By Roger Burkhardt, Head of AI and Chief Technology Officer at Broadridge.
We are the final generation of trading leaders to manage organizations made up solely of humans. As agentic AI – autonomous systems that can decide, learn, and collaborate – moves from experimentation to enterprise scale, then leadership, governance, and workforce design must fundamentally change.
The tasks we give to junior traders such as researching executions for a client will increasingly be executed by teams of agents. More senior traders will use these agentic technologies to rapidly respond to anomalies such as a flurry of algo order rejections. They will research client impact and multiple potential causes in parallel: is there material news, a spike in volatility or volume, or is it just simple client data entry error. Speed is critical in these situations, and traders who are adept at managing a team of digital trading assistants will have a real edge in investigation, triage, and rapid resolution.
Beyond these examples, Agentic AI is enhancing workflows for pre-trade analytics, post-trade cost attribution, anomaly detection, risk management, and surveillance.
These new capabilities require us to rethink career paths to provide richer, more challenging roles for entry level traders and a learning path for senior traders who may have greater confidence in their knowledge of markets than in their ability to manage the new technologies. The ability to constantly learn and adapt will become a primary hiring and promotion criteria.We need to invest in leveraging new types of AI or we will become uncompetitive, but we are an industry that has been using algorithms and AI for automated trading for decades and we need to avoid becoming overly fixated on the newest forms of AI. That is to say with large language models and loosely defined “agentic” systems – without sufficient discipline in when and how to apply them to high-stakes trading decisions in our regulated environment.
LLMs are remarkable at reasoning over unstructured information, generating hypotheses, and supporting human decision-making. They are powerful copilots but inherently non-deterministic. But in automated trade execution – where financially consequential decisions must be fast and repeatable – we must ask a harder question: Do we need probabilistic, opaque reasoning here, or do we need deterministic, explainable behavior?
In most execution and market making scenarios, the answer is the latter. When an algorithm decides to cross the spread, pull liquidity, or aggressively unwind a position, traders, risk managers, and regulators need to understand why. Deterministic and explainable AI models – grounded in well-defined objectives, constraints, and market signals – offer predictability, audit-ability, and trust. They can be stress-tested, bounded, and governed in ways that LLM based systems cannot. These deterministic AI systems also demonstrably meet the latency demands of modern electronic markets.
We should not underestimate the power of proven Reinforcement Learning (RL) techniques to create deterministic learning systems. These systems continuously improve their behavior by analyzing historical outcomes and recent market responses. They can refine parameter selection, venue routing logic, execution timing, or even strategy selection itself. Over time, this creates a compounding advantage: the more the system trades, the more it learns about the specific instruments, venues, and conditions in which it operates.
Market making and liquidity provision are good examples. Intelligent quoting agents dynamically adjust spreads and sizes in response to order flow toxicity, inventory risk, and short-term volatility. Rather than static rules or manually tuned parameters, these agents learn how to balance competitiveness with risk, optimizing capital usage while maintaining compliance with internal and external constraints.
At Broadridge, we see the future not as a single GenAI paradigm winning, but as a layered AI architecture, with each layer optimized for a distinct class of trading decisions. At the core sit deterministic, low latency engines responsible for high stakes execution and market making decisions, where predictability, explainability, and control are nonnegotiable. Above that are adaptive learning systems, including reinforcement learning, which continuously refine parameters, routing logic, and timing within clearly defined objectives and risk constraints.
Sitting on top is agentic, nondeterministic AI acting as a Copilot – providing context, surfacing risks not captured in underlying models, and assisting traders in configuring, launching, and supervising strategies. In this role, agentic workflows add value not by replacing proven quantitative models, but by monitoring when assumptions break down, highlighting emerging risks, and recommending actions that remain firmly under human control.
The firms who win will not be those who adopt the most fashionable AI – but those who choose wisely and invest in their traders training and career paths. Training and governance will set clear policies on which type of AI to use for each workflow, and as a result drive improved trading costs, client service and risk management. These firms will also attract and retain top trading talent which creates a virtuous flywheel effect on those same benefits.
Subscribe to our newsletter
