Lorikeet
for Best Practice Accounting Group
Contact us
Fielding the central support inbox and the 1300 line across the 5-firm national group. Wired into Nagaris so it can look up a client's tax return status, reschedule an appointment across accountant availability, and route complex work straight to the right accountant without attempting to answer it. Every conversation reviewed by Coach at 100% coverage.
Loaded with a mock Nagaris — 4 clients (Sarah/ABC Plumbing, Marcus/Chen Consulting, Olivia/Reed Family Trust+SMSF, James Patel individual), 5 accountants (Nick Adelaide, Karla, Belinda, Tim, Lucy), realistic appointments, tax return states, and outstanding invoices. Any email routes to the canonical demo client (Sarah). Open chat and try a hero flow.
CTO-relevant capabilities, not marketing surface. Every claim below is testable in the chat widget in the next 60 seconds.
One conversation, 4-5 Nagaris tools: verify → look up client → check accountant availability → reschedule with a real write → confirm. Try the "Reschedule with Nick" pill. Every tool call, LLM prompt, guardrail check, and decision is visible in the timeline UI.
Ask "should I set up an SMSF to buy the workshop?" The agent doesn't guess. It acknowledges, captures the exact question in the client's own words, and files a Nagaris case for the right accountant with SLA. CTOs judge AI on refusals, not resolutions.
Every reply has an inspectable timeline: what workflow matched, what tools ran with what args, what the LLM saw, what got steered, what got closed. Debug any conversation the same way you'd debug any other production request — with a stack trace, not a black box.
The 1300 phone line, the support@bestpractice.com.au inbox, and the chat widget all run the same workflows against the same Nagaris mocks. Voice renders natural cadence and full sentences; chat renders bullets and links. No second build. Same brain.
Every ticket, every call, reviewed against policy at 100% coverage. Patterns surface — which categories are resolving cleanly, where the agent's drifting, where the KB is thin. No human QA queue. The reactive→proactive quality loop for a national group processing thousands of conversations a month.
Every workflow is a markdown document — Nick or a Nagaris engineer can open, edit, save, test in under 15 minutes. No YAML, no visual builder maze, no engineering ticket to change a routing rule. Ask us to walk you through the appointmentReschedule source live.
Most "AI support" demos stop at retrieval — search the KB, return the article. BPAG's needs go further: a client saying "reschedule my Thursday with Nick" should look up the appointment, check Nick's availability, offer real slots, write the reschedule, send the calendar invite — end-to-end, in one conversation.
The Reschedule pill above does exactly that against a mocked Nagaris. Open the timeline after the reply and you'll see: verifyClient, getClientProfile, getUpcomingAppointments, getAccountantAvailability, rescheduleAppointment. Five tool calls, one turn, one write.
Built off the 22 June intro call — Robert's actual asks + Simon's technical concerns, wired for Nagaris.
All 54 Q&A articles Robert authored for the Proactive sandbox — BAS deadlines, PAYG, SMSF, Xero, ATO scams, deceased estates, Zero Penalties Guarantee — rebadged for BPAG and pinned as INTEGRATION-type reference material so they can be resynced when Robert edits them.
Eight tools shaped to Nagaris's likely API surface: verifyClient, getClientProfile, getTaxReturns, getUpcomingAppointments, getAccountantAvailability, rescheduleAppointment, getOutstandingInvoices, routeToAccountant. Rich seeded data: 4 clients with different entity structures, 5 accountants incl. Nick Adelaide, realistic tax-return states.
Multi-tool appointment reschedule (5 Nagaris calls, one turn) — and the judgment escalation (agent captures a complex question in the client's own words and files a Nagaris case for the right accountant, refusing to answer the substance). Both live now.
Three STEER guardrails active: no personalised financial or tax advice, customer demands human twice, hostility. All STEER (no auto-close, no black-box escalation). Twelve brand guidelines including SCAN-FIRST, voice contact-detail capture, and complete-flow-with-wait — the same fleet-wide canon that runs on regulated demos.
The honest comparison: what your Nagaris engineers would rebuild if you stitched OpenAI + LangChain + vector DB in-house.
The 8 mocked tools are the shape your real Nagaris endpoints would take. Swap the mock for a real HTTP call and the same workflow still runs — no re-authoring. Compared to building an agent SDK yourself: 3-4 weeks of scaffolding and prompt patterns you don't need to write.
Same workflows, three channels. Voice rendering vs chat rendering vs email rendering is Lorikeet's problem, not yours. Building voice properly (natural cadence, mid-utterance interrupts, TTS-aware phrasing) is its own multi-month project.
Auto-QA at 100% coverage, replacing what a homegrown build usually punts on for 18 months. Not just "did it answer?" — did it hit the right escalation, use the right tool args, follow the fee-disclosure policy. Otherwise this is engineering time you'd never get back.
"No personalised financial advice" is a codified guardrail with a specific prompt, examples, and steering — not a prompt substring lost in a system message. Trivially auditable. Trivially testable. That's the discipline a homegrown agent takes months to build.
Every conversation is an inspectable trace: workflow match, tool calls, LLM prompts, guardrail evaluations, decisions. Not a black box. Debugging in production is the same shape as debugging any request — not a bespoke tracing system your team has to build.
Lorikeet is built on this category long-term, with regulated-vertical customers (super, banking, healthcare) already in production. The Nagaris team stays focused on the product, not on rebuilding LLM plumbing every 12 months as base models turn over.
Four steps from this sandbox to a national deflection layer on the BPAG central inbox and 1300 line.
Robert, Nick, and the three Nagaris engineers get direct MCP + web access. Poke at workflows, add new ones, break things, see the timeline. Roughly 1-2 weeks of unstructured exploration — no cost.
Swap the 8 mocks for real Nagaris HTTP endpoints. Zero write actions in prod yet — pure deflection on current inbound. Measure resolution rate on FAQ + client-status lookups against the current baseline.
Enable each Nagaris write with confidence gates: routeToAccountant first, then rescheduleAppointment, then the more sensitive ones. Each action earns its unlock via measured accuracy on the Coach loop.
Turn on the voice channel against the same workflows. Route the central support@bestpractice.com.au inbox through the email channel (needs a light ticketing layer — Intercom fine as an interim, discussed with Robert on the call).
Happy to jump on a screen-share and walk through the timeline UI, the workflow markdown source, and the Coach eval loop with you and the Nagaris engineers. Robert has the sandbox invite in his inbox.
Book the walkthrough