Coltbridge×AI 01 / 25
Swipe to navigate
Briefing

Coltbridge×AI
A practical guide.

Twenty-five slides. About 8 minutes. We'll cover what AI is, where your data goes, why a normal Claude subscription is fine for Coltbridge, and how to start safely — with a free policy template you can use today.

Use the arrows or swipe to advance · Press → to start

Ch 1 · What is AI

It's very smart autocomplete.

Modern AI tools — Claude, ChatGPT, Copilot — are sophisticated text generators that have read most of the internet. You give them text. They give you text back.

That's the entire trick. The magic is in how good they've become at it.

Ch 1 · What is AI

Three things to know

  • It runs in a datacentre, not on your laptop. Prompts travel to a building somewhere, get processed on specialised chips, and come back.
  • It doesn't remember you. Every conversation starts fresh by default. Memory features bolt on top, but the model itself is amnesiac.
  • It can be confidently wrong. "Hallucination" — fabricated facts, names, numbers. Always verify before relying on it for client work.
Chapter 2

Where does your
data go?

The conversation everyone in finance asks about.

Ch 2 · Where data goes

When you type something, it takes a journey

You Perth IN TRANSIT (encrypted) DATACENTRE GPU · INFERENCE your data is decrypted here STORAGE · AT REST history, logs IN TRANSIT (encrypted) Response back to you ↑ COUNTRY OF DATACENTRE = JURISDICTION OVER YOUR DATA ↑
// the three states data passes through
Ch 2 · Where data goes

What this means for Coltbridge

  • Your clients have expectations about confidentiality that don't easily fit "we sent your financials to a US AI service" — even if no rule strictly forbids it.
  • The MFAA AI guidance puts privacy front and centre — privacy of client data is one of the five principles for safe and ethical AI use in finance.
  • Be deliberate, not paranoid. Pick an option where you can answer the "where does our data go" question clearly.
Deeper dive
Chapter 3

Industry
guidance.

The MFAA's framework for safe and ethical AI use in finance.

Ch 3 · Industry guidance

Five principles, one rule

The MFAA discussion paper identifies five themes for safe AI use in broking:

  • Privacy — be deliberate about what data goes where.
  • Bias & accuracy — verify outputs, especially numbers and names.
  • Accountability — "the AI did it" is not a defence. Humans own outputs.
  • Transparency — be clear with clients about how AI is used in their service.
  • Human element — AI augments judgement, never replaces it.

The one rule underneath all five: a human reviews everything before it reaches a client.

Ch 3 · Industry guidance

A starting policy, free to use

We've drafted an MFAA-aligned acceptable-use policy template you can adapt for Coltbridge — no charge, no strings.

Download · Markdown
AI Acceptable Use Policy Template
8 sections, ~3 pages. Privacy rules, accountability, incident handling. Replace bracketed fields with your specifics.
Download

Aligned to MFAA principles. Not legal advice — review with your compliance adviser.

Chapter 4

Yes, you can
do this.

A subscription like Claude Pro is genuinely fine — with the right habits.

Ch 4 · Yes, you can

Claude Pro / Team is fine for Coltbridge

ACL-licensed firms are not bound by APRA-style data residency rules. With basic obfuscation habits, a standard Claude subscription sits comfortably inside the MFAA principles — and inside the Privacy Act.

The mechanism: de-identified data isn't "personal information" under the Privacy Act, so APP 8 (cross-border disclosure rules) doesn't apply. Anthropic's commercial terms also explicitly state they don't train on customer data.

The honest read

Subscription Claude with sensible habits is the right starting point for 95% of firms in your position. You don't need onshore Bedrock to be safe. You need a policy, an AI lead, and the discipline to obfuscate before pasting.

Ch 4 · Yes, you can

The habits that make it work

Do Replace client names with generic placeholders. "Client A", "Lender B", "the Borrower." Round identifying figures slightly.
Do Use it for structure, drafting, review, and analysis — anywhere the client identity isn't essential to the task.
Do Keep prompts portfolio-generic. "Mid-market property refinance, $40m, three-year term" instead of "Smith Family Trust, 12 Acacia Street."
Don't Paste full identity documents, bank account numbers, TFNs, or anything where a leak would matter even without names.
Don't Combine client name + financial position in the same prompt. The combination is what makes data identifying.
Don't Use free or personal accounts for work. Different data terms apply. Stick to the firm's paid subscription.

This isn't legal advice. Confirm with your compliance adviser before formally adopting. But yes — your instinct is right. You can use this.

Chapter 5

Four real
options.

Everything else is a flavour of one of these.

Ch 5 · The options

At a glance

Claude Direct

Best for drafting
Reads your M365 data
Cheapest to start
Inference in US

ChatGPT Enterprise

Polished out-of-box
AU data-at-rest
Inference in US
More expensive

M365 Copilot

Lives in Word, Excel, Outlook
Can write/edit docs directly
Less flexible

Bedrock Claude (Sydney)

Inference in Australia
Strongest residency
Engineering setup needed
Ch 5 · The options

Claude can now read your M365 directly

Released late 2025 by Anthropic. Available on all Claude plans. Once enabled, Claude can search and read your SharePoint, OneDrive, Outlook, Teams, and Calendar — without copy-paste.

An advisor can ask: "Find the IM template we used for [Client X] last quarter and adapt it for [Client Y] using their briefing notes in SharePoint."

Read-only · Mirrors existing M365 permissions · Setup ~5 minutes · Doesn't change residency

Chapter 6

What does it
actually cost?

Cheap. Cheaper than most people expect.

Ch 6 · Costs

What you'd pay AI vendors

Claude Pro
$30/mo
One person · Try-it-out tier
Claude Team
$230/mo
Five seats · M365 connector included
Bedrock Sydney
$130–200/mo
If you graduate to onshore inference

All AUD. Paid directly to vendors.

Ch 6 · Costs

In practice — real Coltbridge tasks

Drafting an IM section
$0.08
~4 pages of output
Reviewing 50pg term sheet
$0.15
Flagged covenants + recommendations
30-min iteration session
$0.60
Back-and-forth on a structure
The takeaway

If all 5 advisors used AI heavily every working day, raw costs would be $200–400 AUD/month total. Subscription pays for polish and predictability.

Chapter 7

How to
start.

A simple, low-risk path you can run yourselves.

Ch 7 · How to start

Five steps over three to four weeks

  1. Pick an AI lead. One person, curious about tools, willing to learn. They own this.
  2. Adapt the policy template we provided. Fill in the brackets, agree the rules as a team.
  3. Trial Claude Pro ($30/month, one person, two weeks). Use it for 30 minutes a day on real but sanitised work.
  4. Show and tell. AI lead presents 3–5 examples to the team — what worked, what didn't.
  5. Decide. Roll out Claude Team for all five (~$230/mo) and turn on the M365 connector. Or pause if it's not landing.
Chapter 8

Things to
watch out for.

Common ways AI gets used badly. Avoid these.

Ch 8 · Pitfalls

Six common traps

  • Trusting numbers. AI fabricates figures confidently. Verify every financial value, percentage, date, and citation before it leaves the firm.
  • Pasting client-identifying data into free tools. If it's not on your approved-tools list, treat it as the public internet.
  • Letting AI hit "send." Drafting with AI is fine. Sending without a human review step is not.
  • Believing it understands your firm. AI is great at general knowledge, average at specifics. It will get lender names, regulations, and product details wrong.
  • Skipping the policy. "We'll write it later" tends to mean "we'll write it after the incident." Do the policy first.
  • Treating it as autonomous. AI is a junior colleague who needs supervision, not a replacement for judgement.
Chapter 9

When you
want help.

Optional. Hourly. No retainers, no minimums.

Ch 9 · When you want help

If you hit something you can't figure out

Most firms can run the five-step path on their own. But if you find a workflow you'd like to automate and can't quite get there — we're available ad-hoc at $200/hr + GST.

Common things people ask us for:

  • Prompt engineering for a specific recurring document or task.
  • Workflow design — turning "wouldn't it be great if AI could..." into something that actually works.
  • Tool selection & setup when something more advanced than Claude Team is genuinely needed.
  • Compliance review of how you're using AI, against MFAA principles.

No minimums, no retainers. Engagements quoted upfront before work begins.

Final · Questions?

Got questions?
Want a hand?

45 minutes, free, no obligation. Bring questions, situations you're not sure about, or a workflow you'd like to think through.

Book a Call ▸ Read full briefing