An AI Voice Receptionist for a Law Firm: What a Real Build Looks Like

Law firms miss calls for structural reasons: reception is busy with the person standing in front of them, fee earners bill by the hour and can’t answer phones, and the calls keep coming after 5:30pm. Every missed call is potentially a missed matter. This is what an AI voice receptionist actually looks like deployed — drawn from a build I did for a UK solicitors’ practice — as opposed to what vendor landing pages imagine.

What it does on a real call

The agent answers instantly, any hour, and does four jobs:

1. Answers the questions receptionists answer all day — practice areas, how fees work, what a first consultation involves, what documents to bring, office hours and directions. It answers from a knowledge base built from the firm’s own material, so the answers are the firm’s answers, not a language model’s improvisation.
2. Triages the matter — asks what the call concerns and classifies it by practice area (conveyancing, family, immigration, litigation…), because “new conveyancing enquiry” and “existing family matter” need entirely different handling.
3. Captures the caller properly — name, number, matter type, urgency — straight into the CRM as fields, creating the enquiry in the right practice-area pipeline. No message pads, no “someone called about a house”.
4. Books the consultation — offers real slots from the right fee earner’s calendar and confirms by SMS.

The guardrails are the build

The engineering effort isn’t making the agent talk — that part is nearly free now. It’s making the agent refuse correctly:

• No legal advice, ever. The agent is scoped to information and logistics. Anything that smells like “what should I do about…” gets a courteous redirect to a consultation. This line is written into its instructions in a dozen different phrasings, because callers ask in a dozen different ways.
• Fail towards humans. When it can’t answer or the caller is distressed, it transfers during office hours or takes a call-back commitment after hours. A receptionist who guesses is a liability; the same is true of an agent.
• Confidentiality-shaped memory. It never repeats details from one call into another, and what it stores goes to the CRM under the firm’s retention rules, not into a vendor’s training data.

What it costs to run

The moving parts are per-minute AI voice pricing plus telephony — for a small firm’s realistic call volume, running cost lands in the tens of dollars a month. The comparison isn’t really against that number, though: it’s against the value of the after-hours conveyancing enquiry that currently rings out, and against reception salary hours spent repeating the fees explanation for the ninth time that day.

Honest limits

It mishears occasionally — strong accents on bad lines are still hard, which is why the handoff rules matter more than the happy path. It shouldn’t handle existing-client matter updates without careful identity checks. And it will not make a bad intake process good: if enquiries currently vanish into an unmonitored inbox, fix the pipeline first — the receptionist (human or AI) is only the front door.

If your firm — or your agency’s client — needs this, book a call. I’ll show the triage flows from the real build and what a version for your practice areas looks like. The CRM side of the same build pattern is described in the accountancy intake write-up — professional-services intake is the same machine wearing different robes.