AI triage for legal helplines: what works
By the ChamberOne team 5 minute read
A helpline lives or dies on the first minute. Someone rings, frightened or confused, and the service has to decide: is this urgent, who can help, and what should happen next? Most callers have common problems. A few have emergencies that cannot wait until Thursday. Technology can sort the first group faster, but only if the design protects the second.
Key points
- Base triage on a script your own experts have written, and use AI to carry it out, not to invent it.
- Write an urgent-case list, over-flag, and route any match straight to a person.
- Use AI for transcription, summaries and suggested categories, with a human confirming each.
- Always offer a clear route to speak to a person, and keep other channels open.
What triage means for a helpline
Triage is a short, structured conversation that places an enquiry in the right category. It is not advice. The person taking the call, or the form the person fills in, gathers a few facts and decides where the enquiry goes: to an adviser now, to a callback, to a self-help resource or to another service.
Done well, triage shortens waiting times, makes sure urgent matters reach a lawyer first and stops people repeating their story three times. Done badly, it puts barriers in front of the people who most need help.
Start with a script, not a model
The most reliable triage systems are built on a clear script that the service's own experts have written and tested. The script asks the same questions in the same order: what kind of problem is it, is anyone in immediate danger, are there time limits close, is a court or tribunal date set, does the person have somewhere safe to stay?
Software, including AI, is best used to carry out the script and to help with the messy parts around it, rather than to invent the questions. A script can be reviewed, audited and corrected. A free-form model's behaviour is harder to predict.
Urgent-case flags
Every helpline has a short list of situations that need a person immediately. The exact list depends on the service and the place, and your own experts should write it. Typical categories include risk of violence, loss of a home within days, a court or tribunal deadline that is about to pass, detention or arrest and risk to a child.
- Write the urgent list down and have senior advisers approve it.
- Build it so that any match produces an immediate route to a human, not a queue.
- Err towards over-flagging. A false alarm costs a few minutes. A missed emergency can cost much more.
- Test with realistic examples, including badly worded ones, since people in distress rarely use neat phrases.
- Review flagged and missed cases every month and update the list.
Where AI helps
AI can do useful work around the script. It can transcribe a call so that the adviser does not have to type. It can summarise a long free-text enquiry into a few lines for the case worker. It can suggest which category a message belongs to, so that a human can confirm quickly. It can translate between languages, which matters for services that deal with people who speak many languages.
In each case, a person looks at the result. A summary can leave out the one sentence that mattered. A translation can distort a threat or a date. A suggested category can be wrong. The role of the technology is to speed up the human reader, not to replace them.
Where AI should not decide
- Whether someone is eligible for help, where the decision affects access to a service and is not purely mechanical.
- Whether an enquiry is urgent, when the system is unsure. Uncertain cases go to a person.
- What the legal merits of a person's problem are. That is advice, and it needs a qualified adviser.
- Anything involving a vulnerable person, a child, or a person who says they are in danger.
- Transcribe a call
- Summarise a long enquiry in a few lines
- Suggest a category for a human to confirm
- Translate between languages
- Eligibility for help
- Whether an uncertain case is urgent
- The legal merits of the problem
- Anything involving a vulnerable person, a child or someone in danger
A human for the edge cases
Every automated route needs an obvious way out. Put a plain option on every screen and in every phone menu to speak to a person. Staff the option properly. A button that leads to a voicemail box nobody checks is worse than no button.
Train staff to override the system. An adviser who senses something is wrong must be able to escalate, without having to justify it to a form. Keep a record of overrides, since they show where the system falls short.
Accessibility and language
Callers include people with low literacy, limited English, disabilities, poor connectivity or no smartphone. A web form is not an alternative to a phone line for them. Keep several channels open: phone, form, in person where possible, and assisted completion for those who need it. Use plain words in scripts, and test them with the people the service is meant to serve.
If AI is used for translation or transcription, check its accuracy in the languages and dialects your callers use, and have fluent speakers review a sample regularly. Performance in widely used languages can be much better than in smaller ones, and the gap is easy to miss if nobody checks.
Privacy and consent
Helpline enquiries often contain highly sensitive information about families, immigration status, health and criminal matters. Before any enquiry text or call audio goes to a third-party tool, check where it is processed, whether it is kept or used for training, and what the service's own data protection and funder obligations require. Tell callers, in simple words, what is recorded and why, and what choices they have. Check the guidance from your regulator or funder if you are unsure.
How to pilot it
Start small. Run the new triage alongside the existing process for a few weeks, and compare outcomes. Check that urgent cases reached an adviser as fast as before or faster. Ask staff and callers what felt wrong. Only widen the use when the evidence is good, and keep checking afterwards, because enquiries change over time and so do the people who make them.
Handling enquiries in ChamberOne
- Each firm has a public enquiry form with its own link.
- Enquiries arrive in an inbox with a name-based conflict warning, and someone at the firm converts one to a client, a lead or a matter.
What it does not do
- Nothing from the enquiry form becomes a client until someone at the firm converts it.
- A conflict warning is based on names. The decision to act stays with the lawyer, and it does not give legal advice to the person enquiring.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
AI and access to justice: an honest assessment
Where AI can lower the cost of routine legal help, where it cannot replace advice, and who bears the risk when it is wrong.
5 minute readClient portals: what self-service does for clients and the firm
Which questions a portal answers, what to keep out of it, and how it reduces routine calls.
5 minute readAI chatbots on law firm websites: benefits, risks and rules
A chatbot can help with intake, but it can also say too much. Decide its limits before it goes live.
4 minute read