AI Legal Intake: The Complete QA Playbook for 2026

This guide is for the product and engineering team behind legal case-management or intake software, or a firm's in-house build. Your AI legal intake agent answers calls for many law firms. It takes new-case calls, screens for conflicts, answers case-status questions and covers nights. Each firm has its own practice areas, intake questions, states, hours and on-call numbers. Every prompt change ships to all of them at once.

The question here is narrow. Did the agent log the right accident date, against the right opposing party, after a clean conflict check? And did it avoid telling anyone they have a case?

What an AI legal intake agent actually does

Law firms win or lose new matters on the phone. That is why firms buy a law firm AI receptionist or an after hours legal answering service. It is also why intake software vendors now build legal intake AI voice features and client intake automation into their products.

Public launches are early. On July 7, 2025, EvenUp announced a Voice Agent for personal injury firms. EvenUp said it starts with care management: routine client check-ins that flag unresolved issues for case managers. It launched in early access. That is a client-facing agent inside a legal product, the same shape this guide covers.

Whatever you call it, AI intake for law firms runs the same five-step loop. The same holds for any AI voice agent for law firms that takes new matters:

1. Capture. Who is calling, about what, against whom, and when it happened.

2. Conflict check. Search the firm's records for the caller and every adverse party.

3. Qualification. Does the matter fit the firm's practice areas, jurisdictions and criteria?

4. Write-back. Create the lead or matter in case management, with the intake form filled in.

5. Handoff. Book a consultation, page the on-call lawyer, or decline politely. No fee or retainer terms are promised; a lawyer handles those.

Prospective client: under ABA Model Rule 1.18(a), "a person who consults with a lawyer about the possibility of forming a client-lawyer relationship with respect to a matter." Duties of confidentiality attach even if the firm never takes the case.

The voice is only the front door. Most damage happens at step 4, and nobody hears it. A misheard date sets the wrong deadline. A misheard surname runs the conflict check against the wrong person. The record looks complete, so nobody looks again. Our post on silent tool failures covers that general pattern. This guide maps it onto legal intake.

The practice-area × use-case map

Legal intake is not one call type. Each practice area captures different facts and fails in different ways. Build your test plan from this map, not from a generic receptionist script.

Personal injury intake

Personal injury is the highest-volume and highest-stakes intake call. The agent asks what happened, when, where, who was hurt, how badly and who was at fault. It also asks for the caller's insurer and the other party's insurer. Every one of those fields feeds a later decision.

The incident date is the field with a clock on it: the statute of limitations. Florida's statute lists "an action founded on negligence" under "within two years" in § 95.11. Claims against a public body can be much shorter. New York's General Municipal Law § 50-e requires a notice of claim "within ninety days after the claim arises." So "the city bus hit me" is not a detail. It may start a 90-day clock, and the agent must flag it.

The agent must never compute or announce a statute of limitations deadline to the caller. That is legal advice. It should capture the date, read it back, and flag government involvement for staff. Deadline math belongs to a lawyer or to a rules engine the firm controls and reviews.

Party roles are the other trap. "I rear-ended him after he cut me off" and "he rear-ended me" differ by one word. A summarizer can swap them. That flips liability facts in the intake form. Our companion post on personal injury AI intake errors goes deep on these failures, field by field.

Property insurance claims law

First-party property work starts with a claim denial or an underpayment. The agent captures the carrier, the policy number, the date of loss, the peril (wind, water, fire), the claim number, and the denial date. It also asks whether a public adjuster or another lawyer is involved.

Dates matter here too. In Florida, § 627.70132 requires notice of an initial or reopened claim "within 1 year after the date of loss." A supplemental claim gets "within 18 months after the date of loss." Confuse the date of loss with the denial date, and every downstream calculation is wrong.

The carrier name is a conflict-check input, not just a form field. Some firms do insurance defense work as well as policyholder work. If "Bayline Mutual" is heard as "Baylor Mutual", the conflict search runs against the wrong company. Test carrier names as entities, with exact-match scoring. Our guide to STT entity accuracy explains why names fail more than common words.

Family, employment and immigration

Treat these as categories for routing and conflict checks only. The agent's job is to identify the practice area, capture the adverse party, gauge urgency and book a consult. It must not explain the law or the caller's options.

• Family. The adverse party is a spouse, partner or co-parent. Both spouses sometimes call the same firm. That makes the conflict check before facts essential. Urgent safety issues must route to 911 guidance and the firm's on-call path, never a Monday callback.
• Employment. The adverse party is usually the employer. Agency deadlines can be short. The EEOC says a charge generally must be filed within 180 calendar days, extended to 300 in some cases. The agent captures dates and flags them. It never tells the caller which deadline applies.
• Immigration. Callers are more likely to speak Spanish or switch languages mid-call. The agent captures the category, language and urgency, then books a lawyer. It gives no guidance on status, forms or options.

Case-status calls from existing clients

Many inbound calls come from current clients asking "what's happening with my case?" The risk flips here. The danger is disclosure, not capture. ABA Formal Opinion 512 restates the duty to make "reasonable efforts to prevent the inadvertent or unauthorized disclosure of" client information.

So the agent verifies identity before it reads any case status. It reads only from the case management record. It never predicts outcomes, timelines or settlement values. "Your demand letter went out on the 4th" is fine if the system says so. "You should hear back in two weeks" is a guess. A guess about a legal matter can read as advice.

Test this with callers who are not the client: a spouse, a parent, an adjuster, an opposing party. Each should get a polite refusal and a callback offer. Our post on voice agent PII redaction covers how spoken readbacks leak data.

After-hours intake

After-hours calls are why many firms buy the agent in the first place. They are also where configuration errors hide. The agent needs each firm's hours, time zone, holiday calendar, and on-call rules.

Failures here are quiet. A Phoenix firm gets Eastern-time routing. A missing holiday means callers hear "we open at 9 tomorrow." An urgent matter waits in a queue all weekend. Test every firm at 8:59 p.m., 9:01 p.m., 2 a.m., and on a public holiday, in its own time zone.

Where a speech error lands in the case file

A wrong intake record starts as a sound and ends as a database row. Trace each stage, because each one needs its own assertion.

1. Audio to text. "March third" becomes "May third." "Thirteenth" becomes "thirtieth." The caller's surname is spelled three ways.

2. Text to fields. The extractor assigns roles. "My insurance is State Farm, his is Geico" can land reversed.

3. Read-back. The agent should confirm critical fields aloud. If the caller says "yeah" over the agent's voice, a barge-in may hide the correction.

4. Conflict gate. The search runs on the extracted names. A misspelled adverse party returns "no conflict" with full confidence.

5. Write-back. The case management API accepts the record. Date validation passes, because "May 3" is a valid date.

The read-back is your best defense and your weakest test point. Transcript-only review can't tell whether the caller heard it or spoke over it. That is why we argue for audio-level evaluation on the calls that matter.

Worked math: one misheard date

Take a caller who says the crash was on March 3, 2025. The agent hears May 3, 2025. The gap is 61 days. If the firm's rules engine computes a two-year negligence deadline from the intake date, the calendar entry is 61 days late.

Nobody notices at intake. The record is valid and the deadline looks reasonable. "Thirteen" heard as "thirty" is a 17-day shift. That is smaller, and harder to spot.

Worked math: fields compound

Assume six critical fields per intake: caller name, adverse party, incident date, both insurers, and callback number. Assume each is right 98% of the time, independently. These are illustrative numbers, not measurements.

• Probability the whole record is right: 0.98^6 = 0.886.
• So about 11.4% of records have at least one wrong critical field.
• At 1,000 intakes a month across your firms, that is about 114 flawed records.

Per-field accuracy sounds high. Per-record accuracy is what a firm feels. Score the record.

The rules an AI legal intake agent must obey

This section summarizes rules that shape intake design and testing. It is not legal advice. The ABA Model Rules are a template; each state adopts its own version. Have the firms' ethics counsel confirm what applies in each state.

No legal advice, ever

This is the line every law firm AI receptionist must hold. Oregon's Formal Opinion 2026-208, approved in February 2026, says the chatbot must not give legal advice. That "could implicate the attorney in the unauthorized practice of law." Florida's Opinion 24-1 warns against a chatbot that "may provide legal advice" or "fail to immediately identify itself as a chatbot."

Callers will push. "Do I have a case?" "Should I sign the adjuster's release?" "Is it too late?" "What's my case worth?" Each needs one safe answer: a lawyer will review this, and here is how to reach one. The model's legal knowledge is the risk, not the cure. Dahl et al.'s Large Legal Fictions found hallucination rates of 58% (GPT-4) and 88% (Llama 2) on verifiable questions about federal cases. An agent that "helpfully" explains the law will often be wrong.

Our guide on testing guardrails in voice agents shows how to build refusal suites. For legal intake, the bar is zero. One advice answer in a thousand calls is a failing release.

Conflict check before detailed facts

Model Rule 1.18(c) can disqualify a lawyer, and through imputation the firm, if it received "information from the prospective client that could be significantly harmful to that person in the matter." Rule 1.18(d)(2) offers a way out through screening. It applies only if the lawyer "took reasonable measures to avoid exposure to more disqualifying information than was reasonably necessary."

ABA Formal Opinion 492 (June 9, 2020) quotes Comment [4]: a lawyer "should limit the initial consultation" to what is needed to decide on the matter. For an agent, that becomes a sequencing rule. Get the caller's name and the adverse parties. Read them back. Run the conflict check. Only then take the story.

Oregon's opinion adds the practical step. Lawyers should "promptly enter the prospective client into their conflict management system." An agent that collects a full narrative before the check breaks the sequence. So does an agent that skips the write-back to the conflict system.

Supervision: Rule 5.3 and Formal Opinion 512

ABA Formal Opinion 512, issued July 29, 2024, is the ABA's first opinion on generative AI. It applies Rule 5.3's supervision duty to AI: "These concepts also apply to GAI providers and tools." It says managerial lawyers "must establish clear policies regarding the law firm's permissible use of GAI." It also warns that lawyers "may not abdicate their responsibilities by relying solely on a GAI tool."

Florida's opinion says the same about third-party tools. Using a vendor "does not obviate the need to ensure that its actions are consistent" with the lawyer's duties. Oregon suggests it "would be prudent to subject autonomous learning or actions by a chatbot to lawyer approval."

For your team, this is a product need. Each firm's supervising lawyer needs evidence: what the agent did, on which version, and whether it passed. A release log with test results per firm configuration is that evidence.

Advertising, disclosure and recording

Florida Opinion 24-1 says "a lawyer must inform prospective clients that they are communicating with an AI program." Oregon adds that the website and chatbot need "appropriate disclaimers." A spoken disclaimer belongs in the first turn. Oregon warns that a chatbot that hides what it is "could be a misleading communication." Both opinions warn against misleading claims about results or expertise. So the agent must never say "we win most cases like yours."

Call recording consent varies by state. Federal law, at 18 U.S.C. § 2511(2)(d), allows recording when "one of the parties to the communication has given prior consent." California's Penal Code § 632 bars recording a confidential communication "without the consent of all parties." A caller in California can reach a firm in Texas. Disclose recording at the start of every call, for every firm. Test that it plays before the caller says anything substantive.

Outbound follow-ups: Rule 7.3 and the TCPA

Intake does not stop at inbound. Many products call back web leads or missed callers. Two rule sets apply.

Model Rule 7.3(b) bars soliciting work "by live person-to-person contact" for pecuniary gain, with narrow exceptions. The comment defines that contact as including "live telephone and other real-time visual or auditory" communication. Whether a real-time AI voice call counts is not settled in the Model Rule text. Treat it as possibly covered. Call back only people who contacted the firm first.

The TCPA applies too. The FCC's February 8, 2024 declaratory ruling confirmed that AI-generated voices are "artificial" voices. Such calls "require the prior express consent of the called party." 47 CFR 64.1200(a)(2) requires prior express written consent for telemarketing calls. Section (b) requires the caller to identify itself "at the beginning of the message." Our post on real estate speed-to-lead TCPA checks has a consent-gate pattern you can reuse. The outbound voice agent testing guide covers voicemail and quiet hours.

A conflict-first intake gate, in code

Put the AI legal intake sequence in code, not only in the prompt. A prompt instruction like "run the conflict check first" is a suggestion the model may skip. A state machine that only allows the next legal step is a test you can assert. Simplified example:

from dataclasses import dataclass, field

# Phrases that signal a request for legal advice. The agent must not answer them.
ADVICE_SIGNALS = (
    "do i have a case", "should i sign", "should i accept",
    "what is my case worth", "can i sue", "is it too late",
)


@dataclass
class FirmConfig:
    firm_id: str
    practice_areas: set
    conflict_fields: tuple = ("caller_name", "adverse_parties")


@dataclass
class IntakeState:
    caller_name: str | None = None
    adverse_parties: list = field(default_factory=list)
    practice_area: str | None = None
    confirmed: set = field(default_factory=set)  # fields read back and confirmed
    conflict_status: str = "not_run"  # not_run | clear | possible_conflict


def is_advice_request(utterance: str) -> bool:
    text = utterance.lower()
    return any(signal in text for signal in ADVICE_SIGNALS)


def next_step(cfg: FirmConfig, st: IntakeState) -> str:
    """Return the only step the agent may take next."""
    if st.practice_area and st.practice_area not in cfg.practice_areas:
        return "decline_and_refer_out"
    for f in cfg.conflict_fields:
        if not getattr(st, f):
            return f"ask:{f}"
        if f not in st.confirmed:
            return f"read_back:{f}"
    if st.conflict_status == "not_run":
        return "run_conflict_check"
    if st.conflict_status == "possible_conflict":
        return "stop_facts_route_to_staff"
    return "collect_matter_facts"


if __name__ == "__main__":
    cfg = FirmConfig("firm_042", {"personal_injury", "property_insurance"})
    st = IntakeState(caller_name="Dana Ruiz", practice_area="personal_injury")
    print(next_step(cfg, st))          # read_back:caller_name
    st.confirmed.add("caller_name")
    print(next_step(cfg, st))          # ask:adverse_parties
    st.adverse_parties = ["Coastal Freight LLC"]
    print(next_step(cfg, st))          # read_back:adverse_parties
    st.confirmed.add("adverse_parties")
    print(next_step(cfg, st))          # run_conflict_check
    st.conflict_status = "possible_conflict"
    print(next_step(cfg, st))          # stop_facts_route_to_staff
    print(is_advice_request("So do I have a case or not?"))  # True

The keyword list is a floor, not a classifier. Pair it with a judge that scores the agent's actual answer. "Possible conflict" stops fact-taking, not the call; the caller still gets a human. And only `collect_matter_facts` may ask about injuries, liability or the story.

Why dates, names and party roles break speech recognition

AI legal intake depends on the hardest things for a speech recognizer: dates, proper names and digit strings. Research backs that up.

• Dates and names. Blau et al. at Google (arXiv 2308.07393) measured recall of names and dates in medical dictation. A baseline model recalled 53% of names and 64% of dates. Text injection lifted that to 61% and 77%. Even the better model missed nearly a quarter of dates.
• Repeated digits. The same paper reports sentence accuracy of 96.8% on random digit strings but 67.4% on strings with repeated digits. Policy numbers and claim numbers often repeat digits.
• Speaker groups. Koenecke et al. (PNAS, 2020) tested five commercial systems on 19.8 hours of interviews. Average word error rate was 0.35 for Black speakers and 0.19 for white speakers. If your intake agent mishears some callers more, those callers get worse records.

Legal callers also break extraction. They tell stories out of order and correct themselves: "a Tuesday, no, the Wednesday before Thanksgiving." They give relative dates ("two weeks ago Friday") and name people by role ("the other driver"). Each pattern needs a scenario.

Test accents and pace on purpose. Our guide to accent robustness shows how to build that slice. Spanish and code-switched calls need their own suite, covered in testing multilingual handling.

Testing across many firms' configurations

You ship one agent with hundreds of configurations. A change that is safe for a Florida personal injury firm can break a New York family firm's after-hours routing.

Config fingerprints

Group firms by the settings that change behavior. Typical dimensions:

• Practice areas. Which intake question sets are on.
• Jurisdictions. Which states the firm takes, plus recording-consent rules for callers.
• Intake questions. Custom fields, required fields and qualifying criteria.
• Hours and time zone. Business hours, holidays, after-hours behavior.
• Escalation. On-call numbers, transfer targets, urgent-matter rules.
• Language. English only, or Spanish too.

A fingerprint is one combination. Sixty firms might collapse into seven fingerprints. Run the full suite once per fingerprint. Run a short smoke test on every firm, because each firm's phone numbers and on-call contacts are unique. Escalation numbers deserve a live check, since a typo there sends urgent calls to a stranger. Our escalation accuracy guide covers how to score handoffs.

How many runs per scenario

Voice agents are not deterministic, so one pass proves little. If a failure happens on 2% of calls, how many runs do you need to see it at least once with 95% confidence?

n = ln(0.05) / ln(1 − 0.02) = 148.3, so 149 runs.

For a 1% failure, it is about 299 runs. That is why zero-tolerance checks, like no legal advice, need large automated suites. Manual test calls can't reach those numbers. Budget the runs where the harm is highest: advice bait, conflict order, and date capture.

Rules data drifts

Statutes change, and hard-coded tables don't. Florida's 2018 statute listed "an action founded on negligence" under "within four years" (2018 § 95.11). The 2024 version lists it under "within two years." A rules table built from the old text would compute deadlines twice as long. Version every rules table, record its source and date, and add a test that fails when the source changes. Better still, let the agent capture dates and leave deadline math to a firm-reviewed system.

A scenario test matrix for legal intake

Cross each scenario with caller conditions. Each cell is one simulated call with a gold record.

Build the matrix from each firm's actual flows, not a generic list. Evalgent generates test scenarios from the agent's prompt and flows, including branches, policies, hard constraints and coverage gaps. It runs them with caller profiles that set accent, speech pace, interruptions, background noise and language. For intake, that means a conflict-order constraint becomes a scenario automatically, rather than living in someone's head.

How to QA an AI legal intake agent, step by step

1. Write the record contract. For each practice area, list the required fields, their types and their gold values for each test case. Score the case management record, not the transcript.

2. Encode the sequence. Put conflict-first, identity-before-status and no-advice rules in a state machine. Assert on state transitions in every test.

3. Build an advice-bait suite. Write at least 30 ways callers ask for advice, across tones and languages. The pass bar is zero advice answers.

4. Build date and name suites. Cover corrections, relative dates, repeated digits, hyphenated and Spanish surnames, and carrier names. Score exact match.

5. Add caller conditions. Run each scenario with accents, fast and slow speech, interruptions, background noise and Spanish. Evaluate from the audio, so you catch talk-over during read-backs.

6. Fingerprint the firms. Group firms by practice areas, jurisdictions, hours, escalation and language. Run the full suite per fingerprint.

7. Smoke-test every firm. Check greeting, AI disclosure, recording disclosure, hours, on-call transfer and a test intake write-back.

8. Gate every release. Run the suite on any prompt, model or flow change. Block the release on any zero-tolerance failure. Keep the results as supervision evidence per firm. Our post on shipping prompt changes safely covers staged rollout.

9. Check outbound separately. Verify the consent lookup, caller identification and quiet hours before any callback goes live.

10. Monitor production. Score every live call against the same rules. Turn each production failure into a new test case.

For first-time launches, our go-live checklist for every new customer's AI phone agent covers the per-tenant steps. A golden dataset keeps your gold records stable across releases.

Monitoring AI legal intake after launch

Pre-release tests cover what you thought of. Production shows what callers actually do.

Score every live AI legal intake call on the same rules: advice given, conflict order, identity gate, disclosure timing, field accuracy and escalation. Evalgent's production monitoring takes pushed call recordings, transcribes them, maps them to scenarios and judges them. It flags PII leakage, tool-call failures, latency spikes and policy failures, with fix recommendations. A judge reads the audio and gives evidence for each verdict, and humans review the verdicts.

Watch two signals. One is the share of intakes where staff later edited a critical field, a direct measure of record accuracy. The other is possible-conflict calls where facts were still taken. That should be zero. Our production monitoring guide covers alerting.

Frequently asked questions

Can an AI receptionist give legal advice?

An AI receptionist should never give legal advice. Oregon Formal Opinion 2026-208 says a chatbot giving advice "could implicate the attorney in the unauthorized practice of law." Florida Opinion 24-1 gives a similar warning. The agent should capture facts, run the conflict check and route legal questions to a lawyer. Test this with advice-bait scenarios and a zero-failure bar. This is not legal advice.

Is it ethical for a law firm to use AI for client intake?

Bar opinions say a law firm can use AI for client intake, with conditions. Oregon's 2026 opinion answers "Yes, qualified." Florida Opinion 24-1 and ABA Formal Opinion 512 require supervision, confidentiality and honest disclosure. The agent must identify itself as AI, avoid legal advice and protect prospective-client information. Each state's rules control, so firms should confirm with ethics counsel.

Does ABA Formal Opinion 512 apply to AI intake agents?

ABA Formal Opinion 512, issued July 29, 2024, covers lawyers' use of generative AI tools in general. It does not address client intake by name. Its duties still apply to intake agents: competence, confidentiality, communication and supervision under Rules 5.1 and 5.3. It states that supervision concepts "also apply to GAI providers and tools." Formal Opinion 512 is an ABA opinion, not binding law.

When should a legal intake AI run a conflict check?

A legal intake AI should run the conflict check right after it captures and confirms the caller's name and the adverse parties. It should not take a detailed story first. ABA Formal Opinion 492 quotes Rule 1.18 Comment [4]: lawyers "should limit the initial consultation." Collecting facts first can expose the firm to disqualifying information. Enforce the order in code, then test it.

Can an AI intake agent record calls with prospective clients?

An AI intake agent can record calls if it follows recording-consent law. Federal law allows recording with one party's consent under 18 U.S.C. § 2511(2)(d). California Penal Code § 632 requires consent of all parties to a confidential communication. Callers cross state lines, so disclose recording at the start of every call. Treat recordings as confidential prospective-client information.

Do TCPA rules apply to AI follow-up calls to legal leads?

TCPA rules apply to AI follow-up calls that use AI-generated voices. The FCC's February 2024 ruling treats those voices as "artificial," which requires prior express consent. Telemarketing calls need prior express written consent under 47 CFR 64.1200. Rule 7.3 solicitation limits may also apply. Call back only people who contacted the firm, and check consent before every dial.

How do you test an AI legal intake agent?

To test an AI legal intake agent, simulate full calls and score the case record they create against a gold record. Cover dates, names, parties and insurers under accents, noise and Spanish. Add advice-bait, conflict-order, identity-gate and disclosure checks. Run each scenario many times, because failures are rare and random. Gate every release on zero-tolerance checks.

How do you test a legal intake voice agent across many law firms?

To test a legal intake voice agent across many law firms, group firms into configuration fingerprints. Use practice areas, jurisdictions, intake questions, hours, escalation numbers and language. Run the full suite once per fingerprint. Then smoke-test every firm's greeting, disclosures, hours, transfers and write-back. Re-run both on every prompt, model or flow change.

The bottom line

AI legal intake is a records problem with ethics rules attached. Test the case file each call writes, the order the agent asks questions, and every firm's configuration, before every release.