AI Receptionist for Family Law Firms: Handling Callers in Crisis

Treat crisis as a handoff problem, not a conversation: the receptionist asks whether it is safe to talk, tells anyone in danger to call emergency services, collects only a name, a safe contact method and the deadline, then alerts a person. It never advises, predicts outcomes or confirms who is a client. Budget roughly $75 to $265 a month.

Most failures come from the receptionist being helpful. A frightened caller asks "can he take the children?" and a friendly assistant answers. The firm then owns a legal opinion it never gave, delivered to someone who may act on it tonight. Have a domestic-abuse specialist read your scripts before launch. This tutorial gives the structure, and they know what works for people who are in danger.

Follow me on Instagram@sagnikteaches

Sort calls into four kinds before writing a word

Pull two weeks of voicemails and call notes, or ask whoever answers the phone to tally, and sort every call into one of four colours. The colours drive everything else: the greeting, the questions, the alert and the promise. Set aside about an hour for this.

Connect on LinkedInSagnik Bhattacharya

An invented tally of 80 calls over two weeks might read 3 red, 9 amber, 28 yellow and 40 green. Three red calls in 80 is under 4 per cent, which sounds small until you remember that one red call a month is enough to design for. The yellow calls are the most common and the ones an automated receptionist handles worst, because its habit is to start collecting details while the caller is still trying to find words.

Subscribe on YouTube@sagnikteaches

Safety triggers that end intake and start a handoff

A trigger is a phrase or meaning that makes the receptionist stop its normal flow and run a different script. Write them from your own calls, with a lawyer who takes these enquiries. Most AI receptionists match by meaning rather than exact words (test that yours does), but real examples teach the assistant what to look for, so include the ugly, indirect ones. Set aside about two hours.

RED, stop intake and give the emergency instruction:
"he's outside", "she's here now", "I'm hiding", "he said he'd
kill me", "he's got a gun", "he's taken the kids", "the police are
here", "I can't do this any more", "I want to end it".

AMBER, minimal intake and urgent alert:
"I've been served", "hearing is tomorrow", "the hearing is on
Thursday", "she's taking the kids abroad on Friday", "he didn't
bring the children back", "he's emptied the account", "the house
is being sold next week".

YELLOW, slow down, acknowledge, offer a call-back:
"I don't know what to do", "he left me this morning", "I just
found out", "I can't stop crying", "I need to talk to someone".

OTHER, never confirm or deny:
"Has my wife called you?", "Is [name] a client of yours?",
"I'm calling for my sister", "I'm her husband".

Two design choices matter. First, accept false alarms. A red flag that turns out to be nothing costs one phone call; a missed one can cost far more. Second, put the OTHER group in the same list. In family work the person asking whether someone has been in touch may be the person the caller is afraid of, so the receptionist never confirms or denies that anyone is a client, and says the same thing whether or not it is true.

The first twenty seconds, scripted

The opening decides whether a frightened person stays on the line. It must disclose that the caller is talking to software, mention recording, give the emergency instruction once, and ask whether it is safe to talk before asking anything about the case. It must not open with "How can I help you today?", because a caller in danger may answer in detail. Here are six short scripts to adapt, with the placeholders filled in by your firm and checked by a specialist:

OPENING
"Thanks for calling [Firm]. I'm an automated assistant, and calls
are recorded so a lawyer can follow up. If you or anyone with you
is in danger right now, please hang up and call [emergency
number]. Otherwise, is it safe for you to talk?"

IF "NO", WHISPERING OR UNCLEAR
"Thank you. I'll end this call now so you're not overheard. If
you're in danger, call [emergency number]. You're welcome to call
us again whenever it's safe. We won't call you back unless you
tell us a safe way to reach you."

RED, DANGER NOW
"I'm sorry you're going through this. I'm an automated assistant
and I can't help in an emergency, so please call [emergency
number] now. I'm alerting a member of our team as well. If you
can't talk, you can hang up, and you're welcome to call again
whenever it's safe."

RED, CALLER MAY HARM THEMSELVES
"I'm really sorry you're feeling like this. I'm an automated
assistant, so I can't give you the support you need right now,
but people can. Please call [emergency number] now, or [crisis
line name and number, checked by the firm]. I'm letting a member
of our team know you called."

YELLOW, DISTRESS
"I'm sorry, that sounds hard. I'm an automated assistant, so I
can't give advice, but I can ask a lawyer to call you back. There's
no rush on my side. May I take your first name and the best way
to reach you?"

AMBER, DEADLINE
"Thank you for telling me. I'll mark this urgent. I need three
things: your name, a number where it's safe to leave a message,
and the date of the hearing or the court reference if you have
it. Is it safe to leave a voicemail or a text on that number?"

Each line has a job. "Is it safe for you to talk?" gives the caller a way to say no without explaining. Ending the call when the answer is no protects someone who may be overheard. The red script contains exactly one instruction and one promise, that someone on the team has been told, and it never implies the firm is an emergency service. The self-harm script differs from it in one way: it never ends the call or tells the caller to hang up. It gives the emergency number and the crisis line your firm has checked, says a person has been told and, if the caller keeps talking, repeats those two numbers kindly rather than asking what they intend to do. Have a mental-health professional approve its exact wording before it goes live. The yellow script slows the pace deliberately; "no rush on my side" is the line that stops the assistant sounding like a form. The amber script asks only for what a lawyer needs to act on, plus the safe-contact check. After any of them, close with the same sentence: "A lawyer will review this and call you back [time promise]. I can't say whether the firm can take your case until they've spoken to you."

Handoff to a person: who is alerted and how fast

The alert is the part that does the real work, so give it a ladder with timings and a named last resort. The firm cannot promise a rescue, and should not let the script suggest one; the alert exists so a person can follow up quickly when the caller has said contact is safe.

The alert text should be short enough to read on a lock screen and specific enough to act on. Two illustrative examples, one for each of the first two rows:

URGENT-RED 21:42. Caller said someone is outside and she is
hiding. Emergency instruction given. No details taken. No safe
call-back confirmed. Do not call back unless she rings again.
Recording: [link]

URGENT-AMBER 08:15. [First name], safe number [number], OK to
leave voicemail: yes. Hearing Thursday 10:00, [court]. Child not
returned after contact weekend. Wants a call before 09:00.
Recording: [link]

Notice what the red alert leaves out: the caller's narrative, which the assistant should not have collected, and any instruction to ring her back. The tutorial on writing call scripts and escalation rules for an AI receptionist covers the general mechanics of rules like these; the family-law specifics are what this page adds.

What the receptionist must never say to a caller in crisis

Write these into the instructions as a closed list, with the replacement line beside each. The aim is a receptionist that is kind in tone and empty of content.

• No legal advice and no predictions. Never "you'll get the house", "he can't take the children" or "you should apply for a protective order". Say: "A lawyer can go through your options. I can arrange a call-back."
• No instructions about leaving, moving the children or contacting the other person. Apart from the emergency instruction, say: "That's a decision to make with a lawyer's advice."
• No statement on whether the firm will act. Never "we can definitely help" or "that's not something we do". Taking a case needs a conflict check and a lawyer's judgement.
• No minimising and no hurrying. Never "calm down", "I'm sure it's not that bad" or "it'll be fine". Avoid "I understand how you feel", which software cannot honestly claim; "I'm sorry you're dealing with this" is truer.
• No counselling and no request for the story. "Tell me what happened" produces a narrative that then sits in a vendor's transcript. Say: "You don't need to give me the details. The lawyer will ask when you speak."
• No confirmation or denial that anyone is a client or has called, to anyone, in the same words every time.
• No detail in voicemails or texts until the caller has said contact is safe, and then only a neutral request to ring the firm.
• No promised times the firm cannot meet. "A lawyer will call within ten minutes" is a promise; "a lawyer will review this and call you back today" is one you can keep.
• No pretending to be a person, a lawyer or a counsellor.

The tutorial on a family law intake questionnaire with urgency triage covers what the lawyer's team asks next, once a person has taken over from the receptionist.

Safe contact, and the call from the other side

Ask every caller, after the name, whether it is safe to leave a voicemail or send a text, and record the answer in the alert. Make "no voicemail, no text" the default until the caller says otherwise, and make staff obey it. Also think about what the caller's phone will show: a call log or bill that says "[Firm] Family Law" may be a risk where a partner checks the device. Asking "Would our firm's name appearing on your phone be a problem?" costs ten seconds and may prevent a serious mistake.

The other side of the problem is the caller who is not who they seem. The person asking "Has my wife called you?" may be the person she fears, or someone trying to learn whether she has taken advice. The receptionist should give the same answer whether or not it is true: "I can't confirm whether anyone has contacted this firm. I can take a message for the office manager." The same rule protects the conflict check, which needs names collected on every call and compared inside the firm, never shared outward.

Existing clients need a different route. If a caller says "I'm a client of [lawyer]", send the alert straight to that lawyer rather than the intake pool, and say only "I'll let them know you called". The receptionist does not need to confirm they are a client in order to route the call, and it should not.

Recording and disclosure when the caller is frightened

The opening script already says the call is recorded. For a frightened caller that sentence may not register, so a short disclosure is a floor, not a defence. Where the law requires every party to consent to recording, ask your regulator or your indemnity insurer whether the opening line is enough, or whether the assistant must pause for a yes. Do the same for the wider question of whether a prospective client's call reaching an AI vendor fits your duty of confidentiality, and get the answer in writing before launch.

Next, decide what you keep. You have three options: keep audio and transcript for a set period, keep only the structured summary, or switch recording off where the vendor allows it. The Upfirst and Rosie pages list recordings as part of every plan and Smith.ai lists recording within its AI plans, but none of the pages I read says you can switch recording off for a single call, so ask each vendor before you rely on it. One legal-specific vendor, Reflekt Legal, says its data is not used to train public AI models and that retention limits follow the firm's policy; as with any such claim, get it in writing for your plan. The general method of weighing these promises is in the tutorial on AI legal answering: cost, confidentiality and urgent handover.

Monthly cost for a small family firm

Price it for a four-lawyer firm taking about 120 calls a month at three minutes each, which is 360 minutes (illustrative figures). The table prices the cheapest published combination of plan and overage at that volume, from each vendor's own page in November 2026, and shows what each plan lists for handing a call to a person.

That is where "about $75 to $265 a month" comes from. Two things the table hides. First, the cheapest plans are the ones with the least handoff: for a firm whose first duty is getting a flagged call to a person, Upfirst and Rosie, which list transfer at modest prices, deserve a look before Dialzara's entry plans. Second, the law-specific products I checked (Reflekt Legal, Coregent and Call Setter) publish no prices. Call Setter's page says only that its service costs "under $400/month per agent" as a comparison claim, so treat all three as quote-only and put the same questions to them.

For comparison, Smith.ai's human plans would cost $1,125 for the same 120 calls: the 90-call Basic plan at $810 plus 30 calls at $10.50, which is $315. That is between four and fifteen times the AI range, and it buys people who handle ambiguity better. Many firms therefore keep a human route for flagged calls and use AI for first contact and out-of-hours messages. The general pricing arguments sit in the tutorial on AI voice agents for lawyers and safe handling of intake calls.

The year-one total, priced in full

Choose Rosie Scale at $149 for its warm transfer. The subscription over a year is 12 × $149 = $1,788. The set-up time from the stages above is about nine hours: one for sorting calls, two for triggers, two for scripts, one for the alert wiring and three for role-plays. At an internal $80 an hour that is $720 once. Reviewing transcripts takes four hours a month for the first three months and two hours a month after that: 12 + 18 = 30 hours, or $2,400. Add them up: $1,788 + $720 + $2,400 = $4,908. A specialist's review of the scripts is extra, so ask for a quote.

If a new matter brings $3,000 in fees (use your own figure), two extra matters a year cover it. Count the calls you currently miss in a fortnight before you commit, because a firm that rarely misses a call gains little from the subscription and a lot from the safeguards.

Eight role-plays to run before go-live

Ring the assistant from a phone it has never seen, using someone who has not read the scripts, and run each of these twice. Allow about three hours. Ask your specialist to listen to two of the recordings, because they will hear tone problems that a transcript hides.

Rerun the full set after every change to the instructions, because a tweak made to fix one role-play can quietly break another. For the first month, read every red and amber transcript the same day and sample five yellow ones a week, noting any sentence where the assistant said more than it should.

The voicemail that reached a shared phone

This failure is invented, but every part of it is plausible. A caller said she could be reached on her mobile and, asked nothing about safety, was promised a call-back. The next morning a staff member rang, got her voicemail and left the default message: "This is [Firm] family law returning your call about your separation." The phone was on a shared family plan, and her partner heard it.

Nothing in the script was unkind. The fault was the default: leaving voicemail was allowed unless forbidden. The fix had three parts. The receptionist now asks "Is it safe to leave a voicemail or a text on that number?" on every call and records the answer in the alert. The default flips to no voicemail and no text until the answer is yes. And the message left when the answer is yes is neutral: "Please call [Firm] on [number] when it suits you." Test 8 in the table covers the text version of the same risk, and should be rerun whenever staff change the call-back procedure.

Questions family law firms ask about crisis calls and AI

Should the receptionist ever stay on the line with a caller in danger?

No. It cannot summon help, assess risk or steady a caller the way a trained person can, and it should not imply that it can. Its job is one clear instruction: contact emergency services now. After that it alerts a person straight away, and ends the call if the caller says it is not safe to talk, so they are not overheard. The person decides how to follow up once the caller says contact is safe.

Can it tell a caller whether the firm can take their case?

No. Taking a case needs a conflict check and a lawyer's judgement, and a caller may act on a premature yes. It can say that a lawyer will review the details and call back, and it can collect the names needed for the conflict check. It must never say we will act for you, we cannot help, or that the matter sounds strong.

What if the caller says they want to hurt themselves?

Treat it as a red flag in its own right. The script should tell the caller to contact emergency services or a crisis line straight away, give helpline details your firm has checked, and alert the on-call person at once. The assistant must not attempt counselling or ask for detail, and the firm should have those lines reviewed by a mental-health professional.

Is a human answering service safer than AI for these calls?

For callers in danger, trained people handle ambiguity better, and they cost far more: Smith.ai's human plans list $300 for 30 calls and $810 for 90, against about $75 to $265 a month for AI plans at 120 calls. Many firms use AI for routine and after-hours first contact, with a human route for flagged calls. Decide by how often your calls turn out red or amber.

Further reads

• Family Law Client Communication With AI: Calm, Consistent Emails — Calm, consistent written updates after the first call
• Family Law AI Software: What a Small Practice Should Check Before Buying — What a small family practice should check before buying AI
• Setting Up an AI Answering Service That Never Misses an Emergency — Setting up an answering service that never misses an emergency
• AI Receptionist for Churches: Office Phone Cost and Pastoral Call Rules — Another sector where callers phone in distress
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Qualify enquiries out of hours without a live call
• Family Law Retainer Agreements: AI Drafts, Lawyer Sign-Off — The paperwork that follows once a lawyer accepts the case
• AI Receptionist for Law Firms: Cost and What It Must Never Say — Published prices for five AI receptionist services, a worked monthly bill for a small firm, and the sentences a legal front desk must never speak.
• Best AI for Family Law in Australia: Privacy, Court Rules and Cost — A family law shortlist with prices, what the Family Court's AI practice direction requires, and the privacy rules that keep client data out of free tiers.
• Best Legal AI for Family Law: What to Pay For and What to Skip — General assistants, practice-management AI, research and drafting tools and form automation compared by job, with a budget and trial plan for a small practice.
• 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.
• Best AI for Family Law UK: Form E Prep and Client Updates — A safe route to Form E bank statement summaries and calmer client updates: what to redact, which tools to use, what to check and what the SRA has said.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Upfirst, Smith.ai, Goodcall, Dialzara, Rosie, Answrr and CallJolt pricing and home pages; Reflekt Legal, Call Setter and Coregent home pages (all checked November 2026).