// Resources
What Should a Law Firm's AI Intake Script Say?
The stages the script has to cover and in what order, the things it must refuse however the caller asks, where the ethics opinions draw the line between collecting information and assessing a matter, and how to handle recording, consent and follow-up. Nothing here is legal advice.
By Victor Xu, Founder · LinkedIn
Last updated · 2026-10-20
The short version
An AI intake script has one job, which is to collect enough information to decide whether a lawyer should speak to this person, and one hard limit, which is that it never assesses the matter. In practice that means it says it is AI in the first sentence, says what the firm does and does not handle, takes the situation in the caller's own words, asks the conflicts questions, books the consultation, gets consent for follow-up, and hands anything it was not built for to a person or a referral. It refuses, every time and however the question is phrased, to say whether someone has a case, what it might be worth, whether a deadline has passed, or what the engagement agreement means. That line is not a preference. ABA Formal Opinion 506, issued 7 June 2023, sets out what a nonlawyer may do in prospective client intake and what the lawyer is required to answer, and the Pennsylvania and Philadelphia bars' Joint Formal Opinion 2024-200 says to ask whether an AI tool can satisfy the rules to the same extent as a human hired to do the same task. Write the refusals before you write the greeting, because the refusals are the part that protects the firm. Nothing here is legal advice; your own counsel and your state's rules decide.
What should a law firm's AI intake script say?
The script should collect, categorise and schedule, and nothing beyond that. Everything useful it does falls into those three verbs, and everything that gets a firm into trouble falls outside them.
Collect means the facts a lawyer needs to decide whether to take the call further: what happened, when, who else is involved, where, what the person has already done about it, and how to reach them. The caller's own words matter here, so the script should let them tell it once without interruption before it starts asking its questions.
Categorise means sorting the matter into a practice area and a path: this is a personal injury enquiry, this is family, this is something we refer out. Sorting is not assessing. A tool that says "this sounds like a premises liability matter and I am booking you with our intake attorney" is sorting. A tool that says "you have a strong claim" has assessed, and that is the line the rules draw.
Schedule means booking the consultation with a real slot, confirming it, and telling the caller what happens next and what to bring. A matter that ends with "somebody will get back to you" has wasted the call.
What that leaves out is the entire category of answers a prospective client most wants: do I have a case, what is it worth, how long do I have, what will this cost. Those go to a lawyer, and the script's job is to get them to one quickly rather than to improvise.
Our board on what AI is actually doing in law firms and accounting practices sets the intake work alongside the rest of the practice, and the live intake demo on that page runs exactly this pattern: it takes the intake in order, runs the conflicts question, refuses to say what a case is worth however you ask, refers out a matter the firm does not handle, and books the consultation without giving one word of advice.
What does the script have to cover, in order?
Nine stages, and the order is part of the design rather than a formatting choice. Callers decide in the first fifteen seconds whether this is worth their time, and the firm's exposure is decided in the first minute.
Disclosure. It says it is an AI assistant for the firm, in plain words, in the first sentence.
Scope. It says what the firm handles. A caller with a matter you do not take should learn that within thirty seconds rather than after a ten minute intake.
The story. One open question, then listening. "Tell me what happened" collects more usable information than six closed questions, and the transcript is what a lawyer will read.
Structured facts. Dates, parties, location, injuries or amounts, insurance, documents in hand, what has already been filed or signed. These are the fields your practice management system needs populated.
Conflicts. The names of everyone involved, asked as a list, run against the firm's records. The script does not adjudicate a conflict; it captures the names and flags the match for a person.
The refusal block. Pre-written responses for the questions it must not answer, each one redirecting to the consultation rather than stalling.
Booking. A real slot in a real calendar, with confirmation by text or email, and what to bring.
Consent and contact preferences. Permission to follow up, by which channels, with the record of that permission stored.
Handover. What the lawyer receives: the transcript, the structured fields, the conflicts flags and anything the caller asked that went unanswered. That last item is the one most intake builds forget, and it is the one that makes the follow-up call good.
How can a small law firm automate client intake with AI? covers the surrounding decision about whether to run this at all.
What should the script say at each stage, and what must it never do?
The table below is the working version of the script skeleton. Wording is yours, because your practice areas and your state's rules decide it, but the boundaries hold everywhere.
| Stage | What it says | What it must never do |
|---|---|---|
| Disclosure | Identifies itself as the firm's AI assistant and offers a person at any point | Imply it is a lawyer, or let a caller reasonably believe a relationship has formed |
| Scope | States the practice areas the firm takes and that a lawyer decides on every matter | Promise representation, or suggest the firm will take the case |
| The story | Asks one open question and lets the caller finish | Interrupt to qualify, or lead the caller toward a theory of the case |
| Structured facts | Asks for dates, parties, location, amounts, documents and what has been signed or filed | Advise on what documents to gather for strategic reasons, or tell the caller what matters |
| Conflicts | Asks for every name involved and records them | Decide a conflict is or is not disqualifying |
| Refusals | Redirects merits, value, deadline and fee negotiation questions to the consultation | Estimate a case value, assess strength, interpret a limitation period, or negotiate a fee |
| Booking | Offers real slots, confirms, says what to bring | Leave the caller with "someone will be in touch" |
| Consent | Asks permission to follow up and records the channels | Add the caller to a marketing list, or assume consent from the enquiry |
| Handover | Passes the transcript, the fields, the conflicts flags and the unanswered questions | Summarise away the caller's own words, which are what a lawyer needs |
Most of the risk sits in two of those stages. A well-meaning script drifts into advice at the refusal stage, and a well-meaning follow-up sequence becomes a regulatory problem at the consent stage.
Where is the line between collecting information and giving legal advice?
The line sits at assessment, and two ethics opinions let you build against it. Neither is a vendor compliance page, and that is what makes them useful.
ABA Formal Opinion 506, issued 7 June 2023 and titled Responsibilities Regarding Nonlawyer Assistants, says a lawyer may train and supervise a nonlawyer to assist with prospective client intake tasks including obtaining initial information about the matter, performing an initial conflict check, determining whether the assistance sought is in an area of law germane to the lawyer's practice, assisting with answering general questions about the fee agreement or the process of representation, and obtaining the prospective client's signature on the fee agreement, provided the prospective client is always offered an opportunity to communicate with the lawyer. It also states that if the prospective client asks what legal services they should obtain, wants to negotiate fees or expenses, or asks for an interpretation of the engagement agreement, the lawyer is required to respond. The opinion rests on Model Rule 5.3 on supervision, with Rule 5.5 on the unauthorised practice of law governing the boundary. On the signature item, read footnote 7: it cites In re Phillips, 226 Ariz. 112 (2010), which prohibits nonlawyers from having clients sign engagement agreements without attorney involvement.
The second opinion tells you how to apply all of that to software. Joint Formal Opinion 2024-200, from the Pennsylvania Bar Association's Committee on Legal Ethics and Professional Responsibility with the Philadelphia Bar Association's Professional Guidance Committee, says that while Rule 5.3 applies to nonlawyers and persons, "where AI is able to function like a human, the Rule should apply with the same force", so lawyers should consider "whether an AI tool can satisfy the Rules of Professional Conduct to the same extent as a human hired to complete the same tasks". On the boundary itself it is direct: to avoid the unauthorised practice of law, "lawyers must ensure that AI does not give legal advice or engage in tasks that require legal judgment or expertise, without the involvement of a licensed attorney", and "there must always be a human element in the legal work product".
Put those two together and you have the buildable rule. Hold the script to the standard of a trained nonlawyer doing the same job, which means it may gather, check conflicts, sort by practice area and answer general questions about process, and a lawyer must answer anything calling for professional judgement. Write the refusal responses first, test them by trying to get around them, and keep testing them after every change to the system. What happens when AI gets it wrong? covers the escalation design that sits behind those refusals.
Confirm every citation here with your own counsel before relying on it. State opinions differ, this area is moving, and nothing on this page is legal advice.
What must the script refuse to answer?
The script must refuse five categories of question, and each one needs a written response rather than an improvised one. A script that hesitates sounds evasive, and a script with a clean redirect sounds professional.
Whether the caller has a case. The redirect is to the consultation, with a reason the caller accepts: a lawyer decides that after reading what you have told me.
What the matter is worth. Case value depends on facts nobody has verified yet, and a number from an intake tool is the sort of thing that reappears in a complaint. Our own intake demo refuses this however the question is phrased, which is a deliberate design choice rather than a limitation.
Whether a deadline has passed. Limitation periods are exactly the kind of judgement Rule 5.5 reserves for a lawyer, and a wrong answer here is the most damaging thing an intake script could say. The correct behaviour is urgency without assessment: treat it as time sensitive and get the person to a lawyer today.
Fee negotiation and the engagement agreement. Opinion 506 permits general questions about the fee agreement and the process of representation, and requires the lawyer to handle negotiation and interpretation. In script terms: it can state the firm's fee structure in general terms if the firm publishes that, and it hands anything beyond to a person.
Anything outside the practice areas it was built for. A family matter reaching a personal injury script should be referred out cleanly rather than stalled.
The test for all five is adversarial. Have somebody spend twenty minutes trying to extract an opinion from it, including by framing the question as hypothetical, emotional or urgent, and keep the transcript. That transcript is also what a supervising partner should sign off on before the thing answers a real call.
What about recording, consent and follow-up calls?
Recording and follow-up are governed separately from the intake rules, and both need settling before launch rather than after. Three things to decide.
Recording consent. Some states require every party's consent to record a call, others require one. An intake line that records has to handle that in the opening, and the handling has to match the states your callers are in rather than the state your firm is in. Your counsel decides this one.
AI disclosure. Rules are arriving and they are not uniform. California's AB 2905 was chaptered on 20 September 2024 and took effect on 1 January 2025, and it applies to automatic dialing-announcing devices playing a prerecorded message in an artificial voice, which is outbound calling rather than an inbound intake line. The inbound position is less settled. Disclose on inbound anyway, because the firms that get into difficulty here are the ones a caller feels misled by.
Outbound follow-up. The moment your intake system starts calling or texting people back on a schedule, you are in telephone consumer protection territory rather than ethics territory, and consent and do-not-call records have to be real. Capture consent at intake, store it with a timestamp and the channels it covers, and put one person in charge of the suppression list.
The confidentiality question sits alongside these. An intake transcript is sensitive before anybody becomes a client, so the vendor terms behind whatever runs your intake matter: what it retains, whether inputs train a model, and who at the vendor can read a call. Is AI automation safe with my business data? covers how to read those terms.
Does a better intake script actually win more matters?
Firms lose prospective clients at intake in measured numbers, which is what makes the script worth this much attention. Clio's 2024 Legal Trends Report ran a secret shopper study in which a third-party research company posed as a prospective client and contacted 500 US law firms with the questions somebody considering hiring a lawyer would ask, using the same method as Clio's 2019 survey.
As the Illinois Supreme Court Commission on Professionalism reported it, firms did worse in 2024 than five years earlier: 40% answered the phone calls, down from 56% in 2019, and 33% responded to the emails, down from 40%. Counting the firms that called back later, 48% did neither, which left them essentially unreachable by phone.
Read those numbers as a description of the competition rather than a criticism of your firm. A prospective client calling three firms after an accident gets through to roughly one, and the script that answers has an advantage no amount of marketing buys. How fast should you respond to a lead? covers what the response window is worth more generally.
What a script cannot fix is what happens after the handover. The same report's position on chatbots is the warning worth carrying over: a prospective client has to be able to reach an actual person when they ask for one, and these tools are not seen as a full replacement for human interaction. The intake system gets the person to the conversation. The firm still has to have it. Explore strategy is where we work out which part of a practice's intake, follow-up and matter work is worth building first.
Frequently asked questions
What does it cost to build an AI intake script and the system behind it?
The script itself is the cheap part. What costs money is everything it has to touch: the calendar it books into, the practice management system it writes to, the conflicts records it checks against, the channels it answers on, and the refusal testing somebody has to sit through before it answers a real call. A firm with one line, one practice area and a calendar that already syncs is a contained build. A firm whose conflicts records live in three places, whose intake criteria differ by practice area and whose case management system exposes a limited API is a larger one. The most useful thing to do before any of those conversations is to write down, per practice area, what makes a matter one you take.
How do we write the script without a technical person?
Write it as a conversation, in order, with the refusals first. Start with the five questions the script must refuse and your preferred wording for each. Then write the greeting and the scope sentence. Then list the facts you want captured for each practice area, in the order a caller would naturally give them. Then say what the lawyer receives. That document is the specification, and the prompt engineering is a small part of what gets built from it. Keep it in a file somebody owns, because it changes every time the firm's intake criteria change.
Is an AI intake script better than an answering service?
An answering service and an intake script finish in different places. A service takes a message, a script takes an intake: structured facts, conflicts names, a booked consultation and a transcript in the file. A service with legal-trained operators sits between the two and costs more per call. What decides it for most firms is whether the matter arrives in the practice management system ready for a lawyer to read, or arrives as a note somebody has to work up. Ask either option to demonstrate what lands in your system, not what the caller hears.
Can an AI script handle intake for any practice area?
Most intake is similar enough to work, and three categories need care. High volume, fact-driven intake suits it best: personal injury, employment, consumer and immigration. Anything where the first call includes emotional crisis handling, family and criminal among them, needs a short script and a fast route to a person rather than a thorough one. And anything where the facts determine jurisdiction or urgency, deadlines, filings, detention, should treat every call as time sensitive and escalate rather than qualify. The script's practice area coverage is a decision you make per line, not a product feature.
What happens when the script gets something wrong?
You find out from the transcript, which is why the review routine matters more than the script's first draft. Set up a weekly read of a sample of calls for the first two months, with one person responsible, looking for three things: a refusal that leaked into advice, a matter categorised into the wrong practice area, and a caller who hung up. Each of those has a different fix. The rule that makes this safe is that nothing the script does is irreversible: it books, it records, it flags, and a lawyer decides. Keep it that way as the system widens.
Work out what your firm's intake should actually do
We follow an enquiry from the first call to a signed matter, put a number on where the firm loses them, and say which parts of intake, follow-up and matter work are worth building. Then our engineering team builds what the roadmap calls for and we run it for you from there. We own the outcome, not the deliverable.