𝐓𝐡𝐞 𝐀𝐈 𝐦𝐚𝐫𝐤𝐞𝐭 𝐫𝐢𝐬𝐤 𝐞𝐯𝐞𝐫𝐲𝐨𝐧𝐞 𝐰𝐚𝐭𝐜𝐡𝐞𝐬 𝐢𝐬 𝐚 𝐜𝐫𝐚𝐬𝐡.

Let’s start with a correction, because the headlines often get this wrong.

The Bank of England checked how trading firms actually use autonomous AI agents, and reported in June that agents mostly handle research and other low-risk work rather than buying and selling.

You will also read that roughly 70 percent of stock trading is automated, which is both true and misleading. That number counts execution, and machines have been placing orders for twenty years by following rules that people wrote. Choosing the trade is a different job, and humans still hold it.

So we are overstating today’s risks. But we may be underestimating tomorrow’s.

Three finance professors, at Wharton and in Hong Kong, replaced human traders in a market simulation with agents that learn by trial and error. Nobody instructed them to cooperate, and they had no way to communicate.

𝙏𝙝𝙚𝙮 𝙘𝙤𝙡𝙡𝙪𝙙𝙚𝙙 𝙖𝙣𝙮𝙬𝙖𝙮.

The agents worked out that trading gently paid better than competing hard, then learned to punish any agent that broke ranks. No meeting, no message, no intention, and nothing a regulator would recognize as an agreement.

To be clear, this was a simulation and not an observed market.

It still points somewhere uncomfortable. I argued recently that shared models would make everyone move at once, and this is the mirror image.

A crash is loud, and you know the day it happened.

𝐓𝐡𝐢𝐬 𝐨𝐧𝐞 𝐢𝐬 𝐪𝐮𝐢𝐞𝐭. A market exists to convert scattered private knowledge into a single honest public number, and it could stop doing that while looking perfectly calm. The screen still displays a price, but the price carries less information about genuine supply and demand.

That reaches you even if you never trade. Your hedge assumes the price is informative, and so do your procurement benchmark, commodity budget and next valuation. 𝐏𝐫𝐢𝐜𝐞𝐬 𝐚𝐫𝐞 𝐚𝐥𝐬𝐨 𝐜𝐞𝐧𝐭𝐫𝐚𝐥 𝐭𝐨 𝐡𝐨𝐰 𝐚𝐧 𝐞𝐜𝐨𝐧𝐨𝐦𝐲 𝐝𝐞𝐜𝐢𝐝𝐞𝐬 where capital and materials go, so when such an essential signal quietly degrades, it damages every decision downstream of it.

Yet there is one thing you can still see.

When agents restrain each other, information reaches the price slowly, so the price under-reacts to real events. You cannot spot that from a screen, but you can spot it from a warehouse. Your order book, supplier lead times and freight bookings are an independent read on supply and demand that no market handed you.

When something real changes and the price barely moves, stop assuming your data is the one that is wrong. As the public signal weakens, 𝐲𝐨𝐮𝐫 𝐨𝐰𝐧 𝐩𝐡𝐲𝐬𝐢𝐜𝐚𝐥 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐛𝐞𝐜𝐨𝐦𝐞𝐬 𝐰𝐨𝐫𝐭𝐡 𝐦𝐨𝐫𝐞.

We spent a century building tools to catch bad behavior in markets.

We have built almost nothing to catch impeccable manners.

That risk is coming. Soon. Are we prepared?

#AIAgents #Markets #RiskManagement #Leadership #AICollusion

Contact

bruno.gentil@sherpaconsultingasia.com

© 2026. All rights reserved.