Enterprise DNA
Key Findings

Set four governance rules before staff deploy agents, covering client data, human approval, accountability, and practical controls for law firms.

Govern Employee-Built AI Agents at Your Law Firm
Insight ai

Govern Employee-Built AI Agents at Your Law Firm

Sam McKay

Staff may test agents to summarize files, sort intake, and draft routine messages. The problem isn’t that they’re experimenting. It’s that an agent can touch client information, create work product, or send something outside the firm before anyone has decided what it’s allowed to do.

For a law firm, that’s a governance problem from the first day of deployment. A useful agent needs a defined job, limited access, a human approval point, and a person accountable for its actions. Without those rules, a well-meaning experiment can expose privileged information or quietly create work that no one checks.

This matters financially, too. Document review, intake, and matter administration can consume time that doesn’t make it onto an invoice. The scale of that opportunity varies by firm. The answer isn’t to let agents run freely. It’s to put controls around them so the firm can reduce repetitive work without giving up professional judgment.

Start with the work, not the tool

Employees tend to build an agent around a task they know well. An associate may want a first-pass review of discovery files. An intake coordinator may want every new email sorted and summarized. A partner may want a faster way to scan contracts for a familiar set of clauses.

Each is a reasonable starting point. But “summarize this matter” isn’t a safe or sufficiently defined job. Which matter? Which documents? Can the agent see confidential notes? Does it save a summary to the document management system? Can it email the client? Who checks its work?

These are operating questions, not technical details. Answering them first makes it easier to decide what the agent should do, what it must not do, and what a human needs to review.

A firm can start by choosing one repeated workflow with a clear boundary. For example, an agent can classify new intake submissions and prepare a brief for a human to review. That’s different from having it decide whether the firm should accept a matter or give a prospective client legal advice.

The same distinction applies to document work. A first-pass review that flags clauses for an associate is a bounded task. An agent deciding a legal position or sending an unreviewed conclusion to a client is not.

Four governance steps to set before deployment

1. Define the agent’s job and prohibited actions

Write down the agent’s purpose in plain language. State the task it performs, the output it creates, and the actions it cannot take. Keep the statement concrete enough that a staff member could tell when the agent has gone outside its role.

For an intake agent, the permitted job might be to collect contact details, identify the apparent practice area, and offer a consultation time based on firm availability. It should not promise that the firm will take a case, provide legal advice, or tell a caller that a conflict check is complete unless the firm’s approved process confirms that result.

For a document review agent, the job might be to identify specified clauses and prepare a draft summary for an associate. It should not decide whether the firm accepts a client’s proposed language or update a signed agreement without approval.

Record what the agent is meant to do and what it is not meant to do. That simple boundary helps staff distinguish a useful automation from an unapproved delegation of professional judgment.

2. Limit access to client data

Give an agent access only to the information needed for its assigned task. A firm-wide drive is rarely an appropriate default. An agent that reviews a single matter’s documents shouldn’t be able to browse unrelated client folders or internal partner notes.

Define which data sources the agent can read, where its output can be saved, and who can operate it. Consider matter-level permissions, separate test data, and a process for removing access when an employee changes role or leaves. Keep sensitive credentials out of prompts and agent instructions.

Be particularly careful with privileged material, personal information, and documents involving conflicts or restricted matters. Staff should know what data they may submit, which approved environments are permitted, and how to report a mistaken upload. A policy that says “use good judgment” won’t tell someone what to do when a real client file is on their screen.

A useful next step is to review how law firms can lock down agent access and decide which controls fit the firm’s existing systems. Separate identities for each agent can also help the firm track activity and remove access without disrupting a person’s account. We cover that approach in why each agent needs its own identity.

3. Put a human approval gate on consequential actions

An agent can prepare work. A human should approve actions that affect a client, a legal position, or the firm’s obligations. That includes sending substantive client communications, filing documents, making a commitment on the firm’s behalf, accepting or declining a matter, and changing official matter records.

Be precise about the approval point. “A lawyer checks it” doesn’t say which lawyer, what they check, or whether approval happens before the action is taken. Define who reviews the output, what signs of error they look for, and how approval is recorded. The agent should pause until the required person has approved the action.

For low-risk internal tasks, the review can be lighter. A draft classification of an incoming inquiry, for example, might be checked by the intake team before it’s routed. A legal analysis or client-facing conclusion needs review by an appropriately qualified lawyer. The level of review should reflect the possible harm if the agent is wrong.

4. Name an owner and keep a record

Every deployed agent needs a named business owner. That owner is responsible for the workflow, access, review rules, and periodic checks. IT or a vendor may manage the technical setup, but that doesn’t make them accountable for whether the agent is appropriate for legal work.

Record the agent’s purpose, owner, approved users, data access, human approvals, and changes to its configuration. Keep a record of what it did and what was approved, following the firm’s retention and confidentiality requirements. If the system can’t provide a usable activity history, that’s a deployment concern to address before relying on it for sensitive work.

Review the agent when its task changes, its data access expands, or the firm’s policies change. Also give employees a straightforward way to flag unexpected behavior. A governance policy only works if people can use it during a busy day.

For a broader structure, see this law firm agent governance framework. It can help turn individual rules into a consistent process across practice groups.

What a governed workflow looks like

Consider a prospective client who calls after hours. An intake agent, if configured for the task, could collect basic details and ask approved intake questions. It should not give legal advice or tell the caller that the firm has cleared conflicts.

The intake information then goes to the firm’s approved process for conflict checking. Only after the required person or system confirms the next step should the caller be told what happens next. The agent records what it collected and routes the inquiry to the right staff member. If a question falls outside its approved script, it can explain that a person will follow up rather than improvising.

For email and web inquiries, an agent could be configured to classify the apparent practice area, assess fit against criteria approved by the firm, and prepare a brief for a partner. It shouldn’t reject a potential client or send a substantive response on its own.

The handoff needs enough context to be useful: what the person is asking for, when they need help, and what information is missing. A staff member checks the brief, runs the appropriate conflict process, and decides where the inquiry goes. The agent helps shorten the queue without becoming the decision-maker.

For a matter already accepted, an agent could be configured to perform a first-pass review of contracts, discovery batches, or matter files. A human should check the source documents, correct errors, and decide what matters. The agent shouldn’t make the final legal judgment or send the memo to the client without approval.

That workflow gives the firm a way to reduce repetitive handling while preserving clear human responsibility. It also gives the partner a better question than “Can we automate this?” The question becomes “What part of the process can an agent handle safely, and where must a person take over?”

If you want to map the workflows, data access, and approval points in your firm, Book a 60-min Omni Audit. The session is designed to help you consider the work worth addressing, the controls it needs, and a possible next step. No deck.

Make the economics visible

Manual work is often spread across several people, which makes its cost easy to miss. An associate may spend time on first-pass review, matter administration, and document handling that never becomes billable time. The capacity tied up in repetitive review can add up quickly. The exact opportunity depends on the firm’s rates, workload, review standards, and how much of the task is suitable for automation.

Intake has a different cost. When high-intent calls or form submissions wait hours for a response, the firm risks losing work before anyone has a chance to assess it. After-hours intake can be especially vulnerable. The effect of after-hours response times depends on the firm’s circumstances and follow-up process.

Those figures aren’t a promise that an agent will recover every lost hour or inquiry. They help partners see why it’s worth measuring the current process. Count how many files wait for review, how long intake sits before a response, how often a brief needs correction, and how much staff time goes into moving information between systems. Then compare that baseline with a governed workflow.

This is also why staff-built agents shouldn’t be treated as a small IT experiment. If an agent is doing work tied to client service, privilege, or billable capacity, the practice owner needs to know how it behaves and who is checking it. The law firm AI audit page provides more information about the Omni Audit.

Give staff a practical checklist

A policy can feel abstract until someone is building an actual workflow. Use a short checklist to ask what the agent is for, which client data it can access, what it may write or send, who approves its output, and who owns it after launch.

The AI Client Intake Checklist for Law Firms can be used as a starting point for discussing intake steps and human involvement. You can download the checklist directly and use it with the people who handle calls, forms, and conflict checks. Treat it as a starting point for a firm-specific process, not a substitute for your professional, privacy, or security obligations.

It’s also worth finding out what employees are already testing. A clear route for proposing an agent is safer than a blanket ban that drives experimentation out of sight. For more on the risks of unmanaged experimentation, read about shadow agents multiplying inside law firms.

Decide what is safe before the first live matter

A firm doesn’t need to approve every possible agent in advance. It does need a repeatable way to evaluate each one before it touches live client work. Set the task boundary, limit data access, specify human approval, and assign an owner. Then test the workflow with non-sensitive examples and confirm that staff know how to pause it or report a problem.

Start with one workflow and involve the people who do the work. An intake coordinator may spot a failure mode a partner won’t see. An associate can tell you which parts of a document memo need source citations. IT can confirm whether the proposed access model is workable. The partner responsible for the practice can decide whether the benefits justify the risks.

The goal is not to slow down every experiment. It’s to make sure the firm can explain what an agent does, what it can see, what it can change, and which person is accountable when something needs attention. That’s a better foundation for expanding agent use than discovering the rules after a client file has been mishandled.

To apply these steps to your own intake, matter administration, or document review process, Book a 60-min Omni Audit. The audit is intended to help you consider a practical next step, not a deck. You can also see Omni for law firms to understand how the audit applies to legal workflows.