AI Voice Agent for Lawyers: Cost and Safe Handling of Intake Calls

An AI voice agent for a law firm should greet the caller, say it is automated and not a lawyer, collect contact details, matter type, the other parties' names and any deadlines, then pass the record to a person for the conflict check. Build-your-own platforms list voice minutes from about $0.07 to $0.31; packaged AI receptionists charge per call.

The real risks for a firm are three: the agent collects privileged detail it should not, says something that sounds like advice, or mishears a name the conflict check depends on. So the script limits what callers are asked to say, refuses advice questions in fixed words, and has every surname spelled and read back before the handover.

Follow me on Instagram@sagnikteaches

The call, minute by minute

Intake calls are short and predictable, which is why they suit a voice agent. A workable order, with the time each stage should take:

Connect on LinkedInSagnik Bhattacharya

1. Greeting and disclosure (10 seconds). Firm name, that this is an automated assistant, and that a person will follow up.
2. Safety check (5 seconds). One question: is anyone in immediate danger? If yes, emergency services first, then transfer.
3. Confidentiality framing (15 seconds). What not to say yet, and that no lawyer-client relationship exists until the firm confirms it.
4. Contact details (45 seconds). Name spelled back, best number, email, safe times to call and whether a voicemail may be left.
5. Matter type and urgency (60 seconds). A short menu in the caller's words, with a "something else" branch.
6. Other parties (45 seconds). Names and spellings of everyone on the other side, plus anyone else involved.
7. Dates (30 seconds). Any court date, notice deadline or incident date the caller mentions, recorded exactly as said.
8. Read-back and next step (30 seconds). Summarise, confirm, and say when a person will call.

That is about four minutes and fifteen seconds of talk, which is why the cost sums below use 4.5 minutes per call. The agent never asks a caller to tell the whole story. A long account is a privileged conversation, and it is the lawyer's to hear.

Subscribe on YouTube@sagnikteaches

The confidentiality script

People start narrating as soon as someone listens. The opening has to slow them down. A filled-in example, written for a family-law-and-wills firm and adaptable to any practice area:

Hello, you've reached [firm name]. I'm an automated assistant,
not a lawyer, and I can't give legal advice.

Before we start: if anyone is in immediate danger, please hang up
and contact the emergency services now.

I'll take your contact details and a few facts so that the right
person at the firm can call you back. Please don't go into the
full story yet; the lawyer will go through that with you. Speaking
with me doesn't make [firm name] your lawyers. That only happens
once the firm confirms it to you in writing.

This call may be recorded and transcribed so our team can follow
up accurately. Is that all right?

Three details are deliberate. The consent question sits last, so a "no" can route to a human or a callback request. "Please don't go into the full story" stops the privileged narrative at source. And the sentence about no relationship until written confirmation is something to have your own adviser approve, since what counts as a relationship depends on your professional rules and the facts. The tutorial on AI client intake for law firms covers the online side of the same problem.

Five things the intake agent must never do

• Say whether a caller "has a case", what a claim is worth, or what outcome is likely.
• Quote a fee unless the firm has approved a fixed published fee for that service, and then read it exactly.
• Promise a callback time it cannot guarantee. "Within one working day" is safer than "within the hour".
• Take payment details, passwords, identity numbers or documents.
• Confirm or deny that the firm acts for, or has ever acted for, any named person. That would itself disclose client information.

The last point catches people out. A caller asks, "Do you represent [other party]?" The agent's answer is the same every time: "I can't discuss who the firm does or doesn't act for. A lawyer can talk to you about that." Script it and test it.

Conflict-check handover

The agent does not run the conflict check. It gathers what the person running it needs and passes it on cleanly. Names are the weak spot in any voice system, so the form below treats spelling as a field in its own right.

INTAKE HANDOVER (illustrative, filled in)

Call received: Tue 14:32, duration 4m 10s
Caller: [first name] [surname] (surname spelled by caller, read
        back letter by letter, confirmed)
Phone: 555-0142 (illustrative)   Email: confirmed by SMS link tap
Safe to leave voicemail: no   Best time: after 18:00

Matter type (caller's words): "my business partner wants out"
Category chosen: Business / partnership dispute
Urgency flag: none stated

Other parties
1. [full name] (spelled by caller, read back)  role: business partner
2. [partner's company] (caller unsure of exact name)  role: company
3. No other names given

Key dates mentioned: "the partnership agreement is from 2019";
 caller said a letter arrived "last Friday"

Consent to record: yes
Caller asked about fees: yes - told a lawyer will cover this
Caller asked "do I have a case": yes - refused per script

Flags for human: company name uncertain, letter received recently
Next step promised: call back within one working day

Notice the entries a conflict clerk can act on directly: each name with a spelled version, a role beside each name, and an explicit "company name uncertain" flag instead of a guess. A generic summary such as "partnership dispute with former partner" would have forced someone to ring back just to start the check.

Capturing names reliably

Voice transcription mishears unusual names, so build in two safeguards. First, the agent asks the caller to spell every surname once and reads it back letter by letter. Second, it sends an SMS or email link where the caller can type the names, with the typed version winning if it differs from the transcript. Both add a minute but remove the most expensive error in the process: running a conflict check on a name that was spelled wrong.

Keep the data minimal. Names, roles, matter category and dates are enough to start. A detailed account of events does not need to pass through the agent at all.

Per-minute platforms versus per-call receptionists

There are two buying routes, and they price differently.

Build on a voice-agent platform. Retell lists pay-as-you-go pricing of $0.07 to $0.31 per minute depending on the voice and language model you choose. Its page breaks the minute into parts: voice infrastructure at $0.055, a text-to-speech voice from $0.015, a language model from a fraction of a cent up to well over $0.10, and telephony at $0.015. Optional add-ons include PII removal at $0.01 a minute and safety guardrails at $0.005. Vapi charges a hosting fee of $0.05 a minute on its usage plan, with the speech-to-text, language model and voice costs passed through at cost; its page estimates $82 to $129 a month for 1,000 minutes. A phone number is a few dollars a month.

Buy a packaged AI receptionist. Smith.ai's AI receptionist plans price per call: a free tier with 25 calls a month and $3 a call after that, then a Pro plan at $150 for 75 calls ($2.00 a call), $270 for 150 and about $500 for 300 (about $1.67 a call), with calls beyond the bundle at $2.17 to $2.50 each. Its human-first service for comparison starts at $300 a month for 30 calls, with $11.50 for each call beyond that, and lists a $0.50-a-call add-on in which its receptionists check new callers for conflicts against your CRM. Prices change, so treat these as November 2026 list prices and check the pricing page before you budget.

A four-lawyer firm's month, priced four ways

Illustrative figures for a four-lawyer firm. It takes 220 inbound intake calls a month, of which about 45 percent (99 calls) arrive outside office hours. Average intake call length is 4.5 minutes, so the agent handles roughly 990 minutes.

The sum for the first row: 990 minutes at $0.164 is $162.36. The premium row is 990 at $0.31, or $306.90. For the packaged AI route the per-call rate depends on the bundle you buy: the 150-call tier at $270 plus 70 extra calls at $2.30 comes to $431, and the 300-call tier costs about $500 with room to grow. For the human-first service, the 300-call plan at $2,100 is cheaper than the 90-call plan plus 130 extra calls at $10.50, which would come to $2,175.

The cheapest route is not automatically the best. The platform route is cheapest per minute but needs someone to build, test and watch it, which is hours of work the table does not show. A packaged product costs more per call and gives you a ready-made dashboard and support. If you have no one to own the build, price in outside help; what an AI consultant costs a law firm sets out what to expect.

Hidden costs

• Review time. If a paralegal spends ten minutes a day checking handover forms, that is about three hours a month before any callback.
• Transfers. Warm transfers to a lawyer during office hours use telephony minutes on both legs.
• Testing minutes. Your test calls count as billable usage on most platforms.
• Retention and security options. Some platforms put data-retention controls, recording opt-outs or signed data-processing terms on higher or enterprise plans. Vapi, for example, keeps raw call data for 14 days on its pay-as-you-go plan and includes zero data retention only with a paid package, from $29 a month. Read what your plan includes against what your confidentiality duties require.

Call types the agent must send to a person straight away

Transfers should be rule-based, not judgement-based. Write the triggers as plain phrases:

• A hearing, deadline or limitation date within seven days (or whatever window your firm sets).
• The caller says they have been arrested, are in custody, or a family member has been.
• The caller reports an immediate safety concern, with emergency language first.
• The caller says they are already a client, or asks for a named lawyer. The agent takes a message and flags the lawyer, and does not discuss the matter.
• The caller is another lawyer, a court or an opposing party. These calls do not enter the intake flow at all.
• The caller is upset or asks three times for a person.

Outside office hours, a transfer may not be possible. The script then offers an urgent-callback path with an on-call number or a pager rule the firm already uses, and says honestly when a person will respond.

A test call where the court date lacked a month

An illustrative transcript fragment from a test call, with the problem marked.

Caller: My landlord is evicting me and I've got a court date on the 19th.
Agent: I'm sorry to hear that. Court dates are time-sensitive, so I'm
       going to flag this as urgent for the team. Can I take your name?
Caller: Is it likely I'll lose?
Agent: I can't advise on that, but a lawyer will talk it through with you.
       Could you spell your surname for me?

This is the behaviour you want: the date triggers urgency, and the advice question is refused politely without a lecture. The fix needed in the full test was that the agent logged "the 19th" without a month. The prompt now says: "When a caller gives a date without a month or year, ask for the full date and read it back." Gaps like that are where a missed deadline begins.

An afternoon of test calls with colleagues as callers

Do not go live on a good demo. Run at least twenty calls against a written script, from a second phone, with colleagues playing callers. A shorter model of the same routine, written for a trades business, is in the plumbers' 15-call test; the legal version needs these extra cases:

1. A caller with a surname the system will mishear, spelled slowly.
2. A caller who gives the other side's name differently from how they later spell it.
3. A caller who immediately begins telling a long, detailed story.
4. A caller who asks "Do you act for [named company]?"
5. A caller who asks for a quote and refuses to leave details until they get one.
6. A caller who says "I'm already a client of [lawyer's name]".
7. A caller who says there is a hearing tomorrow.
8. A caller who says they are frightened of someone and cannot talk freely.
9. A caller who refuses recording.
10. A caller on a poor line, with background noise.
11. A caller who asks, "Am I speaking to a real person?"
12. A caller who gives card details unprompted.

Score each call pass or fail on four points: disclosure made, no advice given, handover fields complete, correct transfer or callback. Do not go live until every call passes on all four. Re-run the whole set after any change to the script, voice or model, because a small prompt edit can undo an earlier fix.

Scripting the awkward moments

Most of the risk sits in six or seven sentences, so write them out rather than leaving them to the model. Three examples, in the wording the agent should use:

• Caller asks for a price. "The firm explains its fees to each person on the callback, and a lawyer can tell you how they're charged. Shall I take your details so they can do that?" If the firm publishes a fixed fee for a service such as a simple will, the agent reads that exact line, nothing more.
• Caller asks whether they are speaking to a person. "No, I'm an automated assistant for [firm name]. I can take your details now, or I can arrange for a person to call you back. Which would you prefer?" Never dodge this one.
• Caller gives privileged detail anyway. "Thank you. I'm going to stop you there, because I only take contact details and the basics. The lawyer will want to hear the rest from you directly." The transcript may still hold the sentence the caller said, so set retention short and restrict who can read it.

Keep the whole script in one document the partners have signed off, with a version number. When something changes, such as a new practice area or a different emergency rule, update the document first and the agent second, so the firm always knows what its phone line was told to say on a given date.

Data and outage questions for a voice-agent vendor

Whether you build or buy, put these in writing and keep the answers:

• Where are recordings, transcripts and call metadata stored, and for how long? Can retention be set to zero or a short period?
• Is call content used to train any model? Is that off by default or only on a higher plan?
• Which sub-processors (speech, language model, telephony) touch the audio, and can you see the list?
• Can you opt out of recording and still get transcripts, or the reverse?
• Does the plan include a data-processing agreement suitable for client confidentiality, and who signs it?
• What happens during an outage? Is there a fallback that forwards calls to a person?
• Can you export all your data and delete it on leaving?

If any answer is vague, treat it as a no. The firm's duty of confidentiality does not shrink because the data sits with a supplier.

Two firms, two set-ups

The same technology suits different firms differently. A single-lawyer practice that misses calls while in meetings may only need the greeting, a short contact-and-matter capture, and an email summary within five minutes. Keep the flow short, skip the conflict form and let the lawyer run checks at the day's end.

A six-lawyer firm with a conflicts clerk should send the full handover form straight to the clerk's queue and block the lawyer's diary slot only after the check clears. The route there is usually an integration with the firm's practice management system or, to begin with, a shared inbox with a fixed subject line so nothing is lost. If you are still deciding whether the phone is the right first project, what small law firms should automate first weighs it against the alternatives.

Five intake-line numbers to track after launch

For the first month, read every handover and listen to a sample of calls. Track:

• Handover completeness. What share of forms have all the names with spellings and a matter category.
• Callback time. Median time from call to a person's first contact, compared with what the script promised.
• Abandonment. Callers who hang up before the details stage, which usually means the greeting is too long.
• Escalation accuracy. Calls that should have transferred and did not, which should be zero.
• Conversion. Share of intake calls that become instruction letters, compared with the same count before the agent. Do not compare to voicemail alone, as callers who hang up on a machine were never in your count.

Judge value on the matters that were missed before, such as the after-hours calls. If 99 calls a month arrive outside hours and the agent captures usable details on 80 of them, the firm has 80 callbacks it previously chased from voicemail or lost. The fee value of one of them is your own number to put in, and it will outweigh a few hundred dollars of phone cost in most practice areas.

If you would like help working out what your firm's line should and should not do, my AI implementation consultation covers exactly that kind of decision. For the broader plan, the small law firm 90-day plan shows where a phone agent sits in the sequence.

Questions firms ask about AI intake calls

Can the voice agent tell a caller whether they have a case?

No. Assessing a matter is legal advice, and it belongs to a lawyer who has checked conflicts and agreed to act. The agent should collect facts, explain what happens next and book the callback. Script a polite refusal for questions such as 'Do I have a case?' so the agent never improvises an answer.

Does the caller become a client when the agent answers?

The script should say plainly that no lawyer-client relationship exists until the firm confirms it in writing. Whether a relationship has formed in law depends on your professional rules and the facts, so ask your firm's compliance lead or professional indemnity adviser to approve the exact wording before you go live.

Should the agent record the call?

Only if you have a reason, such as quality review or an accurate matter note, and the caller is told at the start. Call-recording consent rules differ between places, so confirm the rule that applies to your callers. Many firms keep a transcript and summary and set short retention on the audio.

What should it do when a caller is in distress or danger?

Put emergency language first. If someone says they or another person is in immediate danger, the agent says to contact the emergency services now and offers to transfer or notify a human. Test these phrases on every test run, because they are the calls where an unscripted answer does the most harm.

How many test calls are enough before going live?

Run at least 20 scripted calls across your main matter types, including three awkward ones: a caller who gives an opposing party's name in a different spelling, a caller who wants advice immediately, and a caller who is upset. Do not go live until every handover reaches a human with the right fields.

Further reads

• How to Set Up an AI Phone Agent for After-Hours Calls — General set-up for calls that arrive outside office hours.
• AI Call Summaries: Log Every Phone Enquiry in Your CRM — Log each enquiry and summary in your records.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Compare per-seat and pay-as-you-go pricing models.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Choosing between general and legal-specific tools.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Draft the follow-up letter once a matter is accepted.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the firm is ready before adding more AI.
• Best AI Tools for Attorneys: Seven Picks for Research, Drafting and Intake — Seven AI tools for a small law practice, mapped job by job, with a worked example and the checks to run before you pay.
• Best AI for Immigration Lawyers: Intake to Filing, Stage by Stage — A nine-stage map of an immigration case with the AI tool type, the data check and the manual step for each, plus a costed worked example.
• Best AI for Solo Lawyers: A Stack That Fits One Person — Assistant, practice-management AI and contract tool for one lawyer: three priced stacks from $120 to $275 a month, and what to leave out.
• Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist — Which AI to use for bodycam transcripts, research, drafting and intake in a small defence practice, with list prices and a vendor data questionnaire.
• Conflict Check Software for Small Law Firms: Where AI Helps — Where AI helps in a conflict check, where it must not decide, and what a partner signs off, with a worked example and a 20-minute test of your own tool.
• Lawmatics Merlin: Is It Worth Adding Qualify, Engage and Copilot? — What Merlin Qualify, Engage and Copilot do for law-firm intake, what each costs, and a four-week test on real leads before you pay.
• Legal Client Intake Software: What to Compare Before You Choose — How a small law firm should compare intake software with AI: the AI's actual job, the conflict-check step, integrations and a 20-enquiry test.
• Live Receptionist for a Small Law Firm: Cost and When AI Fits — What a live receptionist costs a small law firm against an in-house hire or AI, with plan prices, an intake script and an out-of-hours handover.
• Personal Injury Lawyer Chatbot: Intake Questions, Limits and Cost — An intake script, disclaimer wording, handoff triggers and monthly cost bands from named vendors for a personal injury firm's website chatbot.
• AI Receptionist for Family Law Firms: Handling Callers in Crisis — Safety triggers, handoff rules, scripts, recording consent and a monthly cost range for a family law firm's AI receptionist.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Retell AI and Vapi pricing pages and Smith.ai plan pages, read November 2026; call volumes and figures in examples are illustrative.