Chris MihosHaving access to the same model as everyone else doesn’t make you the best trader in the world, but it can make you a much better judge of who is.
AI makes skilled people faster and more capable, but it doesn’t supply the skill itself, and in markets that distinction matters enormously. When leading AI models were given real money to trade in public contests, most of them lost it, often by trading too much and using too much leverage, which are the same mistakes that cost self-directed traders so dearly. The specialists who consistently do well combine data, infrastructure, risk limits, execution and years of judgement, and AI accelerates every part of that. For most capital owners, the better use of AI is to understand, choose and oversee those specialists, which is what family offices have always done, rather than to try to become one.
What happened when AI models traded real money
In late 2025, the research firm Nof1 ran a contest called Alpha Arena, giving six leading AI models $10,000 each to trade onchain markets for about two weeks. Alibaba’s Qwen3-Max finished with a gain of around 22%, but most of the field lost money, and reported losses for some of the best-known models ran to roughly half of their capital or more. Nof1 itself cautioned that the results might reflect luck, given the short period and the limited information the models had, and commentators noted that the losing models tended to over-trade, take on too much exposure and lean heavily on leverage.
Bloomberg’s review of these contests in May 2026 summed up the pattern simply: most AI bots were losing money, they traded too much, and they often made very different decisions when given identical instructions. Its conclusion was that AI isn’t yet ready to replace a fund manager.
None of this means AI is useless in markets. It means that a general-purpose model, however fluent, doesn’t come with an edge.
Why fluency isn’t edge
A large language model can explain a trading strategy beautifully. It can describe how a basis trade works, why a market-neutral book hedges its exposure or how volatility affects option prices, and it can do so more clearly than many textbooks. That fluency is valuable, but explaining a strategy and running it profitably are very different things.
Markets are competitive in a way most other fields aren’t, because every profitable trade has someone on the other side, and any edge that becomes widely known tends to disappear. When millions of people use the same models with similar prompts, they risk arriving at similar ideas at the same time, which is a recipe for crowded trades rather than an advantage. Self-directed trading already has a sobering record: when European regulators required providers of contracts for difference to disclose how many retail accounts lost money, the figures typically ranged between 74% and 89%. Adding a model to that process doesn’t change the underlying arithmetic, and it can make overconfidence easier, because the model’s answers sound so assured.
What specialists actually bring
The managers who sustain good results over many years tend to combine several things that a general model doesn’t have on its own. They have proprietary data and research processes refined over time, execution infrastructure and relationships with venues and counterparties, and risk limits that they follow even when it’s uncomfortable. They also bring experience of how their strategy behaves in stressed markets, which is something no backtest fully captures, and accountability to the people whose capital they manage.
AI accelerates every one of those, because it can process more research, monitor more positions, test more ideas, write and check code faster, and spot anomalies sooner, which means a skilled team can do more and do it better. That’s the sense in which AI is an accelerant: it multiplies what’s already there. Applied to real expertise, it compounds it, and applied to none, it mostly compounds the mistakes.
How capital owners should use AI instead
The family office model offers a better path, because family offices rarely try to trade every strategy themselves. They choose specialists, oversee them closely and keep the decisions about risk, liquidity and access in their own hands. AI is extraordinarily good at supporting that kind of work.
It can help you understand a strategy before you allocate to it, read the offering terms and summarise the liquidity provisions, compare track records while paying attention to where each number came from, and monitor the custody, drawdowns and redemption queues of the managers you’ve chosen. It can also help you prepare better questions for a manager and notice when their answers don’t match the data. Used this way, AI makes you a much better allocator, which is a more realistic and more valuable goal than becoming a better trader than the specialists.
The risks worth staying across
The first risk is overconfidence, because a model that answers every question fluently can create an illusion of competence, and the people most at risk are those who start to believe that a good conversation about markets is the same as an edge in them.
The second is overtrading and leverage, since the AI contests showed the same patterns that cost self-directed traders money, which are frequent trading, excessive exposure and borrowed money, and an always-available agent makes those behaviours easier, not harder.
The third is inconsistency and fabrication, because models can give different answers to the same question and can state figures that aren’t true, so any number that matters should come from a labelled source rather than from a model’s memory.
The fourth is manipulation, since agents that read external content can be steered by instructions hidden within it, which is why permissions matter so much, as we explain in Permissioned agents.
The fifth is crowding, because if many people use the same models in the same way, their decisions can converge, and crowded positions tend to unwind together.
The last is accountability, because when a model makes a poor decision, there’s no one to answer for it, which is the opposite of what allocators look for in a specialist.
How MakeBanc fits
MakeBanc connects capital owners with independent strategy operators they already know, and gives them the oversight tools to see what those operators are doing. Every number is labelled as self-reported, observed or verified, NAV is calculated by a third-party administrator, the liquidity terms and costs are visible before you allocate, and your assets sit with an institutional custodian in a vault you sign for, where the manager can trade but never move them.
MakeBanc doesn’t select, rank or recommend managers, and it doesn’t trade for you. Its agent, when it launches, only follows the allocation rules you set, and you sign every move. You choose the specialists, and the infrastructure helps you oversee them, whether they use AI, experience or, increasingly, both.
Frequently asked questions
Can AI make you a better trader? It can make a skilled trader faster and more effective, but it doesn’t supply an edge on its own. In public contests where leading AI models traded real money, most lost money, often through over-trading and excessive leverage, and their results varied widely even with identical instructions.
How did AI models perform in the Alpha Arena trading contest? In Nof1’s first Alpha Arena season in late 2025, six leading models each traded $10,000 in onchain markets. Alibaba’s Qwen3-Max gained about 22%, but most models lost money, some losing around half their capital or more. Nof1 cautioned that the results may have reflected luck.
How should capital owners use AI in wealth management? To understand, choose and oversee specialists rather than to replace them. AI can read terms, summarise liquidity provisions, compare track records with attention to their sources, and monitor custody, drawdowns and redemption queues, while the owner keeps the decisions.
What are the risks of using AI to trade? Overconfidence, overtrading and leverage, inconsistent or fabricated outputs, manipulation through prompt injection, crowding when many people use the same models, and the absence of anyone accountable when a decision goes wrong.
Sources
- Bloomberg Government, AI bots auditioning for Wall Street trading are mostly losing, 6 May 2026
- South China Morning Post, Alibaba’s Qwen returns 22 per cent in two weeks in crypto trading showdown
- Decrypt, AI crypto trading showdown, October 2025
- Business Standard, AI bots auditioning for Wall Street trading are mostly losing money, May 2026
- ESMA, Notice of product intervention decisions on CFDs and binary options, 2018
- Giskard, How Grok got prompt-injected, May 2026
- UBS, Global Family Office Report 2026
MakeBanc is a technology platform that connects capital owners with independent strategy operators. MakeBanc does not provide investment advice, custody services or banking services, and does not select, rank or recommend any manager or strategy. Past performance is not indicative of future results.