By Stephanie Dempster, Senior Director, Brokerage Services and AI Automation, TradeStation

On May 6, 2010, nearly $1 trillion in market value disappeared from U.S. equities in a matter of minutes.

The event, known as the “Flash Crash” on Wall Street, was not triggered by historically reliable suspects, like a cyberattack, a geopolitical crisis, or a massive collapse in corporate earnings.

The crash stemmed from a chain reaction in which automated algorithms responded to the market faster than humans could comprehend what was happening. Markets recovered most of those losses within about 36 minutes. But the lesson hit hard for markets and investors: technology can and will amplify small problems into major disruptions when controls fail to keep pace.

Fast forward to 2026, and we’re at an inflection point as AI has quickly become embedded across brokerage operations, trading platforms, compliance systems, and customer-facing tools.

AI Is Not Just Another Form Of Automation

Before we can solve the controls problem, we have to identify the culprits. One of the biggest misconceptions I see is treating AI as simply the next generation of algorithmic trading.

I view traditional automated systems as largely deterministic, operating according to predefined rules. If a certain condition occurs, the system takes a specific action. Historically, the logic is structured, predictable, and relatively straightforward to audit.

AI systems operate differently. They generate outputs based on patterns, predictions, and statistical relationships. They adapt and learn quickly. Yet there’s one major problem. AI models can also produce answers that appear entirely reasonable while being fundamentally wrong.

Exchanges and trading firms need to understand why. A traditional trading algorithm may execute an order incorrectly because a rule was poorly designed. An AI-powered system can introduce additional errors, like making a flawed recommendation, because it misunderstood context, relied on inaccurate data, or generated a hallucinated conclusion. The output may appear convincing enough that a user, or even an operator, initially accepts it as accurate.

For brokerages, that creates a fundamentally different supervisory challenge.

The Risk Is Operational, Not Just Trading

When people hear “AI risk,” they often envision autonomous trading systems placing bad trades.

That’s certainly one concern. But from my perspective inside brokerage operations, some of the most immediate risks involve the processes that support trading itself.

Consider margin management.

Today, many firms still rely on employees reviewing reports, spreadsheets, and account balances to identify margin deficiencies and determine appropriate actions. Naturally, firms want to automate portions of that process using AI. The efficiency gains could be significant.

Yet imagine an AI model performing margin calculations incorrectly because of flawed data inputs, incomplete information, or model errors. Accounts may suddenly be flagged incorrectly. Liquidations may occur unnecessarily. Risk exposure could be miscalculated across thousands of accounts, exposing the firm to reputational and financial harm.

The real danger is that AI can make mistakes at scale, with blinding speed, and an alarming level of confidence that makes those errors harder to detect.

Data Quality May Be The Biggest Challenge Of All

One of the least discussed AI risks is also one of the most important: data integrity.

Every AI model depends on the quality of the information feeding it. If the underlying data is inaccurate, incomplete, delayed, or inconsistent, the resulting outputs become questionable regardless of how sophisticated the model may be.

The old technology principle still applies: garbage in, garbage out. Many firms today are focused on building AI solutions. But deploying them successfully is proving far more difficult than creating prototypes.

Across financial services, organizations are discovering that building an AI model is often easier than establishing confidence in its outputs. Testing, validation, governance, and monitoring all remain significant hurdles that largely haven’t been addressed.

For brokers, the first step is recognizing where the real challenge lies. It has little to do with deploying AI. It has everything to do with knowing when to trust it enough to put it into production.

That trust gap is real, and it’s widening as AI moves to the center of brokerage operations. According to the Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report, data availability and quality remain the leading pain point hindering AI adoption, cited by industry respondents. For broker-dealers, that reinforces a critical point: AI tools are only as reliable as the data, governance, testing and monitoring frameworks behind them.

Existing Control Frameworks Weren’t Built For AI

Fortunately, the brokerage industry has a strong history of building risk controls.

Exhibit ‘A’ is the post-2008 reform era. Regulators introduced Rule 15c3-5, requiring broker-dealers to implement robust pre-trade risk controls. After the Flash Crash, markets strengthened surveillance systems, circuit breakers, and execution safeguards. Additionally, the Consolidated Audit Trail, mandated by the U.S. Securities and Exchange Commission under Rule 613, improved transparency and post-trade visibility.

These frameworks remain valuable, but they were largely designed for human decision-makers and rule-based systems.

AI introduces challenges that are harder to supervise, with these issues at the top of the list:

• Non-deterministic outputs
• Model drift over time
• Hallucinations
• Data poisoning risks
• Adaptive behavior
• Difficult-to-explain decision chains

The problem is that existing controls excel at answering the question: What happened? AI, on the other hand, requires firms to answer a different question: Why did it happen?

Understanding that distinction, and what to do about it, may become one of the defining regulatory and operational challenges of the next decade.

The Industry Needs Governance, Not Fear

I’m optimistic about AI’s future in financial services.

Brokerages should embrace AI to remain competitive. Most industry insiders agree, with 73% of executives saying AI is crucial to their future success, according to NVIDIA’s 2026 State of AI Financial Services Report. Additionally, 65% of respondents said their company is actively using AI, up from 45% in NVIDIA’s 2025 report.

Many brokerage firms may believe they have no choice. After all, financial markets are becoming increasingly global and continuous. Investors expect faster responses, better experiences, and more personalized insights. Operational teams face growing complexity and rising volumes. AI can help firms manage all of those challenges more effectively.

Brokerages should embrace innovation while prioritizing governance.

In practice, that means validating every model against real brokerage data before it touches a live account, not just in a sandbox. It means setting explainability standards, so a flagged margin call or forced liquidation can be traced back to the data and logic behind it. It means building AI-specific audit trails that capture not only outcomes, but also decision pathways and data sources. And it means keeping a human in the loop on any decision that moves money or risk.

One encouraging trend is that many firms are using AI to strengthen compliance rather than replace it. The recent ACA Group and the National Society of Compliance Professionals survey shows brokers increasingly leveraging AI for policy and procedure development, monitoring and testing, and communication and training. Rather than eliminating human judgment, successful implementations are augmenting it.

That hybrid approach may ultimately prove to be the most effective model.

Could AI Trigger A New Market Event?

History teaches us that market structure risks often become visible only after adoption reaches scale.

The 2010 Flash Crash demonstrated what can happen when automated systems interact in unexpected ways. AI introduces an entirely new layer of complexity because its underlying decision-making process is often less transparent than that of traditional algorithms.

A crisis is not inevitable. But the firms that build governance into their infrastructure now, before the next event, will be the ones ready for it. The Flash Crash gave the industry 36 minutes and a warning. The next disruption may not be as forgiving.