Jefferies builds AI trade assistant to give traders instant data insights

Jefferies, a global full-service investment banking firm, recognized that its equities trading desks faced a structural efficiency problem: traders needed real-time insights into client behavior, trade patterns, and market trends from millions of rows of data to make split-second decisions, but traditional approaches forced them to rely on subject matter experts for analysis or wait in IT queues to build custom dashboards—a process that took days or weeks. Traders themselves rarely had the time during the trading day or the coding ability to build and maintain such systems, creating a widening gap between available data and the decisions it should inform.

To solve this, Jefferies built an agentic AI trade assistant on Amazon Web Services (AWS). At the core is a domain-specific agent powered by Anthropic Claude running on Amazon Bedrock (AWS's managed large language model service). When a trader submits a natural language query—for example, "Give me the sector breakdown for trading in the U.S."—the agent orchestrates a multi-step workflow. First, it queries Amazon Bedrock Knowledge Bases, which uses Amazon Titan Embeddings to retrieve relevant database schemas, table relationships, and query patterns. With this context, Claude Sonnet (a Claude variant) reasons over the trader's intent and generates syntactically correct SQL. The Query Agent then evaluates its available tools (each exposed as a Model Context Protocol, or MCP, tool) to determine the right data source: whether the in-memory grid for real-time positions or a historical data store for time-series analysis. The Query Executor then runs the SQL, injecting row-level security filters to ensure traders only access data they are entitled to view. Finally, the LLM selects the visualization format, and a markdown UI library renders charts and graphs.

The system is built using Strands Agents, an open-source agent harness SDK that simplifies agent development by letting developers define a prompt and a list of tools in code, then deploy to the cloud. Jefferies chose MCP tools to expose each data source as a distinct tool, which delivers three advantages: extensibility (new data sources can be added without restructuring the core architecture), separation of concerns (each tool encapsulates the logic for interacting with a specific system), and flexibility (the agent can dynamically select which tools to use based on the query). Security is enforced through Amazon Bedrock Guardrails for content moderation and PII filtering, row-level data entitlements, and conversation logging for audit trails and compliance.

Throughout implementation, Jefferies learned several lessons. First, the team mitigated hallucination risk by deliberately not relying on LLMs to generate visualizations—instead, the LLM handles natural language understanding and query generation while dedicated visualization engines render the actual charts and graphs. Second, they used in-memory databases to maximize response times, essential for split-second insights in trading. Third, they invested in observability and user feedback loops, discovering that real-world usage proved dynamic and trader behavior shifted over time. Fourth, they adopted a deliberate language strategy, using Python for LLM interactions and rapid experimentation, and Java for complex business processing to leverage its performance characteristics and integration with existing trading systems.

Since launch, the trade assistant has delivered measurable efficiency gains: traders can redirect time toward client relationships and strategic decision-making rather than manual data wrangling, trading desks can dedicate more capacity to onboarding clients and deepening relationships, and IT teams have been freed from repetitive dashboard creation to focus on higher-value strategic initiatives. Most notably, the ability to securely query millions of rows of equities trading data using natural language has democratized data access, fostering a data-driven culture where decisions are informed by real-time analytics rather than intuition or delayed reports.

Jefferies faced a longstanding operational bottleneck: traders needed real-time insights from vast amounts of data to make split-second decisions, but the process of obtaining those insights was slow. Building custom dashboards required traders to either code (which most could not do) or wait in IT queues for weeks. This gap between available data and actionable decisions created competitive friction.

The firm's solution centers on agentic AI—an AI system that can reason, plan, and act by orchestrating multiple tools and data sources. By combining Strands Agents (an open-source SDK for building AI agents), Anthropic Claude (the reasoning model), Amazon Bedrock (the managed LLM service), and Model Context Protocol (an open standard for connecting to diverse data sources), Jefferies built a system that understands trader intent expressed in plain English and autonomously generates the SQL, executes queries, and renders visualizations. The architecture deliberately separates concerns: Claude handles language understanding and query logic, while separate visualization engines render charts—a design choice that mitigates hallucination risk while preserving accuracy.

The business impact flows from this architectural shift. Traders no longer sit in IT queues; they get answers instantly. That time savings compounds: trading desks can reallocate capacity from dashboard requests and data wrangling to client relationships and strategic decisions. IT teams stop repeating the same dashboard-building work and focus on higher-value initiatives. And by letting any trader ask natural language questions of the data, the system has democratized data access, fostering a culture where decisions rest on real-time analytics rather than intuition or stale reports.

No comments yet. Be the first to share your thoughts!