Autonomy isn’t enough: rethinking AI in Finance

Here’s my take: the future of financial research lies in hybrid systems. LLMs bring flexibility, insight, and creativity, enabling professionals to move quickly since so much can be done out of the box. But finance often demands more: rigor and repeatability. Fully autonomous agents are powerful, yet risky on their own. By combining their adaptability with guided workflows implemented as tools, we get the best of both worlds: innovation that professionals can trust.

The promise and limits of autonomous agents

Autonomous agents and Deep Research capabilities open exciting possibilities for finance. They excel at tackling complex, one-off research questions, whether that’s exploring an unfamiliar company, synthesizing emerging themes, or connecting insights across disparate data sources. Their adaptability and breadth make them powerful when creativity and exploration are the priority.

However, many tasks in finance and investing are inherently repeatable. For instance, monitoring liquidity risk requires calculating turnover, bid-ask spreads, or market depth consistently across time to detect shifts in trading conditions. In these cases, the non-deterministic nature of GenAI poses challenges. The same prompt can yield different outputs, and multi-step reasoning compounds the unpredictability.

In finance, where consistency is often paramount, unpredictability is not not just an inconvenience but a potential risk. For mission-critical workflows, maintaining a high degree of control must remain a central principle.

Guided workflows provide the structure autonomous agents lack

They embed financial expertise, enforce consistency, and produce more deterministic outcomes, such as:

Thematic screeners: define a theme, fetch relevant data, measure exposure, and filter assets in a transparent pipeline.

Risk analyzers: systematically map exposures to macro, geopolitical, or ESG factors with fully auditable outputs.

Narrative miners: trace emerging storylines in news or filings, clustering and tracking them across datasets.

By guiding workflows step by step, analysts avoid reinventing the wheel. They gain robustness, transparency, and reusability, while reserving LLMs for tasks that truly benefit from creativity and flexibility.

Where LLMs add value

GenAI is not sidelined in this model; it is applied selectively in areas where its strengths deliver the greatest impact:

Creativity & expansion: generating variations of search queries and constructing mind maps to explore thematic risks or opportunities.

Verification layers: labeling and validating extracted results before publishing.

Summarization & compression: distilling long documents or outputs into concise, decision-ready insights.

Used this way, GenAI amplifies research without compromising the repeatability or auditability finance demands.

Guided workflows in practice

To make this practical, we’ve built our guided workflows around Bigdata Search, an advanced RAG engine operating on institutional-grade data. Unlike generic retrieval approaches, the Bigdata Search API is fully flexible, giving you complete control over query formulation. This combination drives both precision and recall, ensuring accurate grounding and full traceability of outputs. On top of this foundation, we’ve developed cookbooks powered by the Bigdata API and the open-source bigdata-research-tools (BRT) Python library.

These cookbooks:

Provide step-by-step guidance for common financial workflows.

Deliver modular, reusable components that adapt easily to new datasets or strategies.

Guarantee transparency, so you know exactly how results are produced.

In other words, they give you the best of both worlds: robust, auditable workflows where reliability matters most, combined with controlled flexibility where GenAI can add value.

Guided workflows as tools for Deep Research Agents

Guided workflows don’t exist in isolation from agents, they can be wrapped in a tools container and exposed through an MCP (Model Context Protocol) layer, making them callable by a Deep Research agent.

This approach turns workflows into modular, composable elements. Agents can dynamically orchestrate processes, while tools provide the necessary structure and rigor. For example, an agent might call a tool to run a risk model validation or perform compliance checks before proceeding to the next reasoning step. These tools can encapsulate guided workflows that follow strict, step-by-step processes.

This tools-first strategy shifts the paradigm from “either/or” to “both/and.” Agents bring adaptability; guided workflows deliver reliability. Together, they enable scalable, explainable, and resilient Deep Research solutions.

Finance demands rigor

Autonomous agents open new frontiers for discovery, but on their own they lack the repeatability finance requires. Guided workflows strike the balance: embedding financial expertise, delivering robust outcomes, and harnessing GenAI where creativity is needed most.

By integrating guided workflows as tools within a Deep Research architecture, financial firms can combine the best of both worlds, agents for flexibility, guided workflows for rigor. With cookbooks powered by the Bigdata API and bigdata-research-tools (Github), users can move from experimentation to scalable, trusted workflows that generate real business value.