A prospective client calls your firm at 7:00 pm about a car accident. Nobody picks up. She hangs up and calls the next firm on her list.

That call was worth real money, and it’s gone. It’s also exactly the kind of call an AI voice agent is built to catch. But before you put one on your firm’s phone line, there’s a harder question than whether it can hold a conversation. It’s whether it can hold that conversation without drifting into legal advice, and without creating a confidentiality problem the moment it starts recording.

What these systems do differently

Older phone menus break the second a caller says something unexpected. AI voice agents don’t work that way. They use speech recognition to turn what a caller says into text, natural language processing to figure out what she actually means, and a language model to generate the response. The result holds a real back and forth. The better platforms connect directly into your practice management software to act on the call itself. Schedule the consultation. Log the intake.

The intake problem and the confidentiality problem

For a firm, the upside is obvious. Gartner projects conversational AI will cut global customer service costs by $80 billion by 2026, and that figure, enterprise-scale as it is, reflects the same math that applies to a five-attorney firm losing after-hours calls to voicemail. Every missed intake call is a prospective client who called the next name on the list instead.

But a call to your firm isn’t a call to a pizza place. The moment that caller starts describing her accident, her custody dispute, or her business dispute, she’s sharing information relating to the representation, or the possible representation, of a client. That’s the same duty of confidentiality that applies to any AI agent your staff puts to work, and it doesn’t pause because the party on the other end of the line is software instead of a person. Under the ABA Model Rules, Rule 1.6(c) frames that duty around reasonable efforts to prevent unauthorized access to that information. California firms carry a comparable duty through Business and Professions Code section 6068(e), Rule 1.1’s competence requirement, and Formal Opinion 2015-193 on cloud computing, since that call recording is very likely sitting in a vendor’s cloud the moment the call ends.

There’s a second problem sitting right behind it. If the assistant answers a caller’s actual legal question, “do I have a case,” “what’s my deadline,” “am I liable here,” it’s stopped taking a message and started practicing law without a license. That exposure belongs to the firm, not the vendor.

Where the line has to sit

Don’t hand a voice agent the whole intake conversation on day one, the same phased approach that works any time you implement AI in your firm. Scope it the way you’d scope a new intake coordinator’s first week: hours, availability, scheduling a consultation, taking a callback request. Nothing that answers a legal question, nothing that implies an attorney-client relationship has formed. The venture firm a16z made the same observation in a 2025 analysis of the voice AI market: the deployments that hold up start with a narrow, low-stakes slice of calls and expand only once the results prove out.

When you’re evaluating a vendor, ask where recordings are stored, who at the vendor can access them, and whether the platform integrates cleanly with Clio, MyCase, or whatever you’re running, before you ask about price. Test it with your actual practice-area vocabulary and a caller who doesn’t speak in complete sentences, not the vendor’s polished demo.

Back to that 7:00 pm call. The question isn’t whether your firm can afford to miss it. It’s whether you can catch it without opening a confidentiality gap or an unauthorized practice of law problem you didn’t sign up for. Our managed IT services for law firms help you scope voice AI and intake tools that stay inside both lines, and it’s a lot easier to build that in from the start than to unwind it after a bar complaint or a lost client finds the gap first.