Key takeaways
- The SRA holds the firm accountable for what a chatbot says, including errors in advice from a third-party tool[1].
- Keep website chat to firm information, triage and enquiry capture, and refuse fact-specific advice with fixed wording.
- Collect only what a conflict check and a call-back need, and say how the details will be used before asking.
- Give every conversation a visible route to a person: lack of human oversight is a main reason people distrust AI in legal services[2].
- Test against a script of awkward questions before launch and again after every change.
What website chat should and should not do
A chat window on a solicitors' website does three useful jobs. It answers questions about the firm, it works out what kind of help the visitor needs and how urgent it is, and it captures enough detail for someone to call back. The SRA has recognised the appeal: chatbots can respond quickly to enquiries from potential clients, potentially offer a 24/7 service and gather information to triage cases for urgency[1].
The risk sits in a fourth job visitors often expect: answering their legal question. Once a chatbot applies the law to someone's facts, it is giving advice in the firm's name, and the SRA is clear that firms remain responsible for AI outputs, including errors in advice given by a third-party chatbot[1]. Guardrails are the written rules that keep the tool in its lane, and they work best as a short scope table that every fee earner can read.
| Topic | Chat can | Chat must not |
|---|---|---|
| Firm information | Give office hours, locations, practice areas, the SRA number and how to complain | Guess whether the firm takes work it does not list |
| Process | Explain in general terms what an initial consultation involves and what to bring | Say how long the visitor's own matter will take |
| Costs | Point to published price information and say fees are confirmed after a conversation | Quote a fee for the visitor's matter or promise a fixed price |
| Merits | Say that a solicitor will assess the matter | Say whether the visitor has a case, will win or should settle |
| Deadlines | Flag that time limits can apply and move the enquiry up the queue | Calculate a limitation or tribunal deadline for the visitor |
| Urgent risk | Give 999 and emergency routes, then escalate | Keep chatting through a safety issue |
Put the boundary in the opening message
Visitors read the first message; few open a disclaimer link. State the boundary in plain words at the top and repeat it whenever the conversation drifts. Consumers already have doubts: in the Legal Services Board's Legal Needs Survey, lack of trust in AI technology (51%) and lack of human oversight (49%) were the main barriers to using AI to access legal services[2].
Example opening message: 'Hello, I am the firm's virtual assistant. I can tell you about our services, help you book a consultation and pass your details to the right team. I cannot give legal advice about your situation; one of our solicitors will do that. If you are in danger, call 999.'
Keep three fixed refusal lines and use them word for word, so that testing gives predictable results and nobody has to judge improvised wording:
- Advice request: 'I can't advise on your situation, but I can get you booked in with a solicitor who can. Would you like that?'
- Prediction request: 'Only a solicitor who has looked at the details can say how a matter is likely to go. I can arrange that conversation.'
- Fee for their matter: 'Our published prices are on the fees page. The cost for your matter is confirmed once a solicitor understands what is involved.'
Capture the enquiry without over-collecting
The chat should collect what the firm needs to act and nothing more. The firm is the controller of what visitors type, and the Legal Ombudsman warns that not all free AI tools are secure or keep what users type confidential[3]. That is a reason to use a tool under a proper processor contract, and a reason to tell visitors not to paste documents or full case histories into the window.
- Name and preferred contact method, with a call-back window.
- Practice area, chosen from your list rather than typed freely.
- The other party's name for the conflict check, with a note that it is used only for that purpose.
- A one-line summary of the help needed, with a character limit.
- Any date that worries them, such as a hearing, completion or dismissal date.
- Consent to be contacted about this enquiry, kept separate from any marketing opt-in.
Show a short privacy line before the first question: who will see the details, why, and a link to the privacy notice. Do not ask for health details, criminal records or other special category data in the chat; the solicitor can take those on the call, with proper context and a file note.
When chat must stop and a person takes over
Every conversation needs a visible route to a human, not just a form at the end. People expect one: 89% of legal service users expect an AI legal tool to have a complaints process, and 66% expect that process to involve a real human[4].
Set hard triggers that end the automated part of the conversation and hand it to a named person or rota:
- Words that suggest danger, abuse, self-harm or arrest: give emergency routes first, then alert the on-call contact.
- A deadline inside a short window you set, such as a hearing next week or a completion date.
- A complaint about the firm or a named member of staff.
- A visitor who repeats an advice question after two refusals, or says they are confused or upset.
- Anyone acting for someone else who may lack capacity.
When a trigger fires, the chat should say what happens next and when, and the firm must then do it. A promised call within the hour that never comes does more damage than having no chat at all. For the rules on spoken hand-offs, see when an AI assistant should hand a call to a person.
Automated decisions and transparency
If the chat decides who gets a consultation, for example by turning away enquiries outside your practice areas, look hard at the decision rules. Under the Data (Use and Access) Act 2025, significant solely automated decisions need safeguards, including letting people challenge the decision and obtain human intervention[5]. The simplest safeguard is not to decide automatically: route out-of-scope enquiries to a person who signposts them, or to a referral page a solicitor has reviewed.
Keep regulatory transparency in view as well. The SRA requires firms to display their SRA number and digital badge prominently on the website and to publish their complaints procedure there[6]. The chat should be able to point to both whenever a visitor asks who regulates the firm or how to complain.
Test before launch and after every change
Write a test script
Collect real questions from reception notes and emails, anonymised, and add awkward ones: 'Will I win?', 'Can my landlord do this?', 'What will it cost?', 'I have a hearing tomorrow'.
Set pass rules
For each question, write the expected behaviour: answer, refuse with the fixed line, or escalate. A pass is the right behaviour, not a pleasant-sounding reply.
Run it with a solicitor watching
A fee earner from each practice area reviews the transcripts and marks any reply that strays into advice or states the law wrongly.
Check the data trail
Confirm each test enquiry arrived with the right fields, consent record and practice-area tag, and that nothing landed in an unmonitored inbox.
Re-test on every change
Any new wording, setting, model version or practice area triggers the full script again. Keep the results with the date and the reviewer's name.
Sample live conversations
Read a sample of real transcripts every week in the first month, then monthly, and log what you changed and why.
A worked example
Example: a visitor writes at 10pm: 'My employer dismissed me last Friday. Can they do that without a warning? I want to sue.' A guarded chat replies that it cannot advise on whether the dismissal was lawful, notes that strict time limits apply to employment claims and that a solicitor will want to speak soon, then offers the next consultation slots or a call-back request. It collects the employer's name for the conflict check and the dismissal date, tags the enquiry as employment and time-sensitive, and the duty fee earner sees it first thing in the morning.
What it does not do is explain the law on unfair dismissal, estimate compensation or tell the visitor they have a strong claim. For the intake questions behind this flow, see designing an AI intake assistant, and for where the line between general information and advice falls, read information vs advice.
Frequently asked questions
Can a chatbot on a solicitors' website give legal advice?
Technically it can produce advice-like answers, but the firm is responsible for whatever it says, so most firms limit it to firm information, triage and booking. Advice on a visitor's own facts should come from a solicitor who has checked conflicts and taken instructions.
Is a disclaimer link enough?
No. Say in the opening message what the chat can and cannot do, repeat it when a visitor asks for advice, and show a short privacy line with a link to your privacy notice before collecting any details.
What details should a law firm chatbot collect?
Only what you need to call back and run a conflict check: name, contact details, practice area, the other party's name, a short summary and any urgent date. Leave sensitive details for the call.
How often should we test website chat?
Before launch, after any change to its wording, settings or underlying model, and by sampling live conversations weekly in the first month and monthly after that.
References
Solicitors Regulation Authority (2023). Risk Outlook report: The use of artificial intelligence in the legal market (opens in a new tab) ↑
Legal Services Board (2024). Legal Needs Survey 2023: Summary report (opens in a new tab) ↑
Legal Ombudsman (2026). AI and Complaints: A Consumer Guide (opens in a new tab) ↑
Legal Services Board (2026). AI in Legal Services: Research Report (LSB Public Panel) (opens in a new tab) ↑
GOV.UK (DSIT) (2025). Data (Use and Access) Act 2025: data protection and privacy changes (opens in a new tab) ↑
Solicitors Regulation Authority (2025). SRA Transparency Rules (opens in a new tab) ↑