A real multi-agent system built for a family-law and mediation firm. It takes the routine admin — intake, conflict checks, drafting, scheduling, billing — while every safety call and legal judgment stays with the practitioner. Here's exactly how it works, in diagrams.
A coordinator agent reads every incoming message, runs a safety check first, then delegates to the specialist built for that job. Each specialist has a narrow scope — least-privilege by design.
Reads every message · runs the safety check before anything else · routes to the right specialist
SAFETY GATE RUNS HERE FIRSTClassifies the matter, judges urgency, opens the file
Conflict checks + family-violence screening; shuttle vs joint
First drafts, every one marked needs review
Books sessions, writes reminders, safety-aware formats
GST invoices, trust/office separation, arrears
Each specialist reports back to the coordinator, which synthesises the next step for the practitioner.
The safety check is plain, deterministic code, not an AI decision. It can't be talked around, it's logged, and it works even if the AI is switched off. This is the part most AI tools get wrong.
Any inbound enquiry
Threat to life, weapon, child in danger, self-harm AI blocked. Shows 000, 1800RESPECT, Lifeline, DCJ. Flags a person.
AVO, assault, coercive control, stalking flags for review, defaults to shuttle, won't join parties without sign-off.
No critical flags routed to the AI specialists. Elevated factors quietly noted and re-screened as the matter develops.
The safety net is written into the system's foundations, not its personality — testable, logged, always on.
For every task the system asks: can it be undone, and can the outcome be checked? The answer places it in one of four boxes — and only the green box runs without a human.
Agent acts on its own. e.g. drafting a reminder, classifying a matter.
Agent drafts; you review. e.g. an engagement letter or parenting plan.
Agent prepares; you approve before it acts. e.g. sending a client message.
Agent surfaces info; you decide. e.g. issuing a s 60I certificate, advising on settlement, a safety call.
Legal advice or predicting court outcomes · issuing a s 60I certificate or signing a document · the client relationship — trust, empathy, the hard conversation · any safety judgment about a real person. The system flags and escalates; the practitioner decides.
A high-risk enquiry arrives. Here's the path it takes before the practitioner has even opened their inbox.
Reads it, runs the safety screen — flags High risk immediately.
Classifies as a parenting matter, marks it urgent, opens the file.
Conflict check clears; recommends shuttle sessions, no joint meeting.
A briefed summary + recommended next steps land for your review.
Every step above is written to a timestamped audit log — nothing happens invisibly.
The qualities that make an agentic system safe to run inside a regulated profession.
Every screen, draft and invoice logged with a timestamp. Nothing invisible.
The AI key lives in a secure vault, never written into the code.
Client text is safety-screened on-system before anything is sent to the AI.
Safety + admin logic is plain code — it keeps working even if the AI is down.
Before it touches a real client file, three things take a practitioner's sign-off: the document templates, the trust-accounting and costs-disclosure flow (against the Uniform Law), and the privacy/storage setup. The system is built so each is a review step — not a rebuild.
Every judgment, every safety call, and every client relationship stays exactly where it belongs — with the practitioner. This is the pattern BIFF AI builds across every domain.