Voice AI for Law Firm Intake: Real Costs and Ethics Rules
Voice AI can answer every call a law firm misses, but bar ethics rules treat the agent as the voice of the firm. Real platform and answering service costs, plus the compliance design that keeps intake safe.

Most law firms do not have a lead problem; they have an answer-the-phone problem. In a 2025 secret-shopper study of 500 US law firms, only a third of firms answered the call at all, and of those, only 36 percent ran any kind of structured intake. A voice AI intake agent can answer every call, collect the facts in the right order, and book the consultation for roughly $100 to $300 a month, but a law firm cannot bolt one on the way an online shop can. Bar ethics rules treat the agent as the voice of the firm, so the script, the data handling, and the vendor contract are compliance decisions, not just IT purchases.
This guide covers both halves: what the ethics rules actually require of an intake agent, and what the real options cost using primary vendor pricing.
Why intake is the problem worth solving
The numbers behind intake failure are stark. The Clio 2025 Legal Trends Report found that 79 percent of lawyers expect to miss prospective clients each week, 40 percent take two to three days to respond to an inquiry, and about two thirds of firms are effectively unreachable by email. Meanwhile the caller with a DUI arrest at midnight or a wage dispute on a Sunday is dialing the next firm on the list.
A human answering service fixes availability but not consistency, and it scales linearly with call volume. A voice AI agent fixes availability, asks the same screening questions every time, captures names for a conflict check, and writes a structured summary into your case system. What it cannot do, and must never be allowed to do, is the subject of the next section.
What the ethics rules let an intake agent do
The most useful guidance published to date is the Oregon State Bar's Formal Opinion 2026-208, approved in February 2026, which addresses chatbots and AI agents used for client intake and marketing directly. Read together with the ABA's Formal Opinion 512 on generative AI (July 2024), the boundaries are clear.
Jobs the agent can do well:
- Answer 24/7 and collect facts in a fixed order: caller name, contact details, practice area, what happened, key dates.
- Capture conflict-check data: the full names of the caller, the opposing parties, related businesses, and insurers, before any details of the matter are taken.
- Screen and route: apply written criteria to decide whether to book a consultation, and escalate urgent situations (a client in custody, someone who has just been served papers) to a named human immediately.
- Write structured summaries into the firm's case management or CRM system for a lawyer to review.
Hard limits the rules impose:
- No legal advice. The Oregon opinion warns that a chatbot providing legal advice risks implicating the firm in the unauthorized practice of law. The script must decline, politely and consistently.
- No engagement letters, fee quotes, or matter acceptance. Oregon explicitly flags the risk of an autonomous agent sending an engagement letter: a recipient could reasonably construe that as the formation of an attorney-client relationship, even if the lawyer never intended it. Any autonomous outbound action needs lawyer approval under supervisory-duty rules (Rule 5.3).
- The agent must identify itself as an AI. A chatbot that does not identify itself as a chatbot can be a false or misleading communication under Rule 7.1, and Oregon notes that more than one state has taken this position.
- A lawyer must be notified and review. The opinion says it would be prudent to have a responsible lawyer automatically notified whenever the agent interacts with a prospective client, and to review transcripts promptly. One reason: Rule 7.3(a) concerns. A caller whose emotional or mental state is impaired may be obvious to a human receptionist and invisible to a bot.
- Client data must not train the model. Vendor diligence must confirm that intake data is encrypted, isolated per client, and never used to train the provider's models. ABA Opinion 512 requires informed client consent before confidential information goes into a self-learning tool, which is a consent conversation no intake caller can meaningfully have. The vendor contract has to close that gap.
The conflict check comes before the conversation
This is the design decision most intake deployments get wrong. Under Model Rule 1.18, a person who consults a lawyer about possibly hiring them is a prospective client, and the firm owes that person real duties: no use or disclosure of what they said, and a conflict bar if the information could be significantly harmful to them in a related matter. That duty attaches even if the firm never takes the case.
The ABA's Formal Opinion 510 (March 2024) explains what "reasonable measures" look like for avoiding these conflicts: limit the information gathered at first contact, give cautionary statements during initial communications, use clear website disclaimers, and screen before a full consultation happens.
An intake agent is actually the best tool a firm has ever had for this, if it is built in the right order:
- Collect names first. Caller, opposing parties, related entities, insurers. Nothing about the facts yet.
- Run the conflict check against the firm's system while the caller is still on the line or before any detailed questioning.
- Only then collect the story. If the check flags a hit, the agent stops, thanks the caller, and routes to a lawyer for a decline handled by a human.
- Log everything. The Rule 1.18 duty runs on what the firm learned, so the record of exactly what was collected matters.
Build it backwards, collecting the full narrative before the names, and the firm has manufactured its own conflict problem with a recording attached.
Recording, disclosure, and consent
Two separate consent issues sit on top of every intake call, and the agent's opening line has to handle both in the first ten seconds.
Recording consent. Call recording law is a state-by-state patchwork. The Reporters Committee for Freedom of the Press guide tracks it: most states are one-party consent, but eleven states require all-party consent for recording, including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington, with Michigan and Connecticut effectively in the same column through case law and civil liability. A few more states have quirks (Nevada's statute says one-party but its Supreme Court has required all-party). The safe design for a firm that takes calls from anywhere is to announce the recording at the top of every call and treat continuing the call as consent. Kenyan and East African firms face a different framework; we covered those rules separately in our guide to call recording consent for AI voice agents in Kenya.
AI disclosure. The same opening should say the caller is speaking with an automated assistant, not a person. Beyond Rule 7.1's misleading-communication problem, this is simply what the Oregon opinion and the states following it expect.
A compliant opening is short: "Thank you for calling [firm]. This is an automated assistant, and this call is recorded to help our team follow up. I can take your details and book a consultation, but I cannot give legal advice." Then the first question is for the caller's full name, because the conflict check comes first.
What it costs, with real numbers
There are three routes, and the right one depends on call volume and how deep your integrations need to go. (For the general economics of voice agents across industries, see our voice AI implementation cost breakdown; the numbers below are the legal-intake-specific version, from the vendors' own pricing pages.)
Route 1: Per-minute voice AI platforms
You assemble the stack yourself: telephony, speech-to-text, an LLM, and a voice, orchestrated by a platform.
- Vapi: a $0.05 per minute hosting fee with the underlying components (speech-to-text, LLM, voice, transport) passed through at the providers' list prices. Vapi's own pricing calculator puts 1,000 minutes at $82 to $129 all-in with Deepgram, OpenAI, and ElevenLabs, so about $0.08 to $0.13 per minute depending on which models you pick. Transport on Vapi's own telephony or SIP is free; bringing Twilio adds about $0.008 per inbound minute. The Core package ($29 per month) includes 10 concurrent calls and 5 phone numbers; extra concurrency is $10 per line per month. If the firm handles protected health information and needs HIPAA-eligible handling with a BAA, that add-on is $2,000 per month, which changes the math completely.
- Retell AI: pay-as-you-go at roughly $0.07 to $0.31 per minute depending on configuration, with 20 concurrent calls included and $8 per month for each additional concurrency. Volume discounts start around 100,000 minutes.
Per-minute pricing rewards short, disciplined calls. A rambling ten-minute intake costs twice a tight five-minute one, so script design is a cost lever, not just a quality lever.
Route 2: Flat-rate legal answering services
You buy a packaged service with legal intake scripts already built.
- Smith.ai human virtual receptionists: from $300 per month for 30 calls (about $10 per call) up to $2,100 per month for 300 calls (about $7 per call), with overage billed 15 to 25 percent above the plan rate. Useful per-call add-ons include appointment booking at $1.50 and call recording with transcription at $0.25.
- Smith.ai AI Receptionist: a free tier covers 25 calls a month; the Pro plan starts at $150 per month for 75 calls (about $2 per call), with human receptionists available as escalation on hybrid plans.
- Hona, a legal-specific voice AI: about $200 per month plus a $500 setup fee, with 100 call-minutes included.
Flat-rate per-call pricing is easier to budget, but read the definition of a "call" and the overage rate, because after-hours hang-ups and robocalls can consume your allowance.
Route 3: A custom-built intake agent
A purpose-built single-task intake agent, designed around your practice areas, conflict-check order, and case management system, runs $8,000 to $25,000 for the build, plus the underlying per-minute platform costs from Route 1. This is the route that makes sense when the intake flow is genuinely custom: multiple practice areas with different screening logic, direct write-back into Clio or a practice management system, multilingual intake, or integration with the firm's existing telephony.
Worked example: a 20-attorney firm
Assume 100 after-hours and overflow calls a month, averaging six minutes, so 600 minutes.
- Vapi self-managed: 600 minutes at $0.08 to $0.13 is $48 to $78, plus the $29 Core package, for about $77 to $107 per month. Add part-time engineering time to build and maintain the flow, which is the real cost of this route.
- Retell self-managed: 600 minutes at $0.07 to $0.31 lands between $42 and $186 per month, plus the same engineering time.
- Smith.ai human service: 100 calls at the entry rate of about $10 per call is roughly $1,000 per month, with humans on every call.
- Smith.ai AI Receptionist: the $150 Pro plan covers the first 75 calls, plus about $2 each for the remaining 25, so roughly $200 per month.
- Hona: $200 per month plus the $500 setup fee.
- Custom build: a $15,000 midpoint build breaks even against a $1,000-a-month human service in about 15 months, but takes more than a decade to break even against a $100 self-managed platform. Custom pays off when you need the deep integrations or the volume is much higher, not as a way to save money at 100 calls a month.
One cost footnote from ABA Opinion 512: the firm can pass actual tool costs to clients where agreed, but it cannot bill a client for the time spent learning to use the tool. Treat implementation as overhead.
How to build a compliant intake flow
A step-by-step sequence that satisfies the rules above:
- Write the script with a lawyer in the room. Define exactly what the agent may say, the facts it collects, the phrases it uses to decline legal advice, and the triggers for urgent escalation. The Oregon opinion puts ongoing monitoring and the ability to modify responses on the lawyer, so the script is a living document, not a launch artifact. The general design disciplines in our conversational AI development guide (prompt structure, fallback handling, escalation triggers) apply directly here.
- Order the call correctly. Disclosure and recording notice, then conflict-check names, then the narrative, then booking. Nothing else.
- Wire the conflict check. The agent needs API access to the firm's conflict system, or at minimum a same-day human review step before any second call.
- Build the escalation path. Name the human who gets paged for the in-custody caller, and test the transfer monthly.
- Set up lawyer notification and transcript review. Every prospective-client interaction lands in a review queue. Oregon's opinion treats this as prudent practice; treat it as required.
- Do vendor-chain diligence. The platform, the LLM provider, the speech provider, and the telephony carrier all touch the data. Get contractual confirmation that intake data is not used for training and is isolated per customer. Those vendor contracts are worth a careful read; our piece on AI contract review accuracy benchmarks covers how well AI tools do this review and where they still miss. This is also exactly the diligence exercise we run when deploying Claude for law firms and legal teams, where the same confidentiality analysis applies to document work.
- Log versions. Keep dated versions of the script and the prompt. If a state bar ever asks what the agent was telling callers last March, you want an answer.
When voice AI is the wrong choice
Honest limits, because this is not always the right tool:
- Practice areas where callers are in crisis. Oregon's Rule 7.3(a) concern is real: a caller who is distraught, impaired, or in shock needs a human who can hear it. Criminal defense and family law firms should route these calls to humans or use AI only as overflow with fast escalation.
- High-value consults where the first impression is the product. If your average matter is a seven-figure transaction, the $153 a month you saved is not worth a robotic first touch. Use AI for after-hours capture and put partners' prospects straight through to humans.
- Firms that will not do the review work. An unmonitored intake agent is worse than voicemail. If nobody will read transcripts and tune the script monthly, do not deploy.
- Very low call volume. Under about 30 calls a month, a shared human answering service is simpler and comparably priced.
Common mistakes
- Collecting the story before the names. This inverts the Rule 1.18 sequence and creates conflicts the firm then has to paper over with screens and consents.
- Letting the agent quote fees. An AI that quotes a fee has arguably made an offer. Fee discussion belongs to humans; the script should book the consultation and stop.
- No recording disclosure, or the wrong one for the caller's state. Eleven-plus states want all-party consent. Announce it every time.
- Treating the vendor's default prompt as the finished product. The default demo prompt books meetings; it does not run a conflict check, decline legal advice, or escalate a caller in custody.
- Forgetting the training-data clause. If the vendor can train on your intake calls, ABA Opinion 512's confidentiality analysis says you need informed client consent you cannot get at intake. Fix it in the contract or pick a different vendor.
- Building custom to save money. At low volume, a $15,000 build loses badly to a $150-a-month service. Build custom for capability, not cost. For firms that want a managed version of the custom route, an AI employee built for intake and follow-up covers the build, the compliance script, and the ongoing tuning as a service rather than a project. For the general landscape beyond legal, see our overview of how businesses deploy voice AI agents.
Frequently asked questions
Can an AI intake agent give legal advice? No. The Oregon State Bar's 2026 opinion warns that a chatbot giving legal advice risks unauthorized practice of law problems for the firm. The script must decline legal questions consistently and offer to book a consultation with a lawyer instead.
Does talking to an intake bot create an attorney-client relationship? It can, in the caller's reasonable belief. Oregon law, like most states, looks at the client's reasonable belief, and an agent that sounds authoritative and sends follow-up documents can create duties the firm never intended. Clear disclaimers, no engagement letters from the bot, and prompt human review are the defenses.
Do I have to tell callers they are talking to an AI? Yes, in practice. A chatbot that does not identify itself as one can be a misleading communication under Rule 7.1, and multiple states have taken that position. The same opening line should also disclose that the call is recorded, which covers the all-party-consent states.
Is per-minute voice AI cheaper than an answering service? Usually, at intake volumes: 600 monthly minutes costs roughly $42 to $186 on per-minute platforms versus about $1,000 for a human answering service at 100 calls a month. The honest comparison adds the engineering time to build and maintain the per-minute stack, which is why packaged services and managed builds exist.
Can the AI run the conflict check itself? It can collect the names and trigger the check, and it should do both before taking any details of the matter. The check itself runs against the firm's conflict system, and any hit gets handled by a human, because Rule 1.18 duties attach to what the firm learns at this stage.
What if my firm handles medical records or health matters? Then HIPAA likely touches your intake calls, and the economics change: Vapi's HIPAA compliance package with a BAA is $2,000 per month. At that point a human service or a flat-rate legal product without the HIPAA premium is often cheaper.
Where to go from here
Start with the call logs: count how many calls you miss in a week and what a recovered consultation is worth, then pick the route that matches your volume. The compliance design, disclosure first, conflict check second, human review always, matters more than which vendor's voice sounds warmest. If you want help designing an intake agent that satisfies the ethics rules from day one, or you are evaluating AI across the rest of the firm's document and research work, that is the core of our Claude for Legal practice, and we are happy to talk through the specifics of your intake flow.
About AI Agents Plus Editorial
AI automation expert and thought leader in business transformation through artificial intelligence.



