QuantAI · Fractional AI Operations

Turning ad-hoc AI into a governed operating system

A 32-person Melbourne financial-planning firm was using Claude and ChatGPT inconsistently and without safeguards. This is the AI operating system that fixed it, proven on two pilot workflows.
Claude Projects & skills Prompt library AI governance Financial services Change & enablement
BrightPath Financial Services · Prepared by Dave Richardson, QuantAI

The situation

BrightPath grew from 8 to 32 people in four years. Staff already used AI daily, but every adviser had invented their own method, client data was going into tools without checks, and nothing was measured. The problem wasn't "no AI"; it was ungoverned AI: risk without leverage.

The approach, a four-layer operating system

Not software. The small set of standards, configured tools and rules that make good AI use the default. Built on the firm's existing stack (M365, Claude Team, ChatGPT Team), configuration, not new purchases.

1 · Governance
Acceptable use, mandatory human-review gates, data-handling tiers, red lines, fit for regulated advice.
2 · Workflow standards
One agreed method per workflow: where AI drafts, where a named human signs. Documented as SOPs.
3 · Configured tools
Claude Projects per workflow + a shared, versioned prompt library, so quality compounds instead of staying private.
4 · People
Role playbooks, an AI Champion per department, training, and a measurement baseline.
The operating principle: AI drafts, a qualified human decides, and every improvement is captured once and reused by everyone. Any new workflow that can't preserve human accountability and shared reuse doesn't go in.

Two pilots, built end-to-end

Pilot 1 · Advice preparation

A firm-standard Statement-of-Advice first draft from structured inputs, removing the blank-page and boilerplate cost while making the adviser's best-interests review stronger. The AI never issues advice; a licensed adviser verifies every figure and signs.

Pilot 2 · Client communications

Triage and drafting of routine client email in the firm's voice, with an explicit "possible advice → escalate" path so an advice question never gets a casual AI answer. A human sends every message.

Scenario results

Modelled under stated assumptions (see the before/after model). The brief's success bar is ≥20% cycle-time improvement, both pilots clear it.

34%
Advice drafting ↓
30%
Email handle-time ↓
~1,288
Hours reclaimed / yr
~7×
Value vs fee (scenario)

The ≥20% result depends only on per-item drafting/handle time, the most directly measurable input, so it's robust to the volume and dollar assumptions, which scale value, not the percentage. Dollar figures are an indicative value of reclaimed time, not cash savings.

Governance built for regulated advice

Because this is financial advice, the human-review gate is designed to be unavoidable, not optional. Client-identifying data is kept out of tools via a placeholder de-identification habit; only business-tier tools with training disabled are approved; and a short, signed acceptable-use framework maps to Australian obligations (best-interests duty, ASIC RG 175, Privacy Act 1988 / APPs) at a practitioner level. AI drafts · a licensed human decides

What was delivered