Turning ad-hoc use of Claude & ChatGPT into a governed, repeatable operating system for a 32-person Melbourne financial-planning firm, proven on two pilot workflows.
Claude ProjectsPrompt libraryAI governanceBefore/after model
A demonstration built around a representative client scenario · figures are modelled estimates under stated assumptions.
The problem
Ungoverned AI, risk without leverage
BrightPath grew 8 → 32 in four years. Staff already use AI daily, but:
Client emails and advice drafts are rewritten from scratch every time.
Every adviser uses different prompts, quality never compounds.
Some paste client data into tools without safeguards; others avoid AI entirely.
There's no baseline, the firm can't tell if AI is helping.
Maturity assessment: ~1.6 / 5, "Emerging." The fix isn't more tools; it's standards, configuration and governance on the tools they already own.
The approach
A four-layer operating system
1 · Governance Acceptable use · human-review gates · data tiers · red lines
2 · Workflow standards One agreed method per workflow, documented as SOPs
3 · Configured tools Claude Projects + a shared, versioned prompt library
4 · People Playbooks · AI Champions · training · measurement
Operating principle: AI drafts, a qualified human decides, and every improvement is captured once and reused by everyone.
Pilot 1 · Advice preparation
Current state → future state
Today, ~5.0 hrs / advice document, inconsistent:
Gather inputs
Find last SoA
hunt SharePoint/Dropbox
Rebuild boilerplate
from scratch
Write rationale
blank page
Review
slow, inconsistent
Issue
With the operating system, ~3.3 hrs, 34% faster (scenario):
Structured input
placeholders
Advice-Prep Project
firm-standard draft
Best-interests review
adviser verifies & signs
Issue
human owns advice
Pilot 2 · Client communications, AI in action
Triage & draft, with an advice-escalation guardrail
Inbound · client email
"Hi, quick one, my super's just sitting in the balanced option. With rates where they are, should I switch it all to cash for a while? Also can we move Thursday's meeting to Friday? Thanks, [client]"
TRIAGE: POSSIBLE ADVICE → ESCALATE + BOOKING
Holding reply (for staff to send):
"Thanks [CLIENT], happy to move the
meeting to [FRI DATE]. On your super
question, I'd like [ADVISER] to give
you a proper answer given your goals;
they'll call you [TIMEFRAME]."
⚑ ESCALATE TO ADVISER, reason: switch
super to cash = personal advice.
Booking handled; specifics as
placeholders. A human sends.
The advice question is never auto-answered, it's escalated with a warm holding reply, while the routine booking part is drafted. This guardrail is the point of the design.
Scenario results
Both pilots clear the ≥20% bar
34%
Advice drafting ↓
30%
Email handle-time ↓
~1,288
Hours reclaimed / yr
~7×
Value vs fee
The percentages depend only on per-item time, the most measurable input, so they're robust; the volume and dollar assumptions scale value, not the percentage. Every input lives in one editable spreadsheet, so the model becomes a real dashboard as actuals replace estimates.
Scenario estimates under the stated assumptions.
Governance & what's delivered
Built for regulated advice, and reusable
The gate is unavoidable: AI drafts, a licensed human decides. Client data stays out of tools via placeholders; only business-tier tools with training off are approved; a short signed framework maps to the best-interests duty, ASIC RG 175 and the Privacy Act / APPs.