Olga Magomedova
6 min readBy Olga Magomedova

The sound practice and the day the leverage sold itself

On 15 October the Financial Stability Board published twelve sound practices for AI in finance. Three weeks later, leveraged ETFs sold a record amount into a falling market. Olga Magomedova on what links the two events, and what it means for a desk.

A quiet document and a loud week

The document arrived without drama. On 15 October 2026 the Financial Stability Board published its final report setting out twelve sound practices to help financial institutions adopt artificial intelligence responsibly. The practices cover governance across the full AI lifecycle, from development through retirement, and the board encourages senior management to consult them when shaping strategy and risk.

Three weeks later the market produced a practical illustration of why such a document exists. On 5 November the Nasdaq Composite closed down 5.52 per cent, its worst session in over a year, as concerns about AI data centre regulation after the US midterms collided with crowded positioning. Morgan Stanley's derivatives team estimated that leveraged exchange traded funds had to sell roughly fifty seven billion dollars of underlying stock that day, the highest figure in records going back to 2007. Net leverage at US long short hedge funds fell five per cent in a single session, the fifth largest one day drop since 2010.

Neither event was caused by AI deciding anything. Both are the environment in which AI now operates.

Sound practice ten and the pause button

Of the twelve practices in the FSB report, the one that matters most to a trading desk is the tenth. It mandates meaningful human oversight, with risk based triggers that require an AI agent to seek human approval before executing a transaction or changing a strategy in higher risk scenarios. The report is unusually direct about why the trigger is needed. AI agents do not reason like humans, and they are prone to interpreting predefined goals in ways that can be profoundly unexpected.

That sentence is the whole argument. An agent optimises what it was told to optimise. The trader optimises survival. When the two diverge, the trader has to be the one who can still reach the switch.

For Magomedova, the engineering frame is familiar. Every automated system in an aircraft has a defined envelope outside which a human takes the controls. The envelope is not a compliment to the human. It is an admission that no system fully anticipates its environment. The FSB has written the same admission into the financial sector.

What the November rout actually showed

The November 5 session is not usually described as an AI story, because no model placed the orders. But the mechanical selling that amplified the move was itself a form of automated decision making. Leveraged ETFs do not decide whether anything has changed about the underlying companies. They rebalance to maintain a target exposure, and when the market falls hard enough, the rebalance is large enough to move the market further.

The trader at the screen did not need an AI agent to experience a day of automated selling. The day arrived on its own. What the FSB document, and the broader shift toward agentic trading, will add to future sessions of this kind is a second layer of systems that act at machine speed on objectives their operators may not have fully specified.

The question for a desk is not whether that layer will fail. It is which failure mode it will produce, and whether the plan survives it.

The practice that belongs on the desk

The FSB is addressing institutions. The principle translates downward. A retail trader using an AI signal, an auto routing broker, or an agent that can place orders is operating a small version of the same system the regulators are trying to govern. The sound practice that matters is the same. Define the envelope. Decide in advance which actions require the trader's own approval. Keep the authority to intervene.

For Magomedova, the discipline is older than the technology. A trading plan already specifies size, stop, and the conditions for doing nothing. Adding an AI tool does not reduce the need for those lines. It raises it. The tool is faster than the trader, which means the tool's mistakes arrive faster than the trader's corrections unless the corrections were written before the session began.

People are attracted to the idea that software can remove emotion. But discipline is not something you outsource.

The principle the week returns to

The November rout was not a referendum on AI. It was a reminder that automated processes, from leveraged ETF rebalancing to agentic trading prototypes, now sit inside the same tape a discretionary trader is reading. The FSB report is a reminder that the responsibility for what those processes do is still located in a human. The two reminders arrived in the same month.

The practical response does not change with the headlines. Position sizing that survives a mechanical selling day. A plan that already contains the worst session the strategy was designed to produce. An AI tool used as a second opinion rather than as a delegate. A post mortem on the days when the tool and the plan disagreed, written while the memory is honest.

Build the envelope before the market tests it. The regulators have now written down what the desks that survive these weeks already practise.