How to use AI on a matter without losing control: a 5-step workflow

What is a safe workflow for using AI on a legal matter?

A safe workflow treats AI as an assistant that works from a controlled slice of the matter record and whose output you check. It has five steps, summarised below, then explained.

This is general information. Professional duties on competence, confidentiality and supervision vary by jurisdiction, so check your regulator's guidance.

Step 1: How do you classify the task and the data?

Sort tasks by what goes wrong if the AI is wrong or if the data leaks.

High-risk tasks need stronger tool checks and heavier verification, or may be best done without AI.

Step 2: How do you choose a tool?

Before using a tool with client information, read its terms and settings. According to commentary on the ABA's Formal Opinion 512 (July 2024), concerns include whether a tool retains and reuses inputs, and the opinion's position as summarised is that informed client consent may be needed before entering client information into a self-learning tool. Read the opinion itself and your own regulator's guidance, as state and national rules differ.

• Does the provider train on your inputs, and can you opt out?
• Where is data stored and who can access it?
• Is there a business or enterprise plan with different terms?
• What does your engagement letter say about technology?
• Does your insurer have a view?

• Legal AI ethics guide

Step 3: How do you give AI context without oversharing?

Structured records make this easy. If the matter has a clean summary, parties, key dates, issues and a document list, you can paste only the relevant fields instead of whole files. Use roles in place of names where it does not weaken the task.

Asking it to list questions rather than assume is an inexpensive way to reduce invented facts.

• AI for small law firms: structure first

Step 4: What does verification involve?

1. 1

   Check the facts

   Compare names, amounts and dates against the source documents, not against the AI's summary.

2. 2

   Check the law

   Open every cited authority and confirm that it exists, says what is claimed and is current. Never file an unverified citation.

3. 3

   Check the reasoning

   Does the conclusion follow, and are counter-arguments missing?

4. 4

   Check the tone and promises

   Especially for client-facing text.

• Legal research with AI: a verification-first guide

Step 5: What should you record on the matter?

Add a short note or database row: date, tool, task, what data went in, who verified the output and any decisions. If a question arises later, you can show how work was done. It also reveals which tasks AI helps with and which it does not.

• Audit trail in Notion

Where does MatterOS fit?

MatterOS includes an AI prompt pack for summaries, client updates and next-step checklists, and structured Matter, Actions and Evidence records to draw context from. It does not decide whether a tool is safe for client data; that check, and the verification, remain yours.

• AI drafting prompts
• MatterOS overview

Frequently asked questions

Can lawyers put client information into ChatGPT or Claude?

It depends on the tool's terms, your plan and your jurisdiction's rules. Check whether inputs are retained or used for training, whether business terms apply, and whether client consent is needed. When unsure, anonymise or avoid client information.

What does ABA Formal Opinion 512 say about generative AI?

Issued in July 2024, it covers competence, confidentiality, communication, candor to tribunals, supervision and fees. Summaries note that informed consent may be needed before inputting client information into some tools. Read the opinion directly and check your own state's guidance.

How do I stop AI inventing facts or citations?

Give it a structured context, tell it to list missing information rather than assume, and verify every fact and authority against the source. Never rely on AI output for citations without opening the cited case or statute yourself.

Should I tell clients I use AI?

Many regulators and commentators expect transparency or consent in some circumstances, especially when confidential data enters the tool. Check your jurisdiction's rules and consider covering technology use in your engagement letter.

Why do structured matter records help with AI?

They let you pass a small, accurate slice of context instead of dumping whole files, which reduces both data exposure and errors. A summary, parties, dates and issues fields make prompts consistent and checkable.