An AI trading bot executes the rules you gave it, so weak risk rules get executed with machine precision, and a software bug can repeat a mistake faster than you can react. That is why risk architecture matters more than signal quality: a mediocre strategy with strong guards can survive long enough to be fixed, while a brilliant strategy with no guards can end an account in one bad session. Three controls are non-negotiable before anything goes live: a maximum risk per trade, a hard daily loss limit, and a maximum drawdown guard that halts new entries. Each one needs a number, an enforcer, and a test, not just a good intention.

Start with per-trade risk, expressed in dollars and derived from the stop distance, not from how much margin the account allows. A common retail convention is risking 0.5 to 1 percent of equity per trade. On a hypothetical 10,000-dollar account, one percent is 100 dollars; divide that by the stop distance in points times the dollar value per point to get position size. Margin is collateral, not a cap on loss, so sizing from margin is how small accounts take positions far too large for their risk tolerance. Add commissions and a slippage allowance to the calculation, because a strategy that ignores costs will always look better on paper than it performs live.

The daily loss limit is the session-level guard. A conservative convention is 1.5 to 2 percent of equity per day; on a 10,000-dollar account that is 150 to 200 dollars, after which the bot stops opening new trades until the next session. Ten straight maximum-loss days at that level would cost roughly 18 to 20 percent of the account, which is painful but survivable, unlike a single day of unlimited revenge trading by an algorithm. Note that the limit is a trigger, not a guarantee: slippage, gaps, or a disconnect can push the realised loss past it, so set the trigger below the loss you can truly tolerate. The drawdown guard works the same way at portfolio level, commonly pausing trading after a 5 to 10 percent fall from the account peak, with a larger 10 to 15 percent threshold as a hard halt for review.

A kill switch that has never been tested is a decoration. Configure the limit in the strategy, the platform, or the broker, and write down which layer enforces it, because a limit enforced only in a script that crashes is no limit at all. In simulation, deliberately breach it and watch what happens: are new orders refused, what happens to open positions, and does the system fail closed, stopping trading when the data feed drops or an error occurs, rather than continuing blind. Restart the bot mid-session in simulation and confirm it cannot double-fire the same signal or reopen a position it just closed on a loss limit. If any test fails, fix it before live trading, not after.

Add two softer guards that prevent the usual failure modes. A consecutive-loss counter pauses trading after three to five losing trades, breaking the algorithmic equivalent of revenge trading before it compounds. A size-reduction rule halves position size after a weekly drop of around five percent, so the bot trades smaller while it is struggling instead of larger while it is chasing recovery. Finally, write a restart rule: after any halt, trading resumes only after manual review, not automatically at midnight. Bots do not get tired, bored, or scared, which is their strength and their danger. The kill switch is what makes that tirelessness safe, and it deserves more engineering attention than the signal itself.