Law Firm AI Receptionist:10 Things to Check

The most famous ethics opinion on legal AI does not squarely address the thing a firm buying an AI receptionist wants to know.ABA Formal Opinion 512, issued July 29, 2024, is the ABA's only formal opinion on generative AI that I could find as of September 27, 2026. It covers competence, confidentiality, fees and supervision for generative AI tools. What it never really does is talk about a bot answering your phone and taking intake from a stranger.

The opinion that does came from Florida, six months earlier. Florida Bar Ethics Opinion 24-1, dated January 19, 2024, says it plainly in its own text: "This guidance is especially useful as law firms increasingly utilize website chatbots for client intake." Opinion 512 even points to it, in a footnote. That footnote is about as close as the ABA gets to intake bots.

I'm Chris Machetto, and I run Frenchy Digital, a senior-led Black-owned agency in Los Angeles that builds AI agents, including a productized AI receptionist. So I have a stake here, and I'll disclose it wherever it matters. This article is the checklist I would want a managing partner to run on any AI receptionist, ours included, before it answers a single call.

There are ten checks. Each one ties to a rule, an opinion or a statute you can open yourself, and every date was checked on September 27, 2026. For the broader category (what these products are, how the vendors differ, how to roll one out) start with our AI receptionist buyer's guide. This piece is the law firm layer on top.

Here is a mental model a sensible buyer could easily bring to this: an AI receptionist is a better voicemail. It picks up, it takes a message, it books a slot. If it gets something wrong, the downside is a missed lead.

At a law firm, that model is wrong in a specific way. The first conversation with a prospective client is itself regulated. What the caller tells you is confidential under Rule 1.18 even if you never take the case. What you learn can conflict you out of another matter. Whether the caller thinks they are talking to a person, or to a lawyer, is an ethics question in at least one state.

The right model: an AI receptionist is a nonlawyer assistant doing intake, and you answer for it under Rule 5.3 the way you answer for a front desk hire. It is priced like software (tens to hundreds of dollars a month for the hosted products) and judged like an employee. Therefore the question is never "is the voice good?" It is "would I let a new hire say these exact words to this exact caller?"

To be clear, the downside of thinking this way is that it slows you down. Every check below adds a step before launch. I think that trade is worth it, because the alternative is discovering your intake script created a conflict after the retainer is signed on the other side.

One more thing before the list, because it shapes how much weight to put on each check. As of September 27, 2026, I found no disciplinary case or sanction involving an AI intake chatbot or AI receptionist. None. The AI discipline that exists is about court filings: Mata v. Avianca in 2023 and People v. Crabill in Colorado, where a lawyer was suspended for a year and a day (90 days served) over ChatGPT generated citations. Those cases tell you firms own unverified AI output. They are not intake precedent, and absence of precedent is not a safe harbor.

The check: does the AI treat every caller as a possible prospective client, with the warnings that implies?

The wrong model is that duties start at the retainer. They don't. ABA Model Rule 1.18 says information learned from a prospective client stays confidential even when no representation follows. The caller who describes their divorce at 9 p.m. to your AI and then hires someone else still gets 1.18(b) protection from your firm.

ABA Formal Opinion 492, from June 9, 2020, sharpens this. Summaries of it say a person becomes a prospective client when a lawyer invites information without clear and reasonably understandable warnings that limit the lawyer's obligations. An AI receptionist that says "tell me everything that happened" is inviting information. The warnings have to come first.

Florida Opinion 24-1 applies all of this directly to AI conducted interviews. It says Rule 1.18 applies, and it warns against a chatbot that lacks clear disclaimers limiting the lawyer's obligations. So the script matters word for word, the same way a paper intake form's fine print matters.

Therefore: before the AI asks a single substantive question, it should say that speaking with it does not create an attorney client relationship, that the firm has not agreed to represent the caller, and that the caller should not share details yet. Then it should ask only what the next step needs. The downside is a slightly colder opening, and some callers will hang up rather than sit through a disclaimer. That is a cost worth measuring, not a reason to skip it.

A useful everyday comparison: a good doorman takes your name and who you're visiting. He does not ask why you're there. The AI should work the lobby, not the consultation room.

The check: does the AI collect party names for a conflict check before it collects any narrative?

This is the one that can cost you a client you already have. Rule 1.18(c) says that if a lawyer receives information from a prospective client that could be significantly harmful to that person, the lawyer can be disqualified from representing someone else with materially adverse interests in the same or a substantially related matter. That disqualification can reach the whole firm.

Rule 1.18(d) gives you a way out: screening and written notice, or informed consent. But (d)(2) conditions that route on the lawyer having taken reasonable measures to avoid exposure to more information than was reasonably necessary to decide whether to take the matter. An AI that happily records a twenty minute story is the opposite of a reasonable measure.

Consider a scenario. Your firm represents a business owner in a contract dispute. At 11 p.m., the other side's CFO calls your line looking for their own lawyer, and your AI receptionist, being helpful, records the whole story, including the settlement number the company would accept. Nothing in that call is dramatic. It is exactly the kind of significantly harmful information 1.18(c) worries about, and it is now in your intake system.

Therefore the order of questions is the control. Names of the caller, the opposing party and any other parties first. Run the conflict check (automatically against your practice management system if you can, or by a person the next morning if you can't). Only after it clears does anyone collect the facts. The downside is that some intakes take two touches instead of one. I'd rather lose that speed than lose a current client.

Florida 24-1 adds a related screen: consider questions that limit the chatbot's communications when the person is already represented. "Do you currently have a lawyer for this matter?" belongs near the top of the script, with a clear path for a yes.

The check: where does the caller's information go, and can the tool learn from it?

Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of client information. An AI receptionist sends every caller's words through at least a telephony provider, a speech model and a language model, and possibly a CRM. Each hop is a place the information lives.

Opinion 512 is specific about one kind of tool. For "self-learning" generative AI whose output could disclose client information to others, the opinion says "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool." And it adds a line every firm with a template engagement letter should read twice: "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient."

Here is the awkward part for intake. A caller at intake has not signed anything. You cannot get informed consent from someone before they speak to your receptionist. So the practical answer is not consent; it's choosing a tool that does not learn from your callers' data in a way that could surface to anyone else, and getting that in writing.

Opinion 512 also says competence under Rule 1.1 means a reasonable understanding of the capabilities and limits of the specific tool, kept current. That is a nice way of saying "read the data terms yourself." If nobody at the firm can explain where a transcript is stored and for how long, the competence box isn't ticked yet.

The downside of being strict here is vendor choice. If a vendor won't put a clear no-training commitment in the contract, or offers one only in a higher tier, that narrows the field or raises the price, and sometimes narrowing the field is the point.

The check: does the AI tell every caller, in the opening seconds, that it is an AI program and not a lawyer or employee?

The wrong model is that disclosure is a branding choice, and a natural voice that callers mistake for a person is a feature. It is an easy thing to want. At a law firm, Florida has said otherwise, and it has said it plainly.

"a lawyer must inform prospective clients that they are communicating with an AI program and not with a lawyer or law firm employee."

The Florida Bar, Ethics Opinion 24-1 (January 19, 2024)

The opinion's summary goes further: "Generative AI chatbots that communicate with clients or third parties must comply with restrictions on lawyer advertising and must include a disclaimer indicating that the chatbot is an AI program and not a lawyer or employee of the law firm." It also flags a chatbot that fails to identify itself immediately as one of the things to be wary of. Immediately is the operative word.

Opinion 512 does not require AI disclosure to callers, and I'd refuse any vendor deck that says it does. What 512 says, under Rule 1.4 on communication, is that disclosure of AI use to clients may be required depending on the facts: when a client asks, when outside counsel guidelines address it, or when the output influences a significant decision. That is a different duty, owed to clients, not to every stranger on the line.

Outside the bar rules, general AI disclosure laws exist but with narrower triggers than people assume. California's bot law (Business and Professions Code 17940 to 17943) targets bots used to incentivize a sale or influence a vote with intent to mislead. Utah's AI Policy Act, as amended by SB 226 in 2025, requires general businesses to disclose on a clear and unambiguous request, with up front disclosure for high risk interactions and regulated occupations. Whether lawyers, who are licensed by the Utah Supreme Court, count as a regulated occupation there is not something I could confirm, so I won't assert it.

If you have callers in the EU, Article 50 of the EU AI Act applied from August 2, 2026 and requires telling people they are interacting with an AI system. The Digital Omnibus deferred other duties but not that one. Colorado replaced its AI Act with a narrower disclosure law (SB 26-189, effective January 1, 2027), per firm alerts I could not fetch directly.

Therefore: disclose on every call, in every state, first sentence. Something like "You've reached the firm's AI assistant. I'm an AI program, not a lawyer or a member of staff." The downside is that some callers will ask for a person right away. Good. Build the handoff and treat those requests as data, not failure.

The check: can you make the AI give legal advice, and what happens when you try?

Picture the questions a worried caller asks a receptionist. "Do I have a case?" "How long do I have to file?" "Can my landlord do this?" A trained human receptionist knows to say "that's a question for the attorney." A language model built to be helpful will try to answer, unless it is designed not to.

Model Rule 5.5 reserves the practice of law to lawyers, and Florida 24-1 names the risk directly: be wary of an overly welcoming chatbot that may give legal advice. The lawyer is responsible if the chatbot misleads, or is intrusive or coercive. That's the vendor's demo voice (warm, eager, helpful) turned into a liability.

A statute of limitations question is the dangerous one. An AI that says "you generally have two years" has given advice that may be wrong for the caller's state, claim type or facts, and a caller may rely on it. Even a correct general answer can mislead a specific caller.

Therefore, test it adversarially before launch. Call it and ask twenty legal questions in different phrasings, including sympathetic ones ("I'm scared, please just tell me if I should sign"). Every answer should route to a lawyer or a consultation, with no substance. If you can get advice out of it in an afternoon, a caller will get advice out of it in a week. The downside of a strict refusal policy is that the AI will sometimes sound unhelpful. That's fine. Unhelpful is recoverable; advice isn't.

To be clear, I don't trust prompt instructions alone to guarantee this, and you shouldn't either; a determined caller can phrase a question in ways a script never anticipated. The mitigation is layered: tight instructions, a narrow set of things the AI can say, a check on its output for advice-shaped content, and human review of transcripts. Frame it as reducing the blast radius, because that's what it is.

The check: who at the firm owns the AI receptionist, and what do they review each week?

Opinion 512 reads Rules 5.1 and 5.3 onto generative AI. Managerial lawyers must set clear firm policies on permissible AI use and train both lawyers and nonlawyers, and the opinion says Rule 5.3 duties extend to third party providers of AI services. Your answering vendor is inside that circle.

The wrong model is "set it and forget it." The whole sales pitch of an AI receptionist is that it runs unattended at 2 a.m. That's true of the answering. It can't be true of the oversight.

In practice, supervision looks like a named owner (a lawyer, not only an office manager), a written policy on what the AI may and may not say, a weekly sample of transcripts read by that owner, and a log of fixes made to the script. It is dull. It is also exactly what you would produce if a bar investigator asked how you supervised your intake.

Let me do the arithmetic on the time, since that is where firms balk. Say the owner reads 20 transcripts a week at about three minutes each. That's 60 minutes, or roughly 52 hours a year. At the fee rate of a partner, that isn't free, and I won't pretend it is. But it is about one hour a week against a system that talks to every new caller after hours, and it is the hour that turns "we trusted the vendor" into "we supervised our assistant."

State guidance lines up with this. California's Practical Guidance (November 16, 2023), the NYSBA Task Force report (April 2024), Texas Opinion 705 (February 2025), North Carolina's 2024 Formal Ethics Opinion 1 and DC Bar Opinion 388 all stress competence, confidentiality and supervision. I could not confirm intake specific content in the Texas opinion, so read your own state's document rather than trusting a summary, including mine.

The check: does the AI announce recording on every call, before anything else is said?

An AI receptionist that hands off with the conversation attached is, by design, capturing that conversation as a recording, a transcript or both. That makes it a consent question in a number of states.

California Penal Code 632 requires the consent of all parties to record a confidential communication. A caller telling a law firm about their situation is about as clear an example of an expected confidential conversation as exists. About a dozen states are commonly listed as all party consent jurisdictions (California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington among them), and the lists disagree at the edges, with a few states often listed with caveats.

Callers don't stay in your state, either. A Los Angeles firm will get calls from people in states with different rules, and the safest assumption is that the strictest rule applies. Therefore: announce recording on every call, everywhere, and if the caller objects, stop recording or hand off to a person. The downside is one more sentence in the opening, stacked on the AI disclosure. Keep both short.

SB 690 matters to a firm's website, not its phone line. If your intake chat widget or tracking pixels have drawn CIPA demand letters, watch the Governor's decision. For the phone, nothing changes, so the recording announcement stays.

The check: does the AI ever place a call, and if so, what consent covers it?

Answering an inbound call and placing an outbound one are different legal events. The tempting feature is an automatic callback: a web form comes in and the AI calls the lead right away. If your vendor offers anything like it, that feature is where the TCPA lives.

The FCC's Declaratory Ruling FCC 24-17, adopted February 2, 2024 and released February 8, 2024, says AI technologies that generate human voices are "artificial or prerecorded voice" under the TCPA. Calls using them need prior express consent. An AI voice calling a lead back is, for TCPA purposes, a robocall.

One misreading to watch for: the Eleventh Circuit's decision in Insurance Marketing Coalition v. FCC (January 24, 2025) vacated the FCC's 2023 one to one consent rule. It did not touch FCC 24-17. If a vendor tells you the AI voice ruling was struck down, they have confused the two.

There is a second layer for lawyers. Model Rule 7.3 restricts solicitation by live person to person contact. Whether an AI voice call is "live person to person contact" isn't settled in any source I found, and I won't pretend it is. The safe reading is that an AI calling someone who never asked to hear from you is a problem under one rule or the other.

Therefore: the AI calls back only people who asked to be called, with the consent language captured on the form that asked, and only about the inquiry they made. No list calling, ever. The downside is that you lose the fastest follow-up on leads who didn't tick the box. That is a small price next to the alternative.

The check: what does the contract say about training, retention, subprocessors and deletion?

The marketing page will say "secure" and "compliant." Neither word does anything for you. What does something is a clause in the contract you signed.

Four questions cover most of it. Does the vendor, or any model provider it uses, train on your callers' content? How long are recordings and transcripts retained, and can you set that? Who are the subprocessors (telephony, speech, language model, hosting)? And can you get everything deleted, and exported, when you leave?

A word on BAAs, since vendors raise them. A business associate agreement is a HIPAA instrument. Whether your firm needs one depends on whether the firm itself handles protected health information as a business associate for a covered entity client, which is a question for your own counsel, not for a receptionist vendor's sales team. A personal injury caller describing their injuries doesn't automatically turn the call into HIPAA territory, but it is still confidential under Rule 1.18, and that is the standard that governs you.

Opinion 512's competence duty cuts here too: you need a reasonable understanding of the specific tool. For a law firm, "we read the data processing terms and here is where transcripts live" is part of that understanding. The downside is legal review time on a contract that looks like a SaaS click-through. Spend it anyway.

For how a legal intake agent can be built so the data stays in systems the firm controls, our walkthrough of Claude legal intake agents for Black-owned law firms in LA goes through the architecture in detail.

The check: how is the AI's cost treated on client bills, and what record exists of each intake?

Opinion 512 addresses fees under Rule 1.5. Hourly billing covers only time actually spent. Tool costs that function like general overhead should not be charged to clients absent advance disclosure, while per use third party costs may be billed as actual out of pocket expenses. The opinion also says a lawyer should "not charge a client to learn about how to use a GAI tool or service that the lawyer will regularly use for clients".

A receptionist that answers every caller looks like overhead, the same way the phone bill does. I'd assume it stays off client invoices unless your ethics counsel says otherwise. The downside is that the tool's cost comes out of margin. That is how the front desk has always been paid for.

The paper trail is the other half. For each intake, you want the disclosure that was played, the recording announcement, the conflict check result and time, what was collected after clearance, and who at the firm reviewed it. If you ever have to show that you took reasonable measures under Rule 1.18(d), that record is the evidence.

Therefore, store the log in your own practice management or intake system, not only in the vendor's dashboard. Vendors change, get acquired or shut down. Clio, for one, closed a $1 billion acquisition of vLex in November 2025, which is a reminder that the company behind your tool can change hands. Your records should outlive your vendor.

This is not a ranking. It is what five vendors a law firm is likely to meet publish about themselves, checked September 27, 2026, so you know where the conversation starts. Every price is the vendor's own published figure or clearly attributed reporting; where a vendor doesn't publish one, the table says so. Full pricing across the category is in our AI receptionist pricing comparison.

A couple of things jump out. Smith.ai publishes clear AI receptionist tiers: a free tier with 25 calls and $3.00 per call overage, Pro from $150 a month and Enterprise from $500 a month. A legal conflict check add-on appears in secondary sources, but it isn't on the AI pricing page, so I don't print its price.

Ruby is the human comparison. Its $395 for 100 minutes works out to $3.95 a minute, and its $720 for 200 minutes to $3.60 a minute. That's the price of a person following your script, and the per minute cost falls only about 9% when you double the bundle.

Clio's Grow AI, launched August 20, 2026, is the entrant I find most interesting because it comes from Clio itself, as part of Clio Grow, rather than from an answering company bolting on a legal integration. LawSites reports it at $25 per converted lead, with US phone agents only. I found nothing published on how it tells callers they're speaking with an AI, which is check number four. Ask.

Outcome pricing sounds like alignment. Do the math before assuming it's cheap: 20 converted leads a month at $25 is $500, the same as Smith.ai's Enterprise starting tier. At 40 leads it's $1,000, or 2x. The better your intake converts, the more you pay, which is fine as long as "converted" is defined in the contract.

The other angle is the broader legal AI market, which runs well beyond the phone. Our ranking of the top AI agents for law firms in 2026 covers research, drafting and e-discovery tools, and applies the same verify-only scoring.

All ten checks collapse into one artifact your ethics partner can sign: a table of what the AI may do alone, what needs a person first, and what it must never do. This is the version I'd start from. Edit it for your state and practice areas.

The middle column is where most of the design effort goes. Everything in it is a handoff, and a handoff is only as good as the context it carries. If the person picking up has to ask the caller to start over, you've lost most of what the AI was for.

The right column is not a wish list. Each line maps to a rule: advice to 5.5, pre-clearance facts to 1.18(c), uninvited calls to the TCPA and 7.3, comparative claims to your advertising rules, which Florida 24-1 says chatbots must follow. Write the rule citation next to each line in your internal version.

Disclosure first: Frenchy Digital sells an AI Receptionist, so weigh what follows accordingly. It's $5,000 for the agent plus a $5,000 one-time setup fee, both paid through Stripe Checkout as one-time payments, not a subscription. The product page says most builds run 2 to 4 weeks. Usage costs depend on the telephony and model providers and aren't included in that price.

It doesn't replace a receptionist. It covers after hours, lunch and busy-line overflow, and hands off to your people with the conversation attached. For a law firm, the setup is where the ten checks get built in: the disclosure script, the names-first conflict flow, the recording announcement and the advice refusals.

The arithmetic against a hosted tier, out loud: $10,000 one time, divided by Smith.ai Pro's $150 starting monthly price, is about 67 months of that subscription fee. So on sticker price alone, the hosted product is cheaper for years, and I'd say so to any firm. The case for a build is control (your script, your data, source code and IP transferred to you), not price.

For firms that need more (intake wired into a practice management system, conflict checks run automatically against the matter database, multi-office routing), our custom engagement bands are below. We work senior-led at $150 to $225 an hour, offer retainers from $2,500 to $9,500 a month, include a 30-day post-launch warranty, and send a fixed-price phased proposal within 5 business days.

The website matters as much as the phone line, because it is where callers come from and where the AI's answers about hours, practice areas and consultations come from. We've rebuilt professional practice sites on that principle, such as the BH Dental Corp website revamp, with one page per service and a contact route at the end of each. That is website work, not a receptionist build, and I cite it only for that. If you want the technical side of how a voice agent answers quickly enough to feel natural, the voice AI agent architecture guide covers latency in depth.

You'll hear some version of each of these. Each one is a reason to slow down, not necessarily to walk away.

• “Our voice is so natural callers never know it’s AI.”: That is a Florida 24-1 problem presented as a feature. Ask how to make it disclose, first sentence, every call.
• “ABA 512 approves AI intake.”: It doesn’t address intake bots squarely. It cites Florida 24-1 in a footnote. A vendor who gets this wrong hasn’t read the opinion.
• “We capture the full story on the first call.”: Without a conflict check first, that is how a firm imports significantly harmful information under Rule 1.18(c).
• “Our AI can answer common legal questions.”: That is legal advice by a nonlawyer. Ask for the refusal policy and test it yourself.
• “Automatic callbacks on every lead.”: AI voice calls need prior express consent under FCC 24-17. Ask where consent is captured.
• “SB 690 fixed the recording lawsuit problem.”: As of September 27, 2026 it awaited the Governor, and it concerns website pen register claims, not Section 632 call recording.
• A published accuracy or answer rate with no method: A vendor’s number about itself is a claim. Ask for the methodology or ignore the figure.
• No clear training or retention clause: Opinion 512’s confidentiality analysis turns on whether the tool learns from inputs. Get it in the contract.

None of these is fatal on its own. A vendor who fixes the disclosure script and puts the no-training promise in writing when asked may be a fine partner. The red flag is a vendor who argues.

This article is only as good as its sources, and some of them had limits. Here they are, so you can weigh the rest.

• The ABA’s own pages block automated fetch: Opinion 512 was read in full from a published mirror; the Model Rules pages and the Opinion 492 summary are cited at their official addresses but were not machine-readable to us. Opinion 492 details come from secondary summaries.
• Whether an AI voice call is live person to person contact under Rule 7.3: Not settled in any source found. We assert neither answer.
• Utah’s treatment of lawyers: Whether lawyers fall under Utah’s regulated occupation disclosure rule after SB 226 could not be confirmed.
• Colorado and EU details: Colorado SB 26-189 and the Digital Omnibus timing come from law firm alerts we could not fetch directly.
• California’s 2026 guidance revisions: Search results report revisions approved in May 2026 adding agentic AI guidance; not confirmed on the primary document.
• The exact all party consent list: Published lists disagree. We say about a dozen and recommend announcing recording everywhere.
• LEX Reception and Alert Communications pricing: Not published by either vendor. Directory figures are not current prices.
• Clio Grow AI’s caller disclosure: Nothing published was found; the $25 per converted lead figure is LawSites reporting of a vendor price.
• SB 690’s final fate: Awaiting the Governor as of September 27, 2026, with a September 30 deadline. Recheck before relying on it.

There is also a broader limit. No disciplinary case about AI intake exists yet, so every check above is an application of general rules to a new tool. The first real case may draw the lines somewhere slightly different. Building to the conservative reading is the hedge.

If you're already running an AI receptionist, or about to sign for one, this is where I'd start.

1. 1.Call your own line after hours and listen to the first ten seconds. If it doesn’t say it’s an AI program, not a lawyer or employee, and that the call is recorded, fix the script before anything else.
2. 2.Read the intake questions in order and move every narrative question below a names-only conflict check. If the tool can’t do that, ask the vendor in writing when it can.
3. 3.Name one lawyer as owner, and put 60 minutes on their calendar every week to read 20 transcripts. Log every change you make to the script.

That's it. The rest of the list can wait a week. Those three can't.