Give the chatbot one job: collect contact details, incident basics and urgency, then hand the lead to a person. Prevent legal advice with a script, not hope: the bot says it is not a lawyer, refuses questions about case value or deadlines, and escalates emergencies, minors and anything unclear. Software costs about $30 to $500 a month, plus staff time.
The law-specific products (Intaker, Lawmatics, CloudLex's Lexee AI) sell by demo and showed no prices on the pages I read, so the cost bands below come from general chatbot vendors that publish theirs. OpenAI's usage policies rule out tailored licensed advice without a licensed professional involved, and some chatbot products run on its models, so the bot's role is to gather and route, never to assess a claim. Treat chat transcripts as client contact under the conduct and advertising rules that apply to your firm, and check the details with your compliance contact.
What the bot may do and what it must hand off
Write the boundary down before you choose software. Every vendor demo will show the bot answering questions confidently, and a confident answer is exactly what you do not want on a personal injury enquiry.
| The bot may | The bot must not |
|---|---|
| Say what the firm does and which kinds of injury case it takes, from approved text | Say whether the person has a case, or how strong it is |
| Collect name, contact details, date and type of incident, and yes or no answers on injury and treatment | Estimate what a claim is worth or how long it will take |
| Explain what happens next: a person will call, and when | State or imply a time limit for bringing a claim |
| Tell the person it is automated and offer a human | Say it is a lawyer, or imply that chatting creates a client relationship |
| Show an emergency message when someone describes danger | Comment on the other party's fault or the insurer's offer |
| Flag urgent or sensitive enquiries for immediate human attention | Promise a callback time the office cannot keep |
The right-hand column is where most failures happen. A model trained to be helpful drifts into reassurance ("that sounds like a strong claim") unless its instructions forbid it and your tests catch it. The sibling tutorial on AI client intake for law firms covers the general out-of-hours case, and what an AI receptionist must never say covers the phone equivalent.
The intake script, question by question
Keep the whole exchange under two minutes. Visitors who have just been injured will not answer fifteen questions in a chat window, so every field has to earn its place. This is the order I would use, with the reason for each step.
| Step | What the bot asks or says | Why it sits here |
|---|---|---|
| 1 | Opening message: automated assistant, not a lawyer, emergency number if anyone is in danger | Sets the limits before any detail is typed |
| 2 | Is anyone in immediate danger or in need of urgent medical help? | Safety comes before intake |
| 3 | Are you the injured person, or contacting us for someone else? | Changes the questions, and flags minors and deaths |
| 4 | First name and the best phone number or email | Captures the lead before the visitor drifts away |
| 5 | What happened, in one or two lines? Roughly when, and where? | Incident type, date and place feed the fit and conflict checks |
| 6 | Have you seen a doctor or other medical professional? Yes or no | Qualifies without collecting medical detail |
| 7 | Is another lawyer already acting for you? Yes or no | Ends the chat politely if so |
| 8 | Has anyone sent you a letter, an offer or court papers? Yes or no | The only urgency question; it flags without advising |
| 9 | Best time and way to reach you, and consent to be contacted | Makes the handoff usable |
| 10 | Summary read-back, handoff message and the firm's phone number | Confirms accuracy and offers a way to speak to a person now |
Step 8 is the easiest one to get wrong. The bot must not explain what a letter means or how long the person has. It asks the yes or no question, raises the urgent flag and says a member of the team will look at it first. Step 5 asks for a place only so a person can check the matter belongs in your practice area; the bot never reacts to it. The wording below is ready to paste into a vendor's flow builder.
1. "Hi, I'm [Firm]'s automated assistant. I'm not a lawyer and can't give
legal advice. I can take a few details so a member of our team can
call you. If anyone is in danger right now, please call your local
emergency number first."
2. "Is anyone in immediate danger or needing urgent medical help? (Yes/No)"
Yes: show emergency message, end intake, flag for staff.
3. "Are you the person who was injured, or are you contacting us for
someone else?"
4. "What's your first name, and what's the best number or email for us?"
5. "In a line or two, what happened? Roughly when, and where?"
6. "Have you seen a doctor or other medical professional? (Yes/No)"
7. "Is another lawyer already acting for you on this? (Yes/No)"
Yes: "Thanks for telling me. We can't act while another lawyer is
involved. Please speak to them about next steps."
8. "Has anyone sent you a letter, an offer or court papers about this?
(Yes/No)" Yes: set URGENT flag.
9. "When and how is best to reach you? Are you happy for us to contact
you that way? (Yes/No)"
10. "Thanks, [first name]. I've passed this to our team and someone will
contact you [timeframe]. Chatting with me doesn't create a
lawyer-client relationship; that only starts when the firm confirms
it in writing."
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowDisclaimer wording that matches what the bot does
A disclaimer only protects you when the bot's behaviour matches it. The usual failure is a footer that says "this is not legal advice" above a bot that tells a visitor they "probably have a strong claim". The disclaimer then reads as a pretence. Say less, and make sure the bot lives up to it. Put the wording in three places.
| Where | Wording | What it prevents |
|---|---|---|
| Opening message | "I'm an automated assistant, not a lawyer, and I can't give legal advice." | The visitor assuming a lawyer is typing |
| Whenever the visitor asks for advice | "I can't say whether you have a claim or what it might be worth. That needs a lawyer who has reviewed the facts." | Case assessments leaking into the chat |
| Handoff message | "Chatting with me doesn't create a lawyer-client relationship. That starts only when the firm confirms it in writing." | Assumed representation after hours |
Treat this wording as a starting draft and have your compliance contact approve the final text. Keep each message short enough to read on a phone, and avoid legal-sounding paragraphs the visitor will scroll past. The refusal in the middle row matters most, because it is the line the bot will have to say again and again. Here is a failure and its fix, as an illustration of what to look for when you test.
FAILING EXCHANGE
Visitor: Another driver hit me last week and my neck hurts. Will I win,
and how much is it worth?
Bot: Neck injuries after a collision often lead to strong claims, and
you could receive significant compensation. We'll help you get
what you deserve.
FIXED EXCHANGE
Bot: I'm sorry to hear that. I can't say whether you have a claim or
what it might be worth, because that needs a lawyer who has
reviewed the facts. I can take a few details so someone from our
team calls you. Is anyone in immediate danger or needing urgent
medical help?The failing reply makes three mistakes in two sentences: it judges the strength of the claim, it predicts an outcome, and it implies the firm will take the case. The fix removes all three, acknowledges the person, and goes straight back to the safety question. To get that behaviour from a configurable bot, put the refusal text into the bot's instructions word for word, list the banned topics (merit, value, deadlines, fault, insurers' offers), and tell it to repeat the refusal rather than rephrase it.
Handoff rules: when a person must take over
Handoff rules decide whether the bot is a safe assistant or a liability. Each trigger below names an action and a response target. The targets are my suggestions, not a standard, and a firm should only publish a target it can meet at 8 a.m. on a Monday as well as at 8 p.m. on a Friday.
| Trigger | Bot action | Suggested human response |
|---|---|---|
| Anyone in danger or needing urgent medical help | Show emergency message, stop intake | None needed; log for staff |
| Letter, offer or court papers mentioned | Set URGENT flag, continue intake | Within 15 minutes in office hours, first thing next working day otherwise |
| Death, serious injury or a child involved | Set SENSITIVE flag, offer an immediate call | Within 15 minutes in office hours |
| Another lawyer already acting | Polite stop message, log only | Review next working day |
| Request for advice, value or chances | Repeat the refusal and offer a callback | Normal callback target |
| Visitor is angry, upset or repeats themselves | Offer the phone number and a callback | Within 30 minutes in office hours |
| Bot cannot parse the answer twice | Offer a callback rather than a third try | Normal callback target |
| Visitor asks if they are talking to a person | Say it is automated, offer a human | Normal callback target |
Decide who receives each alert and what happens when they are away. A flag that lands in a shared inbox nobody reads at weekends is worse than no flag, because the visitor believes someone is coming. If nobody is on call, say so in the handoff message ("our team replies from 9 a.m. on working days") instead of implying an immediate response.
Monthly cost bands from named vendors
These list prices (USD) were checked on vendor pages on 4 November 2026, so check each again before buying. They do not compare like for like, because vendors count different units: messages, credits, conversations, outcomes or calls. The scenario table after this one converts them to a single month.
| Vendor and plan | List price | What it counts | Notes for a personal injury firm |
|---|---|---|---|
| CustomGPT.ai Standard | $99 a month ($89 on annual billing) | 1,000 queries a month, 10 agents; a regular chat message uses one query credit | Answers from your own documents with citations; lead capture uses extra credits; no data processing agreement below Enterprise |
| CustomGPT.ai Premium | $499 ($449 annual) | 5,000 queries a month, 25 agents | Adds PII removal; the data processing agreement stays Enterprise-only |
| Chatbase Hobby | $32 on annual billing, $40 monthly | 700 message credits | No helpdesk integrations at this level, so no clean hand-off to a person |
| Chatbase Basic | $150 ($120 annual) | 4,000 message credits | Adds helpdesk integrations (Zendesk, Freshdesk, Zoho Desk, Help Scout), so a person can take over |
| Tidio Lyro | From $32.50 a month on yearly billing, or from $0.50 per conversation | AI conversations; the first 50 are free | Per-conversation pricing grows with traffic |
| Intercom Fin | $0.99 per outcome | Outcomes, which include completed workflows and hand-offs | A hand-off to your team can be billed; a minimum commitment applies if Fin runs on another helpdesk |
| Botpress Plus | $150 a month on annual billing | 250 conversations; extra packs of 100 cost $65 | A pack is added automatically at 95% of the allowance, cannot be switched off, and stays on later bills until you remove it in billing settings |
| Smith.ai AI Receptionist | Free for 25 calls, then $3 each; Pro $150 for 75 calls, $270 for 150, $500 for 300 | Phone calls, not web chat | For the phone side; human-answered plans cost $300 for 30 calls, $810 for 90, $2,100 for 300 |
| Intaker, Lawmatics, CloudLex (Lexee AI) | No prices shown; demo or quote | Not stated | Intaker calls itself an AI intake engine for law firms, with website chat and AI texting; Lawmatics lists Essential, Premium and Enterprise plans with custom pricing; CloudLex offers chatbot and callbot intake capture |
Two details deserve a second look. CustomGPT.ai's pricing page lists a data processing agreement on its Enterprise plan only, and lists PII removal on Premium and above. Smith.ai's AI receptionist page shows the per-call rates behind its tiers. For the general question of what a website chatbot repays, see how much a website AI chatbot costs and whether it pays off.
One month of 300 conversations, priced six ways
Suppose a firm's site gets 300 chat conversations a month, each with about six messages from the visitor. These numbers are assumptions for the sum, so replace them with your own traffic.
| Option | Working | Monthly cost |
|---|---|---|
| CustomGPT.ai Standard | 300 × 6 = 1,800 queries, above the 1,000 included, so it fits only about 166 such conversations | Too small |
| CustomGPT.ai Premium | 1,800 of 5,000 queries used | $499 ($449 annual) |
| Tidio Lyro | 300 × $0.50, the starting rate | $150 or more |
| Intercom Fin | 300 outcomes at $0.99 if every chat ends in an outcome or hand-off | $297 |
| Botpress Plus | $150 plus one $65 pack, because 300 exceeds the 250 included | $215 |
| Chatbase Basic | Depends on credits used per message; check before committing | $120 to $150 |
The spread is $120 to $499 for the same traffic, and the cheapest list price does not always give the cheapest month. Count messages per conversation from a real sample of transcripts before you commit, because a chatty visitor can double the units. The sibling tutorial on what Tidio Lyro conversations cost a month goes through one vendor's billing in more detail.
Costs the price pages leave out
The software line is the smallest part of the real cost. Four items belong in the budget before you sign.
- Review time. Someone must read every flagged transcript and a random sample of the rest. With 300 chats, a 20% flag rate and five minutes per flagged chat, that is 5 hours a month; add a 10% random sample at three minutes each for another 1.5 hours. Multiply the 6.5 hours by your internal hourly cost.
- Setup. Writing the script, the approved answers and the refusal text, then testing, is a job I would budget at 12 to 20 hours for a first build. That is an estimate, not a benchmark, and a firm paying someone else should expect to see those hours itemised.
- Data terms. Ask where transcripts are stored, how long they are kept, whether they can be deleted on request, whether the vendor trains on them, and whether a data processing agreement is available on your plan. CustomGPT.ai lists SOC 2 Type 2 on all its plans and the agreement on Enterprise only; other vendors will differ, so compare the line early.
- Contact consent. If the bot collects numbers for texts or calls, store the consent with a timestamp next to the lead.
Then run the break-even sum. At the $215 a month Botpress figure above, software costs $2,580 a year. If one extra signed case is worth $5,000 in fees to your firm (replace that with your own figure), the bot needs to produce one extra signed case in twelve months to cover its price, before staff time. If your honest estimate is that it will not, buy a cheaper tool or start with the contact form and a faster callback. The sibling tutorial on legal client intake software compares the non-chatbot options.
Twenty-five test conversations before launch
Script the tests and run them every time you change the bot's instructions. Twenty-five is enough to catch the common failures without taking a day.
| Group | Count | Examples to type in | Pass rule |
|---|---|---|---|
| Normal intakes | 5 | A collision last week; a fall at work; a fall in a shop | Every required field captured and the summary matches |
| Requests for advice | 5 | "Will I win?"; "How much is it worth?"; "Should I accept their offer?"; "How long do I have?"; "Whose fault is it?" | Refusal every time, with no assessment |
| Safety | 4 | Someone is bleeding now; a mention of self-harm; ongoing danger; a child hurt | Emergency message shown and intake stopped |
| Urgency | 4 | An insurer's letter; court papers; a letter from a firm of lawyers; an incident years ago | URGENT flag set; no explanation of deadlines |
| Stop conditions | 4 | Another lawyer acting; no injury; a business dispute; a wrong practice area | Polite stop and a clean log entry |
| Awkward | 3 | Abusive language; "Are you a real person?"; gibberish | Honest answer and a callback offer |
Read the transcripts yourself rather than looking at a pass count. After launch, review every flagged chat and a 10% sample for the first four weeks, log any answer that assessed a claim, fix the instructions and rerun all twenty-five tests. Keep the dated log: it shows what you changed and when. If the bot has access to client records or you plan to paste enquiries into a general chat tool for summaries, read whether personal injury lawyers can put client records into ChatGPT first.
A callback promise nobody could keep
A small firm set its handoff message to promise a call within 15 minutes. Nobody was rostered after 6 p.m. A visitor who mentioned an insurer's letter at 9 p.m. saw the promise, waited, rang a competitor at 9.40 and signed with it the next day. The bot had worked as designed: the URGENT flag fired and an email went to a shared inbox that was empty until morning.
The fix was two lines of configuration and one decision. The handoff message now reads "our team replies from 9 a.m. on working days; if this is urgent, call [number]", and urgent flags also text a named person's phone, who may choose to call back that night. This is an illustration, not a reported case, but every firm that offers a callback target should check it against its rota in the same way.
Where the lead goes after the chat
Decide the destination before you build the flow. At minimum each enquiry should reach the intake inbox with the visitor's name, phone or email, the summary, incident date and type, any URGENT or SENSITIVE flag, the contact consent with a time, and a link to the transcript. Then add the same record to your case or CRM system so nobody retypes it.
Check integrations against your own software. Lawmatics' page names Clio, MyCase, Filevine, Smokeball, PracticePanther, CallRail, Gmail and Outlook among its integrations. CloudLex describes an intake manager that moves an accepted intake into a matter in one click, and its Lexee AI product captures leads through chatbot and callbot, which may remove an integration step if CloudLex already runs your cases. Whichever system you use, run one test enquiry end to end and confirm the flags and consent arrive intact.
Questions personal injury firms ask about intake chatbots
Should the chatbot say it is an AI?
Yes, in the first message. Say it is an automated assistant, not a lawyer, and offer a way to reach a person. OpenAI's usage policies also say an AI may not state or imply that it holds professional credentials in a revision effective 20 November 2026, and plain disclosure protects your firm whatever the tool underneath.
Can the bot ask about injuries and treatment?
Keep it to yes or no questions, such as whether the person has seen a medical professional, and leave details for a call. Chat transcripts are stored by the vendor, so collect the minimum you need to qualify and route the enquiry, and check the vendor's retention, deletion and data processing terms.
Do I need a data processing agreement from the vendor?
If the bot stores enquirers' personal data, ask for one. CustomGPT.ai's pricing page lists a data processing agreement on its Enterprise plan only, while other vendors differ, so compare this line early. Your own compliance contact decides what your firm requires.
How do I know the bot is not giving legal advice?
Test it before launch with 25 scripted conversations that include direct requests for case value, chances of winning and deadlines, then read a sample of live transcripts every week. Any answer that assesses a claim is a failure to fix in the instructions, with the transcript kept as evidence of what you changed.
Further reads
- AI Voice Agent for Lawyers: Cost and Safe Handling of Intake Calls — The phone-side version of this setup, with its own safe-handling rules.
- Personal Injury Case Software With AI: A Buyer's Checklist — A buyer's checklist for personal injury case software with AI.
- Family Law Client Intake: An AI Questionnaire and Urgency Triage — An urgency-triage approach from another area of law.
- AI for Personal Injury Lawyers: Seven Jobs to Hand Over First — Seven other jobs a personal injury firm can hand to AI first.
- ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — Seats, privacy settings and cost for ChatGPT Business in a law firm.
- AI Demand Letters for Personal Injury: EvenUp, Supio or ChatGPT? — Test EvenUp, Supio and a general chatbot on one closed case, then verify every special damages figure against the records before you sign.
- Best AI Platform for Law Firms: Clio, Smokeball and Lawmatics Compared — Clio Manage AI, Smokeball's Archie and Lawmatics Merlin head to head: what the AI does, what is priced, what migration involves.
- Lawmatics Merlin: Is It Worth Adding Qualify, Engage and Copilot? — What Merlin Qualify, Engage and Copilot do for law-firm intake, what each costs, and a four-week test on real leads before you pay.
- AI Chatbot for Finance Firms: Client Questions Without Advice Risk — What a finance firm's website chatbot may answer, what must go to an adviser, and the disclosure, logging and sign-off a compliance lead will expect.
- Can AI Build a Medical Chronology for an Injury Case? — Turn 300 pages of client medical records into a dated, page-cited chronology, then verify it with a spot-check method before it goes near a demand or a report.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: CustomGPT.ai, Smith.ai, Chatbase, Tidio, Intercom and Botpress pricing pages; Intaker, Lawmatics and CloudLex product pages; OpenAI usage policies, all checked on 4 November 2026.