AI Agents for Law Firms: Intake Without the Missed Call

Monday morning the staff returns a stack of voicemails and hears, more than once, that the caller already hired someone else. The call came in the evening, or on Sunday, while the person who needed a lawyer was comparing the first three firms that picked up a search. An AI agent for law firm intake can answer, sort the matter, and collect the names a conflict check needs before anyone offers a consultation time. The Legal Services Corporation Justice Gap Report found that 22% of people with civil legal problems did not know where to find help. We build this kind of intake for clinics and home service companies, where the phone rings while the person who should answer is with someone else. A law office has the same shape of day. This post is about that job. We do not practice law, and the limits below are for the firm to confirm with counsel.

Why does law firm intake keep losing potential clients?

People look for a lawyer when the situation will not wait for office hours: an accident, a hard conversation, a deadline they just understood. They call a few firms. The one that answers books the consultation. The ones that do not get a Monday callback into a decision that is already made.

The window that closes before Monday morning.

Personal injury and family inquiries bunch up in the evening and on weekends, which is when the office is dark. A firm with no coverage after hours hands those callers to whoever does answer.

The Justice Gap Report puts a wider frame on that single missed call. It found that 71% of low-income households had at least one civil legal problem in the past year, that 86% of civil legal problems received inadequate or no professional help, and that only 20% of problems initially prompted someone to seek a lawyer. The same report's 22% figure, people who did not know where to find help, is one reason a caller who misses on the first try often stops.

What a missed evening call costs a contingency firm: an illustrative example.

Swap these placeholders for your own log. They are not a firm we work with. A personal injury office takes 30 inbound calls a week, and 30% arrive after hours. If those nine voicemails convert at 15% on a Monday callback, against 40% when someone answers live, the gap is about two retained clients a week. At an illustrative fee of $8,000, that is $16,000 a month of work the log can test. Your missed-call report is the real number.

What does an AI intake agent actually do for a law firm?

A phone tree routes on a keypress. An intake agent reads what the caller says, classifies the matter, collects what that practice area needs, notes conflict signals, and either offers a consultation time or sends the attorney a summary. It acts on the message. The tree waits for a button.

The five tasks the agent handles without staff.

After-hours answering. It picks up when the office does not, says it is an automated intake assistant, and asks the caller to describe the situation in ordinary words.

Practice area. From that description it chooses a path: personal injury, family, estate planning, a business dispute, criminal defense, or something else. An open first question keeps mis-sorts rare.

The first facts. Each practice area has its own questions. This is not a generic contact form.

Names for a conflict check. It asks for every adverse party and writes the names down so the attorney can check the existing client list before the consultation is confirmed.

The consultation time. If nothing is flagged, it offers open times from the attorney's calendar and confirms before the caller hangs up.

Where the handoff line sits.

The agent stops when the caller describes immediate danger, uses emergency language, gives a name that needs a person to check conflicts, or asks whether they have a case. The attorney gets the transcript and the classification quickly. The GPT-4 Technical Report (OpenAI, arXiv:2303.08774, March 2023) documents a simulated bar exam score around the top 10% of test takers. That score is not why you would use software for intake. You would use it to sort, collect, and hand off cleanly, and to stay out of advice.

How do you build the qualification layer for a law firm agent?

The qualification layer is the question set that decides fit and what the attorney needs before the meeting. Too few questions and the consultation starts blind. Too many and the caller drops off before a time is booked.

The four questions every practice area shares.

Four questions show up in nearly every intake:

1. What is the matter, in the caller's words, and how does the agent classify it?
2. When did the event happen, or when is the filing deadline?
3. Which jurisdiction is involved?
4. Has the caller spoken to or retained another attorney about this?

Then the practice area adds its own. Estate planning asks about asset type and a rough size of the estate. Family law asks whether children are involved and whether anything is already filed. A business dispute asks for both business names. Personal injury asks about insurance and whether emergency services responded.

LegalBench (Guha et al., Stanford, August 2023, arXiv:2308.11462) tested 20 large language models on 162 tasks across six types of legal reasoning. Sorting a description into a matter type, pulling a deadline, and naming a jurisdiction are in that family of tasks. The paper is a benchmark of the models, not a clearance to give advice.

Example: qualifying a personal injury caller step by step.

Step 1. Confirm the caller's name and the best number.

Step 2. Ask what happened and when. Classify auto accident, slip and fall, medical negligence, or other.

Step 3. Ask whether it happened in the firm's state, and flag out-of-state matters for review.

Step 4. Ask whether another attorney has already heard about the incident, and log the answer.

Step 5. Ask for the names of other parties, and log them for the conflict review.

Step 6. Offer open consultation times and confirm.

A flow of about five steps usually takes three to five minutes. Flows longer than seven questions see people drop off. The scoring pattern underneath this, across other kinds of businesses, is in how lead qualification logic works.

What happens when a conflict check or sensitive disclosure comes up?

Conflict checks and sensitive disclosures are where the agent must stop. They are ordinary on inbound calls, and getting them wrong creates ethical exposure. The agent's job is to notice and hand off.

The two scenarios that always go to a human.

Conflict checks. The agent collects adverse-party names and logs them. It does not run the check. The check needs the firm's matter system and, in most states, a licensed attorney. The win is that the attorney has the names before the consultation, not after the client is in the waiting room. A check the software cannot finish waits for the attorney the next morning. It is not a failed intake, and it is not a rejection.

Sensitive disclosures. Immediate danger, an ongoing safety situation, or something that could trigger mandatory reporting goes to a person within two turns. No script overrides a caller in distress.

Two more patterns escalate in every practice area. A caller who says a prior attorney reviewed the matter and declined often has complexity the intake script cannot judge. A call that opens with a specific attorney's name is an existing relationship or a referral, and it needs that person, not a new-matter script.

Which practice areas see the fastest return from AI intake?

The faster payback shows up where inquiry volume is high, the first questions can be written down, and a missed call has a cost you can count.

The three strongest first builds.

Personal injury. Contingency work ties intake to revenue, because a retained client has an expected fee. Accidents happen when the office is closed. An agent that captures those calls and qualifies them before Monday turns a voicemail stack into a list with names, matter types, and conflict flags already written down.

Estate planning. These callers are serious and usually not in a ten-minute race among three firms. They do call after work, when they have time to think about beneficiaries or a will. Answering, sorting by how complex the estate is, and booking the consultation catches the after-hours calls that currently die.

Family law. Divorce and custody calls are sensitive and sometimes urgent. A firm that answers and books on the same call is in a different position from one that returns the call 24 to 48 hours later, because the decision is often made while the situation feels most acute.

How do you measure whether the intake agent is working?

Three rates tell you: contact rate, the share of inbound contacts the agent picks up; qualification rate, the share that finish the questions; and booking rate, the share of qualified callers who take a time. Everything else waits until those three are clean.

What to check in week one.

In week one, watch contact rate and qualification rate. Booking rate depends on the calendar, which takes a week to see against real volume. If contact rate is high and qualification rate is under 60%, the flow is too long or the opening is confusing. Read the transcripts where people stop, and shorten the questions. If both rates are solid, let booking data accumulate for seven days before you judge the calendar.

The Wharton 2025 AI Adoption Report found that 82% of enterprise leaders use generative AI at least weekly, up 10 percentage points from the year before, that 89% agree it enhances employee skills, and that 72% formally measure return. Law firm intake is measurable in a plain way: every contact is logged, every finished question set is a data point, and every booking is a consultation that becomes a retained client or does not. The same three rates, followed through to retained clients, sit next to the inbound follow-up framework. Intake's output is what follow-up starts from.

What should stay with a human in law firm intake?

Some pieces of intake carry ethical and relationship stakes the software should not take.

Legal advice. The agent collects facts. It does not analyze them. "Do I have a case?" gets one reply: the attorney will review what you shared and answer that in the consultation. No opinion on the merits, ever.

The conflict decision. The agent collects names. The decision is the attorney's. Do not automate the screening comparison unless a licensed attorney has reviewed the process under the firm's bar rules.

Fees. Contingency percentages, retainers, and billing structures stay out of the script. A caller who asks about fees hears that the attorney will cover the fee structure in the consultation. The agent books the meeting.

A caller in distress. Immediate danger, an ongoing safety situation, or a caller breaking down gets a person within two turns. Set that rule before go-live.

If you want to see which intake step to write down first, get your free Systems Plan and put the practice area in the note. The plan arrives by email a few minutes after you send it. The build, if you want one, is the same repeated job we do for a clinic desk or a home service office, under your firm's rules.

Common questions.

Is it ethical for a law firm to use an AI agent for intake?

It can be, with hard limits. The agent collects facts, classifies the matter, and schedules consultations. It does not give legal advice, run conflict checks, or judge the merits. It identifies itself as an automated intake assistant at the start of every call. Before you turn it on, read your state bar's current guidance and confirm the operating rules with ethics counsel. Advice and intake are different jobs, and only the second belongs in the script.

How does an AI intake agent handle conflict checks?

It collects the names of adverse parties and logs them for the attorney to review before the consultation is confirmed. It does not run the check. That determination needs the matter system and, in most states, attorney review. The practical gain is names the night before, rather than a surprise in the waiting room. A flagged call waits for a person. The agent does not reject it and does not route around the attorney.

Which practice areas benefit most from an AI intake agent?

Personal injury, estate planning, and family law tend to show the fastest return. Injury firms tie after-hours conversion to contingency fees, and accidents happen when the office is closed. Estate planning catches serious callers who research after work. Family law turns on speed: the decision to retain is often made while the situation feels most acute, and a same-call booking is a different conversation from a return call 24 to 48 hours later.

What happens if a caller describes an emergency during AI intake?

Immediate danger, an ongoing safety situation, or a caller who is clearly breaking down goes to a person within two turns. Set that rule before go-live. The agent listens for emergency language, immediate safety concerns, and mandatory-reporting triggers, and it routes then. No intake flow overrides a person in crisis.

How long does it take to set up an AI intake agent for a law firm?

An intake product with legal templates can be live in one to two weeks. A general platform configured to your practice areas and routing usually takes two to four weeks, including a review of real transcripts and a monitored first week. A custom build into a specific practice system such as Clio, Filevine, or MyCase takes six to twelve weeks from kickoff to production. The calendar follows how tangled the qualification logic is and how many systems the schedule has to read.

Where these figures come from.

The 71%, 86%, 22%, and 20% figures are from the Legal Services Corporation Justice Gap Report. The simulated bar exam result, a score around the top 10% of test takers, is from the GPT-4 Technical Report (OpenAI, arXiv:2303.08774, March 2023). The 162 tasks, six reasoning types, and 20 models are from LegalBench (Guha et al., Stanford, arXiv:2308.11462, August 2023). The 82% weekly use, the 10 percentage point rise, the 89% skills agreement, and the 72% who measure return are from the Wharton 2025 AI Adoption Report. The injury-firm dollars are round placeholders, not a client result.