What Happens When a Bot Acts Before You Review the Trade?

Removing human approval lets a flawed signal become an executed loss because the system cannot pause to check whether the market context, position size, stop, or data still support the trade. An approval gate adds a deliberate decision between an alert and a real order.

On August 1, 2012, Knight Capital Group began trading after a software deployment at its Jersey City operation. The deployment left one server running old code. When the market opened, that code generated a large number of unintended orders. Knight’s systems kept sending them until the problem was stopped, and the firm disclosed a major trading loss soon afterward.

The SEC’s order on the incident documents the operational failure in uncomfortable detail. The relevant question for a retail trader is smaller, but similar: what happens when a system sees a valid-looking input, then acts before anyone checks whether the conditions behind that input still hold?

Educational content, not financial advice.

A signal can be valid and still be wrong for the moment

Imagine checking your account the morning after a sharp market move. An autonomous bot saw a breakout, entered on its rule, and the price reversed after the order filled. The chart may show that the setup existed. That does not mean the trade belonged in your account.

A signal can be flawed in several ways:

• The price moved beyond the entry range before execution.
• Volatility expanded, making the original stop distance too small or the dollar risk too large.
• A correlated position already exposed the account to the same move.
• The strategy used data that was delayed, incomplete, or no longer representative of current conditions.
• The trade came after your daily loss limit or after a string of losses that should have changed your decision.

An autonomous system does not experience hesitation. That can sound useful until hesitation is doing risk-control work. A human reviewer can see that a planned entry at one price became a chase at another. They can reject a trade whose stop would require a position size that exceeds the account’s risk limit.

The point is not to approve every queued trade manually out of fear. The point is to make a conscious decision when the actual order differs from the original idea.

The missing step is context

Knight Capital’s failure was not caused by a market opinion. It was an execution problem: a system acted at scale after a deployment process failed. The system had no meaningful pause between the bad condition and the outgoing orders.

Trading bots can create a similar chain on a smaller scale:

1. A rule detects a setup.
2. The bot calculates an entry and size.
3. An order reaches the market.
4. The market moves before the account owner reviews the downside.

Each step can look reasonable alone. Together, they can turn a modest signal error into a real position.

Approval-gated trading changes the sequence. The AI can generate and queue a signal, but the order waits for review. Before approval, you can compare the proposed entry, stop, position size, and maximum loss with your plan. You can also reject the signal without needing to override an already open trade.

That review can take seconds. Those seconds are enough to ask: Has the price already moved? Where is invalidation? What is the dollar loss if the stop is hit? Does this trade fit the day’s remaining risk?

For a practical version of that review, use the five seconds before approving a trade.

Risk limits need to exist before the alert arrives

The morning-after problem often starts the night before. A bot can only respect limits that were defined clearly enough to apply.

Set a maximum loss per trade in dollars or as a percentage of account equity. Define a daily loss limit. Decide whether correlated positions count toward the same risk budget. Write down what invalidates a setup before a signal appears.

Then make the approval screen answer the same questions every time:

• What is the entry price?
• Where is the stop?
• How many shares, contracts, or units does this order use?
• What is the maximum planned loss?
• What changed since the signal was generated?
• Does approving this order break a limit set before the session began?

A visible record matters here. You can later review queued, approved, rejected, and executed trades to see whether the system found poor setups, whether your approvals ignored risk limits, or whether the strategy needs further testing. A Friday review of queued, approved, and rejected trades can turn that record into a repeatable process.

Approval protects the decision, not the outcome

An approval gate cannot prevent every loss. A trade can meet every rule and still fail. Markets can move through stops, gaps can change realized losses, and a sound strategy can experience drawdowns.

Its value is narrower and more useful: it prevents execution from becoming automatic permission. You retain the chance to reject a trade when the current price, size, or account exposure makes the signal unsuitable.

Knight Capital’s 2012 event shows what can happen when an automated process continues after a bad condition enters the system. Retail trading carries different stakes and different mechanics, but the operational lesson holds. Put a checkpoint between the signal and the order. Make the final decision while the trade is still optional.