AI Receptionist for Law Firms: Intake, Ethics and What It Costs Versus a Human Service

AI receptionist for law firms: what intake can safely look like

A solo family lawyer is in a deposition from 9 to 1. Four people call in that window, each in the worst week of their year. Two leave a voicemail, one hangs up, one gets the machine and tries another firm. By the time the lawyer calls back at 2 pm, the best lead of the week has retained someone else.

An AI receptionist answers those calls, asks the intake questions you would ask, and either books a consultation or sends you a clean summary. It can also create real problems for a law firm if you set it up carelessly, because the person calling is a prospective client and the rules around that are specific. This guide covers the benefits, the ethics checks, the prices and a safe setup.

How many calls law firms actually miss

The best data on this comes from Clio's 2024 Legal Trends research, in which a research firm phoned 500 US law firms as a prospective client. Roughly 40 percent of firms picked up, down from 56 percent in 2019. About 48 percent were effectively unreachable by phone, meaning they neither answered nor called back. Of the firms that missed the call, only 20 percent returned it, and 64 percent of prospective clients received no follow-up at all by phone or email. A separate 2025 study by Hennessey Digital looked at 1,333 firms and found a median response time of 13 minutes, with 26 percent never responding to web inquiries.

I could not find a credible figure for how many people hire the next lawyer after a missed call, and the numbers circulating on vendor blogs mostly have no traceable source. Use Clio's data for the problem and your own call log for your cost: count the calls you missed last month, your consultation-to-retained rate and your average fee.

What an AI receptionist should do at a law firm

The agent has one job: capture a good first impression and the facts you need to decide what to do. A sensible call flow is to answer in the firm's name and say it is an automated assistant, ask the matter type (divorce, injury, criminal charge, estate, business dispute), collect name, phone and a short description, ask the questions that screen out matters you do not take (wrong state, wrong practice area, a statute of limitations that may be close), and book a consultation or promise a callback by a specific time. The summary lands in your email or practice management system.

Urgent matters need a rule. If the caller says they have been arrested, served with papers due tomorrow, or are in danger, route to a person immediately. If it is a safety emergency, tell them to call 911.

What it must never do

It must never give legal advice, estimate case value, promise a result, tell someone whether they have a case, or say that an attorney-client relationship exists. Calling someone a client before conflict checks and an engagement letter is a problem on its own. Have the agent say plainly that it cannot give legal advice and that the firm will review the details.

The ethics checks you cannot skip

The American Bar Association's Formal Opinion 512, issued on July 29, 2024, is its first formal opinion on generative AI. It covers competence, confidentiality, communication and fees, among other duties. Its confidentiality section matters most for intake. It treats self-learning AI tools as a higher confidentiality risk because they can retain and reuse inputs, and says informed consent is needed before putting client information into a tool where outputs could disclose it. Model Rule 1.18 gives prospective clients their own confidentiality protection. The opinion does not address phone intake bots by name, so applying it here is a reasonable reading, not a ruling. It is also persuasive guidance, not law: your state bar's rules and opinions control.

In practice, before you pick a vendor, ask in writing: does the vendor store call recordings and transcripts, for how long, and who can access them? Does it train its models on your calls? Can you delete data on request? Is there a business associate or confidentiality agreement? Where is data hosted? And a few simple practices help: a short disclosure at the start of the call, a prompt that stops the agent from asking for more detail than needed for conflicts and routing, and a check of whether your state requires consent from every party to record a call.

If a vendor cannot answer these questions in writing, pick another. The same goes for a DIY build: you are responsible for the speech, language model and storage providers behind it.

What an AI receptionist costs against a human service

Prices below are as of October 2026, from vendor pricing pages.

Smith.ai's human-first plans are $300 a month for 30 calls, $810 for 90, and $2,100 for 300, with overages from $8.50 to $11.50 a call and a 10 percent discount for annual billing. Legal-friendly extras include free intake questions, extended intake at $1.50, conflict-check lookups at $0.50 and Clio integration. Smith.ai also sells AI-first plans, but their prices are not on the page I checked, so ask.

Ruby's plans are by minute: $250 a month for 50 minutes, $395 for 100, $720 for 200 and $1,725 for 500. Ruby says AI enhancements are included on all plans at no extra cost.

For AI-only options, Goodcall bills by unique caller: $79 a month for 100 customers (about 150 calls), $129 for 500 and $299 for 2,000. Rosie bills by minute: $49 for 250, $149 for 1,000 and $299 for 2,000. Building your own on Vapi runs about $82 to $129 per 1,000 minutes by Vapi's estimate.

Take a firm that gets 120 calls a month at three minutes each, or 360 minutes. Smith.ai's 90-call plan is $810, with 30 more calls at about $10.50 each, so about $1,125. Ruby's 500-minute plan is $1,725 and its 200-minute plan, $720, would run short. Goodcall Growth is $129, and Rosie Scale is $149. The gap is real, and so is the tradeoff: a trained human handles a distraught caller better than any current AI, and some firms use humans for the matters where tone matters most.

A worked example: a two-lawyer injury and family firm

Take a firm with two lawyers, one assistant and 120 calls a month, about a third after hours. These numbers are illustrative. The managing lawyer picks Goodcall Growth at $129, writes a script with five matter types, a list of matters the firm does not take and a rule to transfer arrests and service of process to her cell.

On a Saturday evening, a caller says a car hit them on Tuesday and they are still in pain. The agent says it is an automated assistant and cannot give legal advice, takes name, phone, date and place of the accident, whether they have seen a doctor and whether an insurance company has called them, then offers Monday at 10 am for a free consultation. It sends the caller a text confirmation and emails the lawyer a summary. If the caller asks whether they have a case, the agent says an attorney will review the facts and does not guess.

If one extra retained matter a quarter comes from catching calls that would have gone to voicemail, the software pays for itself many times over. If none does, the call log will show you in a month.

How to set it up safely

Step 1: Count last month's missed calls

Pull your phone records and count unanswered calls and unreturned voicemails. Note after-hours calls and calls from new numbers. This is the case for, or against, spending money.

Step 2: Write the intake script

List your matter types, the five facts you need for each and the matters you decline. Add the transfer rules for urgent matters and the disclaimers the agent must say.

Step 3: Vet two vendors in writing

Send the data questions from the ethics section to two vendors and keep the answers. Check your state bar's guidance on AI and on call recording.

Step 4: Configure the agent and limit what it asks

Tell it to collect only what you need for conflicts, routing and scheduling, to refuse legal advice, and to stop a caller who starts telling the whole story. Connect your calendar for consultations.

Step 5: Test it with ten hard calls

Call as an angry caller, a crying caller, someone asking for advice, a caller from the wrong state, someone who asks if you are a robot and someone who has an emergency. Fix every wrong answer.

Step 6: Start with after-hours and overflow

Forward only calls after hours or after four rings. Read every summary for two weeks and ask staff what they see.

When not to use one

If you handle serious criminal defence or other matters where a first call is usually a crisis, route a person to those calls and use AI only for overflow. If you take fewer than 30 calls a month, a human answering service or a good voicemail with fast text-back may be cheaper and safer. And if you cannot get clear answers about data handling, do not go live.

Frequently Asked Questions