Amundi’s latest working paper and APRA’s April letter arrive at the same conclusion on AI investment management: the technology is ahead of the governance. Here is what Australian asset owners need to build before a supervisor asks the questions.
Amundi Investment Institute’s 192nd working paper landed this month, a survey of agentic artificial intelligence drawing on the finance literature of 2024 to 2026, with direct implications for AI investment management in Australia.
Sitting inside it is a finding Australian asset owners should consider closely. It has almost nothing to do with what the technology can do, but rather concerns whether anyone can prove what it did.
What agentic AI means for investment management
Agentic AI sits a step past the chatbots investment teams have been trialling since 2023. Give one a goal instead of a prompt and it will break that goal into subtasks, reach for tools such as market data APIs and code interpreters, then revise as results come back.
A large model usually runs the show as orchestrator. Smaller and cheaper models handle the repetitive work underneath, with standardised protocols governing how they talk to each other and to the outside world.
Two managers, two approaches
Two of the world’s largest managers have shown their working. BlackRock’s AlphaAgents fields three specialists, one on valuation, one on fundamentals, one on sentiment, each reading a different slice of the evidence before the three argue it out and settle on a signal.
Man Group went the other way with AlphaTrend, a fixed pipeline that generates a trend-following idea, implements it, backtests it and writes it up, with every stage’s inputs and outputs kept retrievable.
The failure mode that counts
Authors Ru Qi Wu and Frédéric Lepetit say something that sits awkwardly against the vendor pitch. More orchestration does not mean better output, and what matters is how capable the individual sub-agents are, and piling on extra coordination layers can make results worse by propagating errors and burning through the context window.
Verification is what binds autonomy, and hallucination remains the failure mode that counts. Here, the paper is unusually direct about the limits of the two fixes currently on display.
BlackRock’s inter-agent debate and Man Group’s parallel querying will both surface an error that varies between agents or between runs.
An error that every agent makes for the same reason, because they share a model prior or a common gap in the source data, sails straight through. Someone still has to check it against ground truth.
Reproducibility is the other half of it. Ask an agentic system the same question twice, on different hardware, and the answers can diverge, because floating point arithmetic accumulates rounding error across long sequences of operations.
Annoying for a research tool. Rather more than annoying for a trustee who has to evidence how a decision was reached. All of which runs into what APRA told the industry on 30 April.
What APRA found when it looked
APRA’s letter on artificial intelligence came out of a targeted supervisory engagement with large banks, insurers and superannuation trustees late last year, and the verdict was that governance, risk management and assurance are trailing adoption.
Testing drew the sharpest comment. Entities were leaning on point-in-time, sample-based assurance, which the regulator called “ill suited to probabilistic models that learn, adapt and degrade over time”. Continuous validation, the kind that catches model drift or a control breakdown while it is happening, was rare.
Internal audit was the next problem. Second-line risk and audit functions frequently lacked the skills and the tooling to assess an AI system independently, APRA found, and the shortfall was worst where agentic behaviour and automated decision-making were in play. Assurance keeps turning up after the deployment it is meant to assure.
Then there is supplier concentration. Some entities lean on a single provider across multiple use cases and have never tested an exit. Upstream, the chain of foundation models, training data and fourth party services is largely opaque.
APRA now expects that chain mapped, and contractual rights held over audit access, model updates and incident notification. Anyone licensing an external investment analytics platform should read that expectation twice.
Where the technology earns its place
Adoption is still the right call. The Amundi survey is at its most convincing where it is least exciting, in the unglamorous middle of the research process.
Agents are already good at pulling figures out of annual filings, turning unstructured disclosure into something queryable, building knowledge graphs and drafting first-pass commentary.
Hours of analyst time go into that work and very little insight comes out of it. Hand it across and a team tests far more hypotheses. That is the case Man Group makes for its own tooling, and it asks nobody to hand a machine a decision.
On where the boundary sits, the paper and the regulator agree. APRA wants humans involved and accountable for high-risk decisions. Amundi concludes that human oversight and explicit verification steps remain necessary in current financial deployments.
The test Australian funds actually face
For Australian funds, responsible AI investment management comes down to a narrower test than the industry conversation suggests. Agentic systems can increasingly produce a defensible investment view.
The harder standard is reproducing that view, explaining the path to it, and putting evidence in front of a supervisor a year later.
Funds that have internalised investment management now own both halves of that. The tooling is here. The assurance capability has to be built.
