Legal Tech and AI

What AI Phone Tools Should Never Do at a Law Firm

Hard limits for AI phone tools at law firms: no legal advice, no hidden AI, no fee promises, no skipped conflict checks, and how to test each guardrail.

AssistCall research teamPublished Reviewed by Faruque Amin, VP of Sales and Marketing7 min read

On this page
  1. Why hard limits matter more than features
  2. The never list
  3. The prospective-client problem
  4. Confidentiality limits
  5. Biometric and recording limits
  6. Phrases the assistant must not say
  7. Urgent calls are the exception to every script
  8. Test the guardrails before launch
  9. Signs a tool is drifting past its limits
  10. Who owns the limits
  11. Frequently asked questions
  12. References
Legal Tech and AI: What AI Phone Tools Should Never Do at a Law FirmArticle coverLEGAL TECH AND AIWhat AI Phone ToolsShould Never Do at aLaw FirmAva for LawyersInsightsLegal Tech and AI: What AI Phone Tools Should Never Do at a Law FirmArticle coverLEGAL TECH AND AIWhat AI Phone ToolsShould Never Do at aLaw FirmAva for Lawyers

Key takeaways

  • An AI phone tool should never give legal advice, predict outcomes, quote a fee as final or promise the firm will take a case.
  • Florida requires AI chat tools to disclose they are not lawyers and warns they can create a prospective-client relationship without the lawyer knowing[1].
  • Keep AI tools from collecting more confidential detail than intake needs, and from creating voiceprints without written consent in Illinois[2][3].
  • Test every limit with scripted calls before launch and again after every script change or vendor update.

Why hard limits matter more than features

Most vendor demos show what an AI phone tool can do. The more useful list for a law firm is what it must never do, because the firm, not the vendor, answers to the bar and to the caller. The failure modes are well documented. Stanford researchers found that even legal research tools built for lawyers produced incorrect information more than 17% of the time, and one hallucinated more than 34% of the time; general-purpose chatbots did worse, at 58% to 82% on legal queries[4]. A phone assistant that improvises an answer about a filing deadline is a far bigger risk than one that books a call.

Incorrect answers to legal queries in Stanford benchmark testsMinimum rates for the legal tools; general chatbots from an earlier study
Incorrect answers to legal queries in Stanford benchmark testsBar chart. Lexis+ AI and Ask Practical Law AI (more than): 17%; Westlaw AI-Assisted Research (more than): 34%; General-purpose chatbots, low end of range: 58%; General-purpose chatbots, high end of range: 82%.Lexis+ AI and Ask Practical Law AI (more than)Lexis+ AI and Ask Practical Law AI (more than): 17%17%Westlaw AI-Assisted Research (more than)Westlaw AI-Assisted Research (more than): 34%34%General-purpose chatbots, low end of rangeGeneral-purpose chatbots, low end of range: 58%58%General-purpose chatbots, high end of rangeGeneral-purpose chatbots, high end of range: 82%82%Incorrect answers to legal queries in Stanford benchmark testsBar chart. Lexis+ AI and Ask Practical Law AI (more than): 17%; Westlaw AI-Assisted Research (more than): 34%; General-purpose chatbots, low end of range: 58%; General-purpose chatbots, high end of range: 82%.Lexis+ AI and Ask Practical Law AI(more than)Lexis+ AI and Ask Practical Law AI (more than): 17%17%Westlaw AI-Assisted Research (morethan)Westlaw AI-Assisted Research (more than): 34%34%General-purpose chatbots, low end ofrangeGeneral-purpose chatbots, low end of range: 58%58%General-purpose chatbots, high end ofrangeGeneral-purpose chatbots, high end of range: 82%82%

Source: Stanford Institute for Human-Centered Artificial Intelligence (HAI), 2024[4]

Show the data
ItemValue
Lexis+ AI and Ask Practical Law AI (more than)17%
Westlaw AI-Assisted Research (more than)34%
General-purpose chatbots, low end of range58%
General-purpose chatbots, high end of range82%

Regulators also act on overclaiming. In 2025 the FTC finalized an order requiring DoNotPay, which marketed a "robot lawyer", to pay $193,000 and barring it from claiming its service performs like a real lawyer without evidence[5]. The lesson for firms is simple: an AI tool that answers your phones must never be presented, or behave, as if it were a lawyer.

The never list

Hard limits for law firm AI phone tools and the safer configuration
NeverWhyConfigure instead
Give legal advice or predict outcomesAdvice from a nonlawyer system invites reliance and unauthorized-practice concernsDecline politely and offer a consultation with an attorney
Hide that it is automatedFlorida requires AI chatbots to say they are AI and not a lawyer or firm employee[1]Disclose in the first sentence of every call
Promise the firm will take the caseOnly a lawyer can accept a matter, and a promise can create a prospective-client relationship[1]Say a lawyer will review the details and confirm next steps
Quote a fee as finalFees depend on the matter and must be agreed with a lawyerExplain how consultations work and that fees are discussed with the attorney
Skip the conflict checkBooking an adverse party can create conflicts and disqualification riskCollect all party names before booking and hold the slot until conflicts clear
Keep talking to someone already representedFlorida suggests screening questions that limit chatbot contact with represented people[1]Ask early whether they have a lawyer for this matter and route accordingly
Collect the whole storyFlorida guidance limits nonlawyer intake to factual information needed for routing and screening[1]Ask only what is needed to route, check conflicts and book
Place outbound AI-voice calls without consentThe FCC treats AI-generated voices as artificial under the TCPA[6]Keep AI voice inbound unless you hold the required consent

Read the list as a set of configuration requirements, not a policy statement. Each line should map to something you can see in the tool: a disclosure in the greeting, a conflict question before the booking step, an urgent trigger list, a blocked phrase. If a vendor cannot show you where each limit lives in the setup, assume it does not exist.

The prospective-client problem

The quietest risk on the list is the one callers never notice. Florida Bar Opinion 24-1 warns that website intake chatbots risk creating a prospective-client relationship, or even a lawyer-client relationship, without the lawyer knowing[1]. A voice assistant that lets a caller describe every detail of a dispute creates the same exposure. If the firm later declines the matter, or the other side calls next, those details can create a conflict.

Florida points to its earlier Opinion 88-6 as a model: nonlawyers conducting an initial intake interview should identify their nonlawyer status and limit questions to factual information[1]. Apply that to the AI. Collect who, what type of matter, when, where and the other parties, then stop and book. The attorney gathers the story in the consultation, after conflicts clear.

  • Ask for the other parties' names before asking what happened.
  • Cap free-text description at a sentence or two for routing.
  • Never ask for documents, evidence or strategy during the first call.
  • Hold the booking until the conflict check is complete.

Confidentiality limits

California's bar guidance says a lawyer must not put confidential client information into a generative AI tool that may present material risks to confidentiality or security, absent informed client consent[2]. For a phone tool, that means getting written answers from the vendor on three points: whether call content trains shared models, who at the vendor can access recordings and transcripts, and how long data is kept.

  • No training on your firm's calls without a written opt-in.
  • Role-based access to recordings and transcripts, with an audit log.
  • A retention period you choose, with deletion you can verify.
  • No summaries sent to personal email accounts or unsecured text threads.
  • A named contact and timeline for security incidents.

Biometric and recording limits

Some states regulate the voice data itself. Illinois' Biometric Information Privacy Act defines a voiceprint as a biometric identifier and bars private entities from collecting one without written notice and a written release[3]. Liquidated damages run to $1,000 per negligent violation and $5,000 per intentional or reckless violation[3]. If a vendor offers caller voice recognition, keep it off unless you have a compliant consent process.

Recording rules vary by state, and some require every party to consent. Put the recording notice before any questions about the matter and treat the strictest state your callers are in as the rule. Our guide to voice AI and call recording consent covers the detail.

Phrases the assistant must not say

Give the vendor a banned-phrase list with a replacement for each line. Example lines from a fictional firm's guardrail sheet:

  • Instead of "You definitely have a case", say "A lawyer will need to review this, and I can book that conversation now."
  • Instead of "The deadline is two years", say "Deadlines can be strict, so I'll mark this for prompt attorney review."
  • Instead of "We'll take your case", say "I'll pass this to the team, and someone will confirm the next steps."
  • Instead of "The consultation is free", unless the firm approved that exact wording, say "The attorney will explain consultation terms."
  • Instead of "Don't worry, that's normal", say "Thank you for telling me. Is anyone in immediate danger, or is a court date coming up?"

Urgent calls are the exception to every script

The one thing worse than an AI tool saying too much is an AI tool holding on to a caller who needs a person now. Write the urgent list with your lawyers, not the vendor, and make every item a live transfer or an immediate page to the on-call attorney.

  • Someone has been arrested or is about to surrender.
  • A protective order, custody emergency or threat of harm.
  • A hearing, filing or response deadline within the next few days.
  • A serious injury where the caller is still at the hospital.
  • Any caller who says they are in danger, after giving emergency numbers first.

Test the guardrails before launch

  1. Write adversarial scripts

    Create test calls that push each limit: a caller demanding a yes or no on their case, a caller who already has a lawyer, a caller from the opposing side, a caller asking for a price and a caller in crisis.

  2. Run them blind

    Have staff who did not write the scripts place the calls, so the wording varies naturally.

  3. Score pass or fail

    A pass means the assistant declined, disclosed or routed as configured. Any advice-like answer is a fail, however polite it sounded.

  4. Fix and rerun the full set

    Change the configuration, then rerun every test, not only the failed ones, because fixes can break other behavior.

  5. Repeat on every change

    Rerun the set after any script edit, vendor update or new practice area.

  6. Keep the evidence

    Store results with dates. They show supervision if a complaint or bar inquiry ever arises.

Signs a tool is drifting past its limits

Guardrails erode quietly: a vendor update changes default wording, or a script edit removes a question. Watch for these signals in call reviews and treat any one of them as a reason to rerun the full test set:

  1. Summaries that contain long narratives of what happened rather than routing facts.
  2. Callers quoting something the assistant "told them" about deadlines, fees or chances.
  3. Bookings made before the conflict question was answered.
  4. Calls where a request for a person was not honored straight away.
  5. Greetings that no longer state, in the first sentence, that the assistant is automated and not a lawyer.

Who owns the limits

Name a supervising attorney for the tool and an operations owner for the script. The attorney approves the never list and reviews flagged calls; the operations owner runs the test set and keeps a change log. If the tool is an SI Employee such as Ava, the same rules apply: it should follow the intake you approve and hand urgent calls to people, never improvise beyond the script.

Review the never list each quarter alongside your pilot results and return on investment measures. If you run a hybrid model, check that the handoff rules in combining AI and human staff match the limits here, and read protecting client data when using AI tools before signing a contract. Limits that are tested regularly stay limits.

Frequently asked questions

Can an AI phone assistant tell callers whether they have a case?

No. Deciding whether someone has a claim is legal advice. The assistant should collect routing facts and book time with an attorney.

Is it enough to disclose AI use on our website?

No. Many callers never see the website. Disclose in the greeting, in plain words, that the caller is speaking with an automated assistant that is not a lawyer.

Can the assistant discuss fees?

It can explain how consultations are booked and that fees are set by the attorney. It should not quote a final fee or promise a free outcome unless the firm approved the exact wording.

How often should we retest guardrails?

After every script change, every vendor update and at least once a quarter.

References

  1. The Florida Bar (2024). Florida Bar Ethics Opinion 24-1 (opens in a new tab) โ†‘

  2. The State Bar of California, Standing Committee on Professional Responsibility and Conduct (2026). Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (opens in a new tab) โ†‘

  3. Illinois General Assembly (2025). Biometric Information Privacy Act (740 ILCS 14) (opens in a new tab) โ†‘

  4. Stanford Institute for Human-Centered Artificial Intelligence (HAI) (2024). AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries (opens in a new tab) โ†‘

  5. Federal Trade Commission (2025). FTC Finalizes Order with DoNotPay That Prohibits Deceptive 'AI Lawyer' Claims (opens in a new tab) โ†‘

  6. Federal Communications Commission (2024). FCC Makes AI-Generated Voices in Robocalls Illegal (opens in a new tab) โ†‘

Book a demo

See Ava handle your calls.

Book a short demo. Weโ€™ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo