On 1 August 2012, a market-making firm in New Jersey deployed an update to a routing system. A configuration was missed on one of eight servers. For forty-five minutes after the opening bell, the firm sent millions of unintended orders into the market. By the time the system was stopped, the firm had taken a four-hundred- and-sixty-million-dollar loss. The capital had been raised in five years. It evaporated in less than a single trading session.
Knight Capital is a story you hear told inside every bank’s engineering org because it answers, in one number, why financial software has the controls it has. The financial system has a long memory of what bad code costs at market open.
The cost of bad software in finance is not measured in user complaints. It is measured in the time it takes to lose the company.
Where AI builds belong on the desk
AI-built software has plenty of room on a trading floor. None of that room sits between a portfolio manager’s decision and the order that lands on an exchange. That path is short, well-controlled, and audited to the millisecond. It is not the path that’s changing because of AI builders.
The path that’s changing is the path beside it. The analyst-tool path. The risk-summary path. The client-letter-generation path. The compliance-monitoring-second- line-of-defense path. The internal-operations-dashboard path. These are the paths where a person who is not, by job title, an engineer, has been wanting tools for years and has been told there is a six-quarter backlog. Those are the tools that AI builders are making feasible. And those are the tools where the press matters.
What the press protects against, in finance
Three things, in the order we see them most often.
The unintended scope. A risk analyst builds an internal tool that summarizes overnight position exposure. The first version works. The second version, six weeks later, has accidentally become the system that drives the morning risk meeting. By the third version, an internal stakeholder uses it for a regulatory report. Nobody intended this. Nobody architected it. The press is for the moment somebody asks can we use this for the regulator report? and a software engineer says not in this shape; let’s harden it before you do.
The vendor-coupling drift. AI tools love third-party APIs. A first-pass version uses a third-party data vendor for a market-data feed; the tool ships; six months later, the firm cannot replace the vendor without rebuilding the tool, because the AI used vendor-specific schema choices that nobody noticed in review. In finance, vendor concentration risk is a regulated concept. The press, here, is for the moment the third-party API is being chosen, not after the dependency has compounded.
The kill-switch absence.If the tool starts behaving badly at 9:31am, who can stop it, and how fast? AI-built first drafts do not have answers to this question. Software engineers do. The press, in many trading-floor sessions we’ve run, ends up being a refactor that adds a circuit breaker, a feature flag, and an oncall path that did not exist in the original build. None of these are exotic. All of them are missing by default.
What this means for buyers in finance
Two things. One: let the analysts and traders build. The productivity is real and the alternative is shadow IT, which is worse. Two: draw a clear line between tools that touch the order path, the books-and-records path, and the regulator-facing path, which need a software engineer in the room before they ship, and tools that do not, which can ship freely. Make the line a policy. Make the press the enforcement.
The financial industry’s instinct to over-engineer the order path was earned the hard way. The instinct will, correctly, extend to AI-built order-path tools, and those will not happen soon. What is happening, today, is the rest of the floor. The press is for the rest of the floor.
Industry essay. We are not investment professionals. None of this is investment advice. All of it is engineering opinion, held loosely.