Anthropic published a case study on Millennium, one of the world's largest alternative investment managers, co-developing a digital risk analyst with Claude: an AI teammate that works alongside the firm's risk managers, not instead of them, to surface risk insights and explain daily changes in risk exposure across asset classes.
The build is a template, not a one-off
The case study is worth reading past the big name at the top, because the shape of the build is the actual template other financial services firms can learn from:
It retains and recalls context over time, so it is not re-explaining itself in every session, a real limitation of chat-only tools.
Every action is logged, tested in a sandboxed environment first, and needs human sign-off before anything is acted on.
It runs on Millennium's own proprietary data rather than a black-box model reasoning in isolation.
Persistent memory, sandboxed testing, human-approval gates, grounding in the firm's own data. None of those four design choices are exotic. They are the same governance pattern any regulated business should demand of an AI system before letting it near anything material, and they are why the build survived Millennium's own internal risk review before it ever reached a live trading desk.
Why this maps cleanly onto APRA and ASIC expectations
For Australian financial services, that combination maps closely onto what APRA and ASIC already expect: an auditable reasoning trail and human sign-off on anything material, the kind of control that survives a regulator asking how a system reached its conclusion. A super fund or wealth manager evaluating an AI risk tool should be asking the same four questions Millennium's build answers: does it remember context across sessions, is every action logged and testable before it runs live, is a human required to sign off before anything material happens, and is it grounded in the firm's own data rather than a generic model's priors.
That last point matters more than it sounds. A risk model reasoning over public market commentary and general training data will produce plausible-sounding output that is disconnected from the fund's actual positions. A risk model grounded in the firm's own portfolio, exposure and counterparty data produces output a risk committee can actually act on, and defend to an auditor.
The Millennium build also settles a question we hear often from AU risk teams: whether an AI system can be trusted with anything material at all in a regulated environment. The honest answer, per this case study, is that it can, provided the governance scaffolding around it does the work a human reviewer would otherwise do by default. The model is not the control. The sandboxing, logging and sign-off gate around the model is the control, and that is the part firms should be scrutinising in any vendor pitch.
What most AU firms can realistically build
Most AU wealth managers, super funds and boutique asset managers do not have Millennium's in-house AI lab or budget, and they do not need one. A scoped build, a Claude-based workflow wired into a firm's own portfolio or risk data with the same logging-and-approval discipline, typically runs A$15,000 to A$35,000 for a first working version, well short of a bespoke research-lab collaboration.
Questions to ask before signing off on a risk-AI pitch
Vendors pitching AI-for-risk tools to AU financial services firms rarely lead with governance, because governance is not the exciting part of the demo. A risk or compliance lead evaluating one of these tools should push past the demo and ask directly:
Where does the model's context come from, is it grounded in our own data, or is it reasoning from general training knowledge dressed up as insight?
What is logged, and can a compliance officer reconstruct exactly why the system flagged, or did not flag, a given exposure change?
Is there a sandboxed test path before anything reaches a live workflow, or does the vendor expect production access from day one?
Who signs off before an action is taken, and can that approval step be turned off later without anyone noticing?
That last question matters more than firms tend to assume during procurement. A human-approval gate that is easy to disable under deadline pressure is not a governance control, it is a governance suggestion. The Millennium build treats the sign-off step as structural, not optional, and that is the detail worth copying regardless of firm size.
The Automata AI take
If you are on a risk, compliance or portfolio team in Sydney or Melbourne weighing where AI actually earns its place in a regulated workflow, the Millennium build is a useful reference point precisely because it is boring in the right ways: memory, sandboxing, human sign-off, real data. That is the conversation worth having before the next audit cycle, not whether AI belongs in a risk workflow at all.
Book a scoping call and we will work through what a Claude-based risk or compliance workflow would look like against your existing audit requirements.
