What an AI receptionist should and shouldn't do at a law firm

AI answering a law firm's phone raises real questions about advice, confidentiality, and disclosure. Here's the line we draw and why.

A firm considering AI on its phone line asks two questions in quick succession. Will it work? And is it appropriate? The second question is the more important one, and it deserves a straighter answer than most vendors give.

The problem it exists to solve

Start with the honest baseline. The alternative to AI answering at 2am is not a person answering at 2am. It is voicemail. And people with urgent legal problems do not leave voicemails — they call the next firm on the results page. A firm sending after-hours calls to a message box is not competing for those matters; it simply never learns they existed.

That framing matters, because the relevant comparison is not AI versus a human receptionist. It is AI versus nothing.

What it should do

• Answer every call, on the first ring, at any hour
• Ask the qualifying questions the firm would ask, consistently every time
• Screen for urgency and transfer genuine emergencies to a person immediately
• Book a consultation on the calendar during the call
• Answer factual questions about the firm — hours, locations, practice areas, what to bring
• Write a full transcript into the CRM so nothing depends on recollection

What it should never do

This list is shorter and considerably more important.

• Give legal advice, or anything a caller could reasonably mistake for it
• Deny being an assistant if a caller asks directly
• Handle an emergency itself when the rules say transfer to a human
• Make representations about outcomes, timelines, or fees it cannot stand behind

Confidentiality is a design requirement

Anything a caller tells an intake system is confidential from the first word. In practice that means scoping what the agent is permitted to collect and store, least-privilege access on the systems holding it, encryption in transit, and no training on your call data. These are architecture decisions made before launch, not policies added after.

Test the hard calls before the easy ones

Any agent handles a straightforward new-client inquiry. The ones worth testing before go-live are the distressed caller, the family member who does not have the details, the person calling about a matter you do not take, and the genuine emergency. We script and rehearse those with the firm's own attorneys, because that is where the reputational risk actually lives.

You can hear where we've landed on this by calling the demo line on the AI Receptionist page. Ask it something it shouldn't answer and see what it does.