Case Studies · 10 min read
Family Law Missed Call Recovery Case Study
The Omni model: one managed AI Employee owns one recurring workflow; specialist employees add capacity around the same business context; your team keeps the judgment calls.
At a four-attorney family law practice we worked with last year, the managing partner noticed that intake conversion had dropped roughly 18% across two consecutive quarters. The firm's marketing spend was steady. Its Avvo rating was strong. Its reviews were solid. The cause wasn't reputation or lead quality. It was the phones. During the lunch hour, after 5 p.m., and across every weekend, calls were going to a full voicemail box. By the time intake staff returned messages the next morning, roughly a third of those callers had already retained another firm.
This is the specific problem we were asked to solve: how does a family law firm recover the calls it can't physically answer, without making the caller feel like they've been handed off to a robot, and without putting confidential information at risk in the process.
Why family law intake is uniquely exposed to missed calls
Family law is one of the most time-sensitive practice areas a consumer calls. Someone considering a divorce, a custody modification, or a protective order is often calling more than one firm in the same hour. They are not shopping for a logo. They are trying to find a competent adult who can talk to them before the panic fades.
Clio's annual Legal Trends Report has consistently shown that law firms, on average, fail to respond to a meaningful share of new client inquiries, and that small and mid-sized firms are hit hardest because they don't have after-hours coverage. The pattern is not unique to this one practice. It's structural.
Two compounding factors make family law worse than, say, estate planning or business law:
- The emotional clock. A caller exploring a prenup can wait a day. A caller with a court date in eleven days cannot. Without an urgency signal captured at the first touchpoint, intake staff spend the next morning triaging callbacks in the wrong order.
- The attorney-client privilege boundary. The first call is where the caller decides whether to disclose facts about their spouse, their children, or their finances. If the intake experience feels careless, they hang up and call the next firm on the list.
So the bar for "handling the call when no one is at a desk" is not just operational. It is a trust test.
The workflow we designed
We mapped the firm's existing intake process before touching any tooling. Here's what we found: two intake coordinators handled roughly 300 inbound calls per month, with a call volume distribution that spiked between 11:30 a.m. and 1:00 p.m., again from 4:30 p.m. to 6:30 p.m., and a baseline trickle on Saturdays. The firm had a phone tree and an overflow voicemail, but no structured way to triage what came back.
The deployed workflow has three entry points and one review point. Every caller ends up in the same case management record, regardless of which entry they used.
Entry point 1: Business hours, intake staff available. Calls route to a coordinator as before. No change to this path. We left it alone because it wasn't broken.
Entry point 2: Business hours, intake staff unavailable. The AI voice agent answers, identifies the firm by name, and asks three questions: the caller's name, the general nature of the matter (divorce, custody, prenup, modification, other), and whether there is a pending court date. If the caller mentions a court date, a protective order, or any urgency keyword in free response, the call is flagged for an immediate human callback. Otherwise, the agent offers two consultation windows from the firm's actual calendar (pulled in real time from the scheduling system), books the slot, and sends an SMS confirmation. The booking is held in a pending state until intake staff approve it on the next login, which typically happens within 15 minutes during the day.
Entry point 3: After hours and weekends. Same intake script, but the consultation booking is constrained to the next two business days, and the SMS confirmation explicitly says the firm will follow up by 10 a.m. the next business morning. Urgent flags bypass the booking flow entirely and page the on-call intake coordinator through an existing escalation channel the firm already used for after-hours emergencies.
The single review point. Every AI-handled call generates a transcript, a structured summary, and a recommended next action. These land in a daily queue for the senior intake coordinator to review before her first coffee is finished. Anything that looks miscategorized, anything involving a flagged urgency that wasn't routed correctly, and any caller who explicitly asked for a human is pulled into a follow-up queue.
Handoffs, review points, and fallbacks
The part of these builds that matters most is what happens at the boundaries. Here is how we set each one up.
AI to human handoff during the call. At any point in the script, the caller can say "representative," "lawyer," "attorney," or "person" and the call transfers immediately. If no staff is available, it goes to a live human queue with a callback promise rather than voicemail. We benchmarked this against the alternative of trapping the caller in the AI loop and there is no defensible reason to do that.
AI to scheduling system. The booking only writes to the calendar after a confirmation step where the caller presses 1 or says "yes." No silent bookings. We have seen too many implementations where the system double-books or books a slot the caller didn't actually agree to.
AI to case management. Every call writes a record to the firm's existing matter intake system with a clear "AI-handled" tag. This is non-negotiable for two reasons: it lets the firm audit every interaction, and it preserves the chain of custody for any privileged information that flows through the call.
Human review of every AI interaction. The daily review queue is the most important part of the system. It is also the part most firms try to skip. Don't skip it. The review catches misclassified urgency (a caller who said "I need help soon" without using a specific keyword), it identifies questions the script can't answer well, and it generates the monthly tuning list. In our experience, this is where the actual return on the system compounds. Without it, you have a phone tree with extra steps.
Failure modes we planned for. The agent fails open to a human callback if the speech recognition confidence drops below a threshold, if the caller stays silent for more than eight seconds, or if the caller's free response contains any of a curated list of high-stakes phrases. We also have a kill switch: a single Slack command from the managing partner sends every call back to the existing phone tree for the rest of the day. We have used it once, during a calendar sync outage, and we were glad it existed.
What changed after 90 days
We won't make revenue claims, because we don't have a methodology that lets us cleanly attribute new matter revenue to this system versus the firm's other intake improvements running in parallel. What we can report is what we measured:
- Inbound answer rate moved from approximately 60% to 96%, measured by calls that reached either a human or a completed AI interaction.
- Median time to first human touch on after-hours calls dropped from "next morning, whenever staff arrives" to under 10 minutes for flagged-urgent calls, and before 10 a.m. for routine matters.
- Intake coordinator hours previously spent on initial phone triage moved into pre-consultation preparation and conflict checks, which the firm had historically under-resourced.
- Missed-flag incidents (urgent calls that were not correctly flagged) in the first 30 days: 14. After tuning the urgency classifier against the first month's transcripts: 2 in the next 60 days, both of which were genuinely ambiguous at intake.
For context on why response time matters across professional services generally, the often-cited InsideSales.com / Oldroyd finding in the sales lead literature — that contacting a new lead within five minutes is roughly 21 times more effective than contacting them within 30 minutes — has been replicated in professional services contexts as well. McKinsey's work on automation in professional services has pointed to similar dynamics in client acquisition funnels. None of this is unique to law, but family law amplifies it because of the emotional-clock problem described above.
Implementation notes for firms considering this
A few things we would flag for any firm thinking about deploying a similar system.
Map the call before you map the AI. Spend a week listening to real calls before writing a single prompt. The actual questions callers ask, the words they use, and the moments they get frustrated will reshape your script more than any vendor demo will.
Confidentiality is a design constraint, not a feature. Decide upfront what the AI is allowed to capture, where transcripts are stored, who can access them, and how long they are retained. Family law callers will sometimes disclose things on the first call that they would not put in an email. Treat that information accordingly. We worked with the firm's outside counsel to write a short data-handling addendum before going live.
Approval gates are not optional. Every booking confirmation that goes to a caller should be reviewable by a human before it is treated as a firm commitment. Every summary that writes to the case management system should be tagged. Every escalation that bypasses the normal intake flow should generate an audit record. This is what makes the system defensible to the bar, to the firm's malpractice carrier, and to the partners.
Plan for the boring parts. The implementation timeline is roughly: two weeks of call mapping and script drafting, one week of integration with the calendar and case management system, one week of private testing with the firm's own staff acting as adversarial callers, and two weeks of parallel running where the AI handles overflow but the human path remains the default. Total elapsed time was about six weeks from kickoff to full deployment. The hardest part was not the tooling. It was writing the script in language that matched how the firm's actual callers talk.
Budget for ongoing tuning. Expect to spend roughly four hours per month reviewing transcripts, updating the urgency classifier, and adjusting the script based on new matter types. This is not a "set it and forget it" deployment. The firms that get the most out of these systems treat them like a junior team member who needs coaching, not like an appliance.
Frequently asked questions
Does the AI sound like a robot to callers?
It uses a natural-sounding voice model, but the more important factor is the script. We wrote the script to mirror how the firm's own intake coordinators actually talk on the phone, including the pauses and the way they confirm information. Most callers do not realize they are talking to an AI on the first turn. Some realize by the second turn and don't care, because the experience felt competent. A small number ask directly, and the agent tells them honestly.
What happens when the AI doesn't understand the caller?
The call transfers to a human. If no human is available, it goes to a live human queue with a callback promise. We did not build a loop where the AI asks the caller to repeat themselves more than once. That is the fastest way to lose a caller.
How is confidential information handled?
Call audio and transcripts are stored in the firm's existing case management system with the same access controls as any other client communication. We wrote the intake script to collect only what is needed to route and schedule: name, callback number, general matter type, and an urgency signal. We do not capture details about children, finances, or specific allegations on the first call. Those happen later, in a privileged setting with a human.
What does the human review process look like in practice?
The senior intake coordinator spends roughly 20 to 30 minutes each morning working through the queue from the previous day. She approves bookings, corrects misclassifications, and flags anything unusual for the managing partner. The monthly tuning meeting is a 60-minute call where we review patterns and update the script and classifier together.
How long does implementation actually take?
About six weeks from kickoff to full deployment for a firm with an existing calendar system and case management platform we already integrate with. Longer if we are integrating against a less common system or if the firm has unusual routing rules. We do not recommend compressing the timeline, because the parallel-running week is where the firm builds confidence in the system.
If your firm is losing calls during lunch, evenings, or weekends — or if your intake staff is spending more time chasing voicemails than preparing for consultations — the fix is usually less dramatic than it feels. It is a workflow problem with a workflow solution. We can walk through your specific intake path on a 30-minute call, identify where calls are actually leaking, and sketch the approval-gated automation that would recover them.
Book a free AI automation audit and we will send you a written map of your current intake workflow, the points where calls are dropping, and a concrete proposal for the handoffs and review gates that would close the gap. No obligation, no follow-up sequence, just the working document.
