A 6-partner UK law firm tracked their paralegal diary in January. Of the 47 minutes spent on phone calls before 10am each morning, 38 were first enquiries: prospective clients describing their situation, answering the same 14 questions — name, contact details, matter type, urgency, opposing party, previous adviser, legal aid eligibility — and being told a solicitor would call back within two working days. The paralegal had the script memorised. Every call followed it without exception.
A script that fixed, executed that consistently, is automatable. A well-designed intake agent covers the same ground in under four minutes with a higher data-capture rate than a paralegal juggling notes and call management simultaneously.
What UK law firm client intake calls actually cover: the 14 standard questions that follow a deterministic script
The exact script varies slightly by practice area, but the core structure is stable across the generalist SME firms we've worked with. Here's what a standard first-enquiry call captures:
| # | Question | Primary purpose |
|---|---|---|
| 1 | Full name | Identity; conflict-of-interest check |
| 2 | Contact phone number | Callback |
| 3 | Contact email address | Confirmation; future correspondence |
| 4 | How they heard about the firm | Marketing attribution |
| 5 | Matter type (property, employment, family, commercial) | Department routing |
| 6 | Brief description of situation | Initial triage |
| 7 | Is there an opposing party? | Conflict check trigger |
| 8 | Name of opposing party (if yes) | Conflict check against client database |
| 9 | Have they instructed another solicitor for this matter? | Prior advice liability |
| 10 | Which solicitor or firm (if applicable)? | Potential referral conflict |
| 11 | Is the matter urgent? (court date, deadline) | Priority queue routing |
| 12 | What is the approximate deadline or court date? | Scheduling |
| 13 | Legal aid potentially eligible? (benefits, income threshold check) | Fee-earner capacity planning |
| 14 | Preferred callback time | Scheduling |
Questions 1, 7, and 8 drive the conflict-of-interest check against the firm's existing client database. A missed opposing-party name creates a professional conduct problem: the solicitor takes the matter and only later discovers they already act for the other side. The agent collects all 14 fields, pushes a structured JSON record to the case management system, and flags it for paralegal review. It does not open the matter file.
SRA compliance for automated intake: what AI voice agents can and cannot do in an SRA-regulated firm
The SRA's Transparency Rules and Standards and Regulations apply regardless of who — or what — conducts the intake call. The firm remains the regulated entity. The SRA's guidance on technology and innovation in legal services is unambiguous: automation does not reduce the firm's responsibility for what gets captured or communicated.
The agent can: collect information following a defined script; read out the privacy notice and recording consent; classify matter type and route to the correct queue; flag conflict-of-interest risks for human review.
The agent cannot: give legal opinions or assess claim strength; open a matter file without solicitor review; provide fee estimates; represent itself as a human.
One compliance step firms consistently skip: the agent must identify itself as automated in its first sentence — not buried in terms disclosure. "Hi, this is an automated intake system for [Firm Name]. I'll ask a few questions before arranging a solicitor callback." Failing to disclose this is a Principle 7 breach and damages completion rates: callers who discover mid-conversation they're speaking to a machine are more likely to abandon the call.
Call routing by matter type: property, employment, family, and commercial — the branches that determine which solicitor sees the summary
After capturing matter type at question 5, the agent routes the intake record to the correct fee-earner queue. The routing logic:
Caller describes situation
|
Matter type classified?
/ | | \
Property Family Employ. Commercial
| | | |
Conveyancing Divorce ET claim Contract
Remortgage Children Redundancy Dispute
Dispute Probate TUPE Debt
|
Urgent flag set?
/ \
Yes No
| |
Duty solicitor Standard queue
(same-day) (2-day callback)
The LLM classifies matter type from the caller's free-text description at question 6 — it does not ask callers to select from a menu. "My landlord is trying to evict me" maps to property/dispute. "I think I was unfairly dismissed" maps to employment. Classification runs as a structured output step; if confidence falls below 0.72, the agent asks one clarifying question, then routes to general enquiries if still unclear rather than looping.
LLM prompt design for legal intake: capturing conflict-of-interest flags and urgency without hallucinating advice
The most common failure mode in legal intake agents is the LLM helpfully extrapolating legal opinions from the caller's description. A caller says "my employer changed my contract without telling me" — a poorly constrained model responds "that could constitute constructive dismissal." That's legal advice. It's prohibited. It's the firm's professional liability.
The system prompt structure that prevents this:
{
"role": "system",
"content": "You are an intake agent for [Firm Name], a UK solicitors' firm. Your sole function is to collect information. You do not assess legal merits, predict outcomes, advise on rights, or comment on the strength or weakness of any claim. If a caller asks for legal opinion, advice, or your view on their situation, say exactly: 'I can note that down, but I am not able to give legal advice — a solicitor will call you to discuss your situation.' Collect the following 14 fields in order: [field list]. At call end, output a JSON record with all fields plus conflict_flag (true if opposing_party matches any name in context block CONFLICT_LIST) and urgent_flag (true if caller mentions a court date, deadline, or imminent legal action within 48 hours). Do not interpret or summarise the caller's situation beyond what they state. Do not speculate on their legal position."
}
The conflict-of-interest list must be injected at call-time, not baked into the system prompt — it changes as matters open and close. For Clio, the Clio Manage API exposes an /matters endpoint returning active client and opposing-party names with appropriate auth scopes. Pull it at call-start and inject as a CONFLICT_LIST context block.
The urgency flag needs few-shot examples, not just the rule. Without them, "I need to get this sorted soon" triggers urgent: true alongside genuine court-date emergencies. See our notes on prompt engineering for voice agent latency and accuracy for the few-shot structure.
Handoff to case management systems: writing structured intake data to Clio, Leap, or a Google Sheet automatically
At call end, the agent produces a structured JSON record and pushes it via webhook. The record we've standardised on:
{
"intake_timestamp": "2026-09-18T09:14:32Z",
"source": "voice_agent",
"caller_name": "Sarah Thompson",
"contact_phone": "07700 900123",
"contact_email": "[email protected]",
"matter_type": "employment",
"matter_sub_type": "unfair_dismissal",
"situation_summary": "Dismissed 3 days ago after 4 years service, no written reason given",
"opposing_party": "Meridian Logistics Ltd",
"conflict_flag": false,
"previous_solicitor": null,
"urgent": false,
"urgency_detail": null,
"legal_aid_eligible": "unknown",
"preferred_callback": "morning",
"consent_recording": true,
"call_duration_seconds": 214,
"routing_queue": "employment_standard",
"review_status": "pending_paralegal"
}
For Clio, this record creates a new Matter in "Intake" status via the Manage API. For Leap, common in UK conveyancing practices, the REST API creates client and matter records simultaneously. For firms without case management software, a Google Sheet append via Apps Script is a short-term stopgap — not suitable for conflict checking, but functional.
The webhook fires at call end. If the push fails, the agent emails the record to the firm's intake@ address. We do not auto-retry — a duplicate matter record is harder to unpick than a delayed push. The voice AI and document analysis pipeline case study shows how we wired a similar structured-output pattern into a professional services document workflow.
Voicemail handling and after-hours intake: capturing matters that arrive outside reception hours
Prospective clients dealing with employment or family matters often call in the evening or at weekends, away from a workplace where contacting a solicitor might be noticed. A firm with standard reception hours misses those enquiries without an after-hours agent.
The configuration is the same intake script with an adjusted opening: "Our office is currently closed. You can leave your details now and a solicitor will call you back on the next working day." The record writes with received_after_hours: true and joins the morning review queue.
Voicemail fallback triggers when a call drops or the caller is silent for more than eight seconds. The agent transfers to voicemail; the recording is transcribed via Deepgram's Nova-3 model and appended to the intake queue with source: voicemail. Lower data quality than a completed conversation, but better than a missed call with no record. For a broader look at intent-driven voice agents replacing IVR menus, see replacing IVR systems with AI voice agents for UK SMEs.
Transfer-to-human triggers: the six situation types where the agent must hand off immediately
The agent transfers to a live person in six situations, without delay or second attempt:
- Caller is distressed or indicates risk of harm — mentions domestic abuse, self-harm, threat of violence, or immediate physical danger
- Urgent legal emergency requiring same-day action — court hearing today, bailiff on premises, injunction deadline within hours
- Caller explicitly asks to speak to a person — no re-prompt, immediate transfer
- Multiple failed matter classification attempts — if matter type remains unclassified after two clarifying questions, route to general enquiries rather than loop
- Language barrier — if the agent cannot conduct a coherent exchange in the caller's primary language, detect and transfer; do not improvise
- Safeguarding concern involving a minor — any mention of a child's welfare or safety triggers an immediate duty-solicitor transfer regardless of matter type
The transfer destination is configured per firm: duty solicitor direct line during hours, emergency voicemail out of hours. The partial intake record gets flagged with transfer_reason and pushed to the receiving solicitor before the call connects. The transfer-to-human handoff mechanics post covers SIP REFER vs. call bridge options and the context-injection pattern.
What changed in 2025–2026: SRA guidance on AI in client-facing processes and Legal AI regulation updates
The SRA published detailed guidance on technology and innovation in legal services in June 2025 — the first time the regulator explicitly addressed AI in client-facing workflows rather than back-office document drafting. The requirements: a human review step must occur before any substantive legal action follows automated intake; AI tools used in client communications must be disclosed in the client care letter; and firms must conduct a proportionate risk assessment before deploying automated intake systems.
The Legal Services Board's February 2026 consultation proposed that AI-generated intake summaries be labelled as such in documents provided to clients. Final rules are expected before year-end — firms deploying now should build labelling into matter templates as a precaution. The ICO's guidance on AI and data protection (updated March 2024) recommends retaining full call transcripts only during an active matter, then moving to pseudonymised summary. The consent capture mechanics — opening disclosure script, explicit verbal consent, and UK data residency obligations — are covered in voice agent call recording consent for UK SMEs.
The counterpoint: the Law Society's 2025 AI practice note recommends that firms seek Counsel's opinion before deploying any AI in client-facing roles — a more cautious position than the SRA's guidance. For smaller practices uncertain about their professional indemnity insurer's position on AI, that's a reasonable first step.
Good / Bad / Ugly: three law firm intake designs and the compliance risk each one carried
Good — the compliant triage agent
A 4-partner conveyancing firm: agent identifies itself as automated in the first sentence, reads the privacy notice and recording consent, writes a structured record to Clio, and flags every record for paralegal review before the fee-earner sees it. No legal comment on the caller's situation. Conflict checks run against a live Clio list before any substantive questions. Transfer-to-human fires on any of the six trigger conditions. Intake capture rate moved from 71% of inbound calls to 94% over 90 days. No SRA compliance issues. The same intake-and-routing model scales to other regulated professional services — voice agents for UK accountancy firms shows how a parallel triage structure operates under different compliance constraints.
Bad — the agent without disclosure
A 3-partner employment firm deployed an agent that did not disclose it was automated. Callers were told they'd reached "the firm's client services team." The Principle 7 exposure is straightforward: representing an automated system as a human team is a breach. The agent also retained call transcripts in the voice platform's default cloud storage without a UK data residency guarantee — a UK GDPR Article 46 issue for special-category data. The firm suspended the deployment after six weeks when a client queried a discrepancy between their intake notes and what they'd told "the person who answered."
Ugly — the agent that gave advice
A commercial law firm deployed an off-the-shelf legal chatbot as a voice intake agent without system-prompt restrictions on legal opinion. The model assessed claim strength and offered indicative timescales for contract disputes. Two callers received statements their solicitors later described as legal advice given on the firm's behalf. The firm's professional indemnity insurer flagged this at renewal review. The model cannot recognise it is doing something prohibited — that constraint comes entirely from prompt design and post-call audit. No prompt restriction, no safety.