From Chatbot to Portfolio Weight — AI Moves Closer to the Asset Allocation Decision
Most AI sceptics are running one prompt and calling it a test. Three institutional signals suggest the real question is what happens after the fifth iteration.
I have spent more time this year arguing about AI than actually using it, and that tells me something.
The debate has two camps. Believers and sceptics. What strikes me is not the divide itself, but what the sceptics are actually testing. One prompt. Minimal input. Then disappointment when the output is thin. That is not a fair test of AI. That is a fair test of a bad question.
The way I think about it: AI is an iteration process. The more context you give, the more specific your framing, the more rounds of refinement you run — the better the output. By the fourth or fifth exchange, you are no longer talking to a tool. You are inside a dialogue that sharpens your own thinking. Whether that counts as AI performance or human performance is a question I have not resolved.
That is the lens through which I read three recent developments.
ABN Amro Investment Solutions (Citywire Selector, paywalled) CEO Marguerite Bérard stated that the firm has been running an AI portfolio since last year and uses AI to influence investment decision-making. The details remain limited — portfolio construction, manager selection, allocation support or a combination is not specified. But the signal matters precisely because it comes from a large institutional platform, not a technology vendor with a product to sell.
In The Self-Driving Portfolio, Andrew Ang, Nazym Azimbayev and Andrey Kim describe an AI-native architecture built around roughly fifty specialised agents. Some generate Capital Market Assumptions. Others construct portfolios, challenge conclusions, vote on alternatives and test robustness. A meta-agent rewrites prompts and code based on observed outcomes.
The authors make one observation that stopped me: the most binding constraint in institutional asset management is not data availability or model sophistication, but the finite bandwidth of human decision-makers.
That sentence reframes the entire debate. The question is not whether AI is smarter than an allocator. The question is whether AI can expand the bandwidth of the allocator who already knows what matters. That is a tool question, not a replacement question.
Danny Wall (Founder, CEO & CTO @ OA Quantum Labs | Angel Investor for AI and Quantum Projects) recently described a forecasting architecture where independent agents continuously challenge each other’s conclusions and learn from errors. One prompt would never surface that, iteration might.
Three signals. One direction. The conversation is moving from productivity tools toward portfolio construction, forecasting and allocation support. Not there yet — but closer than it was.
That still leaves the question I cannot answer.
The iteration process clearly improves AI output in research and analysis. What I cannot yet answer is whether that same logic translates to the allocation decision itself, where the input is not a prompt but a portfolio, and the output is not a paragraph but a weight.
Is the asset allocator who iterates five times with AI making a better decision? Or just a more elaborate one?
New to this? Every week I test whether AI can improve the way professional investors decide where to put their money, and the honest answer is that the quality of the question matters more than the quality of the tool.