sources available across connected assistants, while privileged matter details stay in your approved systems. The goal is a research trail another person can inspect, not a longer chat transcript.

What are the steps for a checked research result?

1. Define the issue, jurisdiction, relevant time period, and procedural posture before asking the assistant for research leads.
2. Use an approved tool with appropriately limited information. Ask for possible issues, search language, and uncertainties rather than an unqualified conclusion.
3. Retrieve every potentially useful authority from an authoritative source or your legal research service. Treat an unavailable citation as unverified.
4. Read the relevant text, check whether the proposition follows, and investigate current treatment and contrary authority.
5. Write the conclusion with its limits, supporting sources, and unresolved questions. Have the responsible lawyer review it before use.

For example, suppose you are researching a federal procedural issue for a training exercise. State the court and the motion involved. Ask for concepts to investigate, then search those concepts independently. A rule relevant to one motion or court may not resolve a superficially similar question elsewhere.

What should you give the assistant before searching?

Prepare a short research question that separates known facts from assumptions. Specify whether you need controlling authority, persuasive examples, historical development, or a list of unresolved issues. If jurisdiction or the operative date is uncertain, make that an explicit question instead of letting the assistant choose silently.

• Issue: the precise legal proposition or question to investigate.
• Scope: jurisdiction, court level, date range, and procedural context.
• Sources: the approved research service and official materials to consult.
• Output: a list of leads, source passages to inspect, and reasons a conclusion may fail.

Keep confidential facts out of a tool unless its use has been assessed and approved for that information. Removing a client’s name may leave a distinctive fact pattern identifiable. The same caution applies when copying prompts, research notes, and handoffs into another service.

The ABA’s Formal Opinion 512↗ (opens in a new tab) discusses competence, confidentiality, supervision, and review when lawyers use generative AI. It is guidance based on the ABA Model Rules, not a nationwide permission to upload client information. Check the professional conduct rules and court requirements applicable to your work.

How do you verify an AI-suggested citation?

Start with existence. Find the case, statute, regulation, or other source independently; confirm the name, citation, court or issuing body, and date. An assistant can combine a real case name with the wrong citation or attribute a proposition to a source that does not discuss it.

Then check the exact proposition. Read the passage and enough surrounding material to understand qualifications, the issue actually decided, and the relevant facts. For cases, distinguish the court’s holding from commentary, a dissent, or a party’s argument. A genuine quotation can still mislead when removed from its context.

Use a research log with separate fields for citation found, passage checked, authority assessed, and current treatment checked. Do not use one “verified” marker for all four. If a source cannot be retrieved, do not rely on the assistant’s reconstructed quotation as a substitute.

A shared AI brain is useful for preserving that distinction: the link is a lead until someone has actually inspected what it supports.

How do you check authority and current treatment?

Ask two separate questions: does this source apply to the research question, and is it still reliable for this proposition? Identify the controlling jurisdiction and hierarchy. Check the effective version of a statute or regulation and investigate later decisions through appropriate research tools.

A negative treatment flag requires reading the relevant treatment, not automatically discarding the case. A clean-looking result also does not prove that the proposition controls. Explain the connection between the source, the legal issue, and the facts rather than treating a label as the analysis.

Search for the strongest contrary position as a separate pass. ABA Model Rule 3.3↗ (opens in a new tab) addresses, among other matters, known directly adverse controlling authority not disclosed by opposing counsel. Check your jurisdiction’s adopted rules; finding only favorable authorities is a poor research method regardless of the tool.

This article covers checking research; the separate guide to AI for legal documents addresses drafting work. Do not let movement from research to a filing bypass the lawyer’s review or the applicable court’s requirements.

What should the final research handoff contain?

Write the conclusion in terms of the checked evidence. State the scope, material uncertainty, contrary authority, and the remaining action. “Two relevant cases located; later treatment still unchecked” is a usable handoff. “The law supports our position” conceals the work that remains.

Keep the detailed matter analysis in the approved matter system. A separate shared work record can hold public source references and neutral progress descriptions where permitted. Our guide to using AI at work explains the broader habit of distinguishing draft output from completed review.

In fenbs, record people’s approved research instructions as Decisions and rules, and make each unresolved verification question a task. The handoff workflow helps a new assistant see where checking stopped without treating the prior assistant’s confidence as evidence.

How to do this with fenbs?

1. Create a free account and add a nonsensitive research project, such as a public-source training question.
2. Record the research rules a person approved: required source checks, confidentiality boundaries, and who reviews conclusions.
3. Add public source links and tasks for the checks still outstanding. Identify the proposition each checked source supports without adding privileged matter details.
4. Connect an approved supported assistant from Integrations. Ask it to read the context first and save where the research stands before stopping.

fenbs is not a legal research database, citation treatment service, or privileged matter management system. Use appropriate dedicated tools when you need those functions. Its role here is continuity for permitted project information, not deciding whether an authority is good law.

Try the process on a public hypothetical question before using it in work. Check whether a fresh session can distinguish a retrieved source from one that has passed every review step. For other professional workflows built around clear responsibility, see the AI by role library.