An AI receptionist is a system that answers the phone and the web chat when a person cannot: out of hours, during the lunch rush, when the receptionist is on another call. In a dental practice it can book the appointment and be done. In a law firm it has to work inside confidentiality, the rules on who can give advice, and the SRA's expectations on supervision. That does not make it unusable. It makes the design decisions matter.
What it should do
- Answer, every time. The firm's greeting, the firm's tone, no hold music. It says it is an automated assistant. Not saying so is the first way to lose trust and the first thing a regulator would ask about.
- Take details. Name, best number, the type of matter in the caller's own words, whether there is a deadline, how they heard of the firm. Nothing more than a receptionist would write on a message pad.
- Book. Offer a call back or an appointment from the fee earner's real availability, confirm it by text and email, and put it in the diary.
- Route. Conveyancing enquiries to the conveyancing team, an existing client to their fee earner, an urgent matter to whoever is on call. Out of hours, the message goes where the firm has decided out-of-hours messages go.
- Answer the questions a receptionist answers. Office hours, address, parking, whether the firm does a type of work, whether a first conversation is free. Questions of fact about the firm, from a list the firm approved.
What it must not do
- Give advice, or anything that could be heard as advice. "Do I have a claim?" gets "That is exactly what the solicitor will go through with you; can I book that call?" Every time. The system has no path to an answer, by design, not by instruction. An instruction can be argued out of; a missing capability cannot.
- Discuss an existing matter. A caller who is a client of the firm is routed to their fee earner. The assistant does not confirm what the firm holds, does not read from the file, does not say whether a letter went out. Confidentiality is simplest when the assistant has no access to matter data at all.
- Pretend to be a person. It identifies as an assistant at the start. It does not have a human name. If the caller asks whether they are speaking to a person, it says no and offers to transfer or to take a message.
- Record without saying so. If calls are recorded or transcribed, the greeting says so. The transcript goes into the firm's own system, not into a consumer AI product, and is kept according to the firm's retention policy.
The SRA angle, in plain terms
Three things the SRA cares about map directly onto those rules. Clients must not be misled, which covers the assistant announcing itself. Advice must come from someone competent and supervised, which is why the assistant cannot answer questions of law. And confidential information must be protected, which is why the assistant has no access to it. The August 2026 warning notice on AI misuse is about unchecked output and client data in unsafe tools; an assistant that cannot advise and cannot see the files has no output to check and no data to leak.
The firm should also write down, in a page, what the assistant is allowed to say and who reviews its transcripts, and keep that with the firm's other policies. The AI policy template for law firms has a section for it.
What it is worth
The value is not in replacing the receptionist. It is in the enquiries that currently ring out, hit voicemail or arrive at 7pm on a web form and wait until the next morning. Those are instructions the firm never finds out it lost. An assistant that answers in seconds, takes the details, and books a call for 9am is the difference between being the first firm to respond and the third. For a firm that depends on new enquiries, that is the whole business case, and it is measurable: count the enquiries that arrive outside the hours a person answers, and count how many of them convert now.
How it fits with the rest of intake is on the intake automation guide, and the law firms page has pricing and the systems it connects to.


