How to Set Up an AI Receptionist for a Law Firm: What It Must Never Do and 14 Test Calls

An AI receptionist for law firms can answer calls, take contact details and the other side's name, book a consultation and page a lawyer. It must not give legal advice, accept a matter, clear a conflict or promise confidentiality it cannot guarantee. The six AI plans in my table run from $99 to $400 a month (read Oct 23, 2026). I have shipped 126 production AI systems, voice agents among them. I have not run this script for a law firm, so the test calls are a method with no results behind them yet.

This is the AI intake companion to my guide on what a virtual receptionist for lawyers legally creates, which covers the Rule 1.18 trap and human services. For live people overnight, read my legal answering service guide. Here I stay on the AI voice agent.

I am an engineer and no lawyer. This is the operational side of AI legal intake, and your state bar and statutes decide. I say where I did not read them.

Before you start

• A named lawyer who owns the decision. Rule 5.3 comment [3] asks for reasonable efforts to make sure services used outside the firm fit your professional obligations, and ABA Formal Opinion 512 applies Rules 5.1 and 5.3 to generative AI tools.
• An on call roster with a backup for each name.

Step 1: Write down what an AI receptionist for law firms must never do

Do this before you open a vendor page. The never list becomes the first lines of the prompt and the last gate in testing.

Two more duties sit with you, not the software. Rule 1.1 asks you to understand your technology, and Opinion 512 describes the standard as a reasonable understanding of the capabilities and limitations of the tool. Rule 1.4(a)(2) asks you to consult clients about the means of their representation.

Step 2: Open every call with the AI and recording notice

The first ten seconds should say what the caller is talking to, say the call is recorded, and stop them narrating their case. Here is my starting draft. A lawyer at your firm should write the final words, including the approved lines for tests 5 and 6 in Step 6. If a caller objects to the recording, the script should offer a human callback. The intake fields I would collect: name, callback number, best time, a one line topic and the names of other parties.

Thank you for calling [Firm name]. I'm an automated assistant, not a lawyer.
This call is recorded and transcribed so the firm can follow up.
I can take your contact details and book a consultation. I can't give legal
advice, and nothing I say means a lawyer has agreed to represent you.
Please don't share details about your situation yet. Is it okay to continue?

The last line before the question is the Rule 1.18 control. A prospective client's disclosures can disqualify the firm, and one way out under paragraph (d)(2) needs reasonable measures to avoid exposure to more disqualifying information than necessary, plus a timely screen and written notice. An agent that opens with "tell me what happened" collects exactly the facts that rule warns about.

Recording consent

Federal law allows a recording when one party consents, and your firm is a party: see 18 U.S.C. 2511(2)(d). States can be stricter. The Reporters Committee guides say all parties must consent to record a phone call in California (which also covers cellular calls), Florida and Pennsylvania, and in Washington, where a clear recorded announcement counts as consent. Those guides carry 2019 to 2021 dates and published lists disagree on how many states count, so read the current statute. I have not read the case law on a caller in one state and a firm in another, so I would announce on every call.

Telling callers it is an AI

I have not surveyed which states require AI disclosure on a phone call, so check yours. Opinion 512 says whether Rule 1.4 requires telling clients about AI use depends on the facts, and it addresses lawyers' work for clients, not intake callers, so that part is my read. California's bot statute is worded for online interactions and spares a bot user who discloses. LegalClerk's own FAQ says most callers cannot tell it is an AI and that this is by design. I would set the opposite rule: say it first, and never let the voice claim to be human.

Step 3: Build the escalation rules before the intake questions

The failure I would test hardest for is an urgent call handled as routine. Write each trigger, the action and the person paged, then set the response times yourself.

"Paged" needs a mechanism: the vendor's live transfer, an SMS plus a call to the on call phone, or a paging service. The AI never promises a callback time the paging system has not confirmed, and paging continues down the chain until someone acknowledges. For the vocabulary, see escalation path, and for how I build the handoff, my voice agent lead qualification guide.

Step 4: Log every call and fix the contract terms

Rule 1.6(c) asks for reasonable efforts against unauthorized access, and Rule 1.18(b) protects what a prospective client tells you. You cannot supervise a call you cannot replay. Keep these for every call:

• The audio, the transcript and the structured summary.
• The version of the opening script that was spoken.
• Every action the AI took, such as a booking or a record it wrote.
• Each escalation: trigger, who was paged, when someone answered.
• The model and prompt version in force that day.
• An access log showing who opened each recording.

Retention is a firm decision. I give no number, because I have not read your state's file retention rules. Ask your carrier and state bar. Then put these in the vendor contract:

1. No training on your calls, and no use of them beyond serving you.
2. A named list of subprocessors (outside companies that touch your data) and the countries where recordings sit.
3. A breach notice deadline in days.
4. Deletion on request and at termination, confirmed in writing, plus an export of your recordings first.
5. Access logs you can pull yourself.
6. Notice before the vendor changes the model or prompt behind your line.
7. A failover rule: if the AI is down, calls ring your phones or a human service.
8. Written confidentiality terms for any human agents who hear your calls.

Treat "ABA compliant" as marketing: LegalClerk's pricing page lists it. The rules bind lawyers, and Opinion 512 says lawyers should read the terms of use and privacy policy of any AI tool they use, or have an expert do it.

Step 5: Price the plans and pick a vendor

I read each plan on the vendor's own pricing page on Oct 23, 2026 and tested none of them. The billing unit differs, so rates are not directly comparable: per call favours short calls, per unique caller favours repeat callers, and flat plans favour a busy line.

Implied rates are my arithmetic: price divided by included units. Cheaper entry tiers exist: Smith.ai has a free plan with 25 calls, Dialzara Lite is $29 for 60 minutes with blind transfer only, and Goodcall Starter is $79 for 100 customers. NextPhone's page has a monthly and yearly toggle.

For live minutes, see AI versus human answering and my AI voice agent pricing breakdown. If you are shopping for an AI answering service for law firms, my short rule is this. Pick a plan with a warm transfer, where the AI stays on until a person picks up, because the escalation table in Step 3 depends on it. Ask each vendor for the Step 4 contract terms before you compare a dollar.

Step 6: Run 14 test calls before go live

I wrote this script as a method. I have not run it for a firm, so there are no results to report. Each vendor in the table lists a free plan, free trial or demo agent on its pricing page, so test there and forward your real line to the vendor number only after the calls pass. Call from a real phone, record every call, and tick each row. Failures on rows 3 to 6 and 9 to 11 block go live.

Row 14 is a prompt injection test, an attempt to talk the AI out of its instructions. In week one after go live, listen to every call. After that, sample a few a week plus every escalation.

Where the AI fails

Accents, dialects and bad audio. A 2020 study of five commercial speech recognition systems found an average word error rate of 0.35 for black speakers and 0.19 for white speakers, in a PNAS paper. Those systems are old, so I treat the number as a reason to test your own callers, not as a current score. A cell phone in a car is the same problem. The speech to text layer mishears a name or a digit, and the intake record is wrong.

Callers in crisis. A person who has just been arrested, hurt or served does not answer in neat fields, and a form is the wrong response to danger. That is row 11.

Several people on the line. A couple on speakerphone, an interpreter, or a relative calling for the injured person blur who the prospective client is. The AI cannot tell.

Calls I would keep with a human

• Any call where the caller is the opposing party or opposing counsel.
• Existing client calls beyond a message, because the AI cannot verify identity.
• Fee quotes beyond what you publish, engagement letters and anything about a conflict.
• Outbound calls to leads, which raise consent questions I cover in the lead qualification guide.

My pages on AI voice agents for law firms and work for law firms show how I scope one.

Frequently asked questions

Can an AI receptionist for lawyers accept clients?

No. Only a lawyer decides whether to take a matter. The AI can collect contact details and the other party's name, book a consultation, and say that a lawyer reviews every request. Anything that sounds like acceptance or rejection should fail your tests.

Do I have to tell callers it is an AI and that the call is recorded?

Federal law needs one party's consent, but the four states I checked (California, Florida, Pennsylvania, Washington) require all parties. I have not surveyed state AI disclosure rules for phone calls. I would announce both in the first sentence of every call. Check your state bar and statute, because I have not read every state.

What should AI legal intake never ask?

It should not ask the caller to describe what happened or who was at fault. Take the name, phone number, best callback time, a one line topic and the names of other parties for conflicts. A lawyer takes the story after the conflict check.

Is an AI answering service for law firms cheaper than a live one?

Goodcall's page puts live services at about $1.50 to $1.90 a minute, a competitor's claim. In my table, AI rates run from 26 cents per customer to $1.64 per AI minute, but the units differ. My AI versus human comparison works the live side out in dollars.

How long should I keep recordings and transcripts?

I give no number, because retention depends on your state's file rules, your malpractice carrier and whether the caller became a client. Decide it in writing before go live and make the vendor delete on schedule.

If you would rather have this built to this spec, with the escalation table and the 14 test calls wired into the agent, that is the work I do: see how I build AI agents.