Family-law intake schema: where the conflicts hide and what the AI agent has to ask early · NeuraVoice

It is 9:07 p.m. on a Tuesday. A woman dials a family-law firm in suburban Chicago. The website has an after-hours AI voice intake. She does not say "I want to file for divorce." She says, "I need to talk to someone, my husband took the kids to his mother's and he is not answering my texts and there is supposed to be an order."

That sentence is the entire problem with family-law intake. There are at least four signals in it. There is a possible active custody interference. There is a possible existing protective or temporary custody order. There is emotional distress. There is no schema yet, no name, no jurisdiction, no matter type. An AI agent that opens with "Can I get your full legal name and date of birth?" is going to lose the call inside ninety seconds. An AI agent that just listens warmly and never gets to the schema is going to leave the firm with an unusable voicemail at 10 p.m.

This post is a working overview of how family-law intake schemas actually have to be built when an AI voice agent is taking the first call. It is not legal advice. Conflict-of-interest rules, ABA Model Rule 1.18 obligations, and state-specific bar guidance all govern what your firm can do with prospective-client information, and you should treat what follows as a structural map, not a compliance opinion.

The conflict surface in family law is wider than any other practice area

In personal injury you usually run conflicts on the caller, the at-fault party, and any insurer. In criminal defense you run on the defendant and possibly co-defendants. In family law you run on the caller, the spouse, the spouse under a maiden name, the spouse under any prior married names, every child by full legal name and date of birth, any family business or LLC the couple owns, any family trust either side is a beneficiary of, and frequently the financial advisor or estate planner who is mid-engagement with both parties.

That last category is the one most intake schemas miss. Mid-divorce, the wealth manager who has been advising "the family" for nine years is now adverse-by-proxy, and if your firm has ever taken a referral from that advisor on the other spouse's matter, you have a problem before the consult ever happens. The conflict-check field set has to capture institutional names, not just human ones. Clio Grow, MyCase, Smokeball, and PracticePanther all support custom conflict fields, but the default intake forms shipped with each of them are written for general civil practice. Family-law conflict screening is a custom build, not a configuration toggle.

Children's names and ages are the other place schemas leak. If the caller has a fourteen-year-old named Jacob and the firm represented Jacob's father in a probate matter three years ago, that is a conflict surface that only shows up if the AI captured the child's full name. "Two kids, ages 8 and 14" is not enough. The schema has to ask for full legal names and dates of birth for every child, even when the caller hesitates, and it has to explain why ("for our conflict check, this is confidential") so the caller does not assume the agent is being intrusive.

Protective-order detection happens in the first ninety seconds or it does not happen

There is one rule that overrides every other rule in family-law intake: if the caller is in active danger, the AI escalates immediately. Not after the schema. Not after the jurisdiction question. Immediately.

The detection signals are not subtle once you list them. The caller says they are "hiding," "in the car," "at a friend's," "scared." They say "he doesn't know I'm calling." They say "there's already an order" or "the order ran out last month." They say a child is missing or has been taken. Any one of those is a hard interrupt. The AI's response is to acknowledge, ask whether the caller is safe right now, and if there is any ambiguity, route to a human on-call attorney or, if the firm's protocol says so, prompt the caller to call 911 and stay on the line.

Building this is not a content problem. It is a routing problem. The intake agent has to have a real escalation path, with a real phone number that rings a real human, and the firm has to staff that path. An AI that says "let me transfer you to someone who can help" and dumps the caller into a voicemail box at 9 p.m. is worse than no AI at all. If your vendor cannot show you the live-transfer behavior under a simulated protective-order signal, the schema is incomplete regardless of how good the rest of it looks.

No-fault is the rule, so intake gates on jurisdiction and complexity, not grounds

Roughly seventy percent of US divorces are filed as no-fault, and every state now permits no-fault filing in some form. The practical consequence for intake is that asking "what are the grounds for divorce" is the wrong gating question. It is not what determines whether the matter is worth a consult. What determines that is jurisdiction, asset complexity, and custody complexity.

Jurisdiction is two fields, not one. State of residence and state of intended filing are sometimes different, especially when one spouse has recently moved or when military service moved the family across states. Most states have a six-month residency requirement for filing, with variations. The schema needs both fields and the duration in the filing state, because a caller who moved to Texas four months ago cannot file there yet, and that is a fact the firm needs before the consult, not during it.

Asset complexity is a flag, not a balance sheet. The schema asks: any business owned by either spouse, any trust either spouse is a beneficiary of, any real estate beyond a primary residence, any retirement accounts requiring a QDRO. A "yes" to any of those routes the matter to a more senior attorney and a different fee structure. A "no" across the board may route to associate-led representation or to a flat-fee track if the firm offers one.

Custody complexity is also a flag. Are there minor children. Are any of them from prior relationships. Is there a current custody order from a prior matter. Is either parent planning to relocate. Is there a special-needs child. Each of these changes the staffing and the fee, and each is a discrete yes/no the AI can capture cleanly.

Marital duration matters because of social security, not because of property

Property division is largely state-driven and the schema does not need to dig into it on the first call. Marital duration matters for a different reason: ten years. Under federal Social Security rules, a divorced spouse is eligible for derivative benefits on the working spouse's record only if the marriage lasted at least ten years. A caller who has been married nine years and ten months is in a meaningfully different posture than a caller married eleven years, and the conversation about timing the filing is one a competent family-law attorney is going to want to have at the consult. The intake schema captures duration in months, not "about ten years."

Pro se filings happen in roughly fifteen percent of US divorces, and the schema has to ask

Self-represented filings are common in family law in a way they are not in PI or business litigation. A meaningful share of callers have already filed something, on their own, and are calling because they hit a wall. Maybe the petition was rejected, maybe a temporary order was entered against them, maybe they have a hearing in nine days and no idea what to do.

The schema has to ask: have you already filed anything in this matter, and if so, what court and what case number. The case number unlocks the docket. Without it, the intake notes are missing the single most time-sensitive piece of information on the call. With it, the firm walks into the consult already knowing what was filed, what the response deadline is, and whether there are pending motions. PracticePanther and Smokeball both support docket-linked matters; Clio supports it through integrations. The AI does not need to look up the docket on the call. It just needs to capture the case number and the court correctly.

Tone is a feature, not a finish

In PI intake the schema-first approach works because the caller is usually calm enough to answer questions in order. In family law that is rarely true. Callers are mid-crisis, often calling outside business hours specifically because they could not sit with the panic until 9 a.m. the next day. A schema-first opening reads as cold, the caller disengages, and the firm gets a partial intake.

The pattern that works is brief acknowledgment, then schema, then more acknowledgment when the schema surfaces something hard. "I hear you. That sounds like a lot to be carrying right now. Can I ask a few questions so I can get you to the right attorney as quickly as possible?" then schema, then when the caller mentions the protective order: "Thank you for telling me that. Are you safe where you are right now?" The acknowledgments are not filler. They are gates that keep the caller engaged through the schema.

This is the part that vendor demos tend to fake. Anyone can script empathic phrases. The question is whether the agent's empathy phrases are deployed conditionally on caller-state signals, or whether they are sprinkled in regardless. The first is intake. The second is theater.

Five common mistakes in family-law intake schemas

Conflict fields scoped to humans only, missing family LLCs, trusts, and shared advisors.
Children captured as a count and age range instead of full legal name and date of birth per child.
No protective-order escalation path with a real human endpoint, or one that fails after hours.
Asking about grounds for divorce as a gating question when seventy percent of filings are no-fault and grounds are not what determines fit.
Treating self-filed status as an edge case instead of a standard schema field, missing the case number that unlocks the docket and the deadline.

Five vendor-evaluation questions specific to family-law intake

Show me the call transcript when the caller says "I have a protective order against him" in the first thirty seconds. What does the AI do, and where does the call go?
How does the schema capture full legal names and dates of birth for every child, and how does it handle a caller who declines to provide them?
What conflict fields does the schema capture beyond the two spouses? Specifically, family businesses, trusts, prior counsel, and shared financial advisors.
How does the agent capture filing status, court, and case number for callers who are already pro se in an active matter?
How does the agent balance acknowledgment and schema progression when the caller is in distress, and can you show me a recording where it goes well and one where it goes poorly?

The closing assertion

Family-law intake is the practice area where the schema and the tone are most in tension, and where getting either one wrong costs the firm the matter. PI firms can absorb a cold intake because the caller is comparison-shopping and will tolerate friction. Family-law callers will not. They will hang up, call the next firm on the list, and the firm with the warmer first thirty seconds will get the consult regardless of which one has the better attorneys.

So the contrarian point is this: in family law, the intake schema is not the moat. The first ninety seconds before the schema starts is the moat. If your AI cannot detect a protective-order signal, acknowledge a child being taken, or pause for a caller who starts crying, the rest of the schema is decoration on a call that already ended.

If you want to see how a family-law-tuned intake agent handles the first ninety seconds, book a call with the team, start a free trial, or look at the pricing page.