Finch: AI-Powered Financial Data Retrieval for Real-Time Insights at Uber

Introduction

At Uber, financial analysts and accountants handle vast amounts of data to support real-time decisions. Traditional methods often require complex SQL queries across multiple platforms, causing delays that can hinder timely financial insights. To solve this, the team created Finch—an AI agent integrated right into Slack®.

Finch converts natural language questions into structured data queries, providing secure, instant financial intelligence. This lets finance teams focus on strategy instead of troubleshooting data access, removing the friction that often slows down data retrieval.

Motivation

Finch was developed to make data retrieval easier and more efficient for Uber’s finance teams. After evaluating several technologies, the team chose a mix of generative AI, retrieval-augmented generation (RAG), and self-querying agents. The priorities were data security, smooth integration with existing systems, and scalability.

This combination allows Finch to deliver real-time, secure, and accurate financial insights directly within Slack, where finance teams already collaborate.

The Problem: Traditional Data Access Is a Bottleneck

Financial analysts rely heavily on data to make informed decisions, but accessing that data has been slow and cumbersome. Analysts like Taya face several challenges:

• Manually searching for datasets. Logging into multiple platforms such as Presto®, IBM® Planning Analytics, Oracle® EPM, and Google Docs™ to find the latest figures causes delays and errors.
• Writing complex SQL queries. Even for those familiar with SQL, building accurate queries requires cross-referencing documentation and troubleshooting syntax issues, which is time-consuming.
• Submitting data requests. When queries are complex or permissions are lacking, analysts need to wait hours or days for the Data Science team, delaying critical reports.

This inefficient process clearly demanded a better approach that simplifies data access without compromising security or accuracy.

The Solution: Finch as an Intelligent Financial Assistant

Finch is an AI agent embedded in Slack that streamlines financial data retrieval. It lets users query data in natural language, eliminating the need for manual SQL or data requests. Finch addresses key pain points with:

• Conversational AI in Slack. Users can ask questions like “What was GB value in US&C in Q4 2024?” and Finch fetches the correct data instantly.
• Uber-specific language mapping. It understands internal terms (e.g., “US&C” for US and Canada, “GBs” for gross bookings) and links them to structured data sources.
• Self-querying agents. Finch scans metadata of accessible tables, selects the most relevant source, and generates accurate SQL queries.
• Security and access control. Role-based permissions ensure only authorized users can access sensitive data.
• Automated export to Google Sheets™. Large datasets can be sent directly to Google Sheets for further analysis.

By combining natural language understanding with metadata-driven queries and secure access, Finch cuts down data retrieval time and complexity.

Architecture

Finch uses a modular design integrating internal and third-party technologies within Uber’s ecosystem. It operates on curated, domain-specific data marts and a semantic metadata layer.

These single-table data marts consolidate frequently accessed financial metrics, simplifying queries and speeding up retrieval. This structure helps the AI quickly generate and refine queries, aided by natural language aliases stored in an OpenSearch™ index. This approach improves the accuracy of SQL WHERE clauses far beyond standard LLM-powered SQL agents that rely only on table schemas.

Key Components and Technologies

• Generative AI Gateway. Finch connects to various large language models, allowing easy updates as AI advances.
• LangChain Langgraph™. Coordinates query flow between specialized agents like the SQL Writer and Supervisor agents.
• OpenSearch index. Stores dataset metadata and natural language aliases for fuzzy matching and accurate query building.
• Slack SDK and Slack AI Assistant APIs. Provide real-time communication, suggested questions, pinning Finch within Slack, and a split-pane interface for seamless interaction.

Finch Data Agent Flow

• User query input. A finance team member asks Finch a question in Slack.
• Agent orchestration. The Supervisor Agent routes the query to the right sub-agent, such as the SQL Writer Agent.
• Metadata retrieval. Agents query the OpenSearch index to fetch relevant metadata and aliases.
• SQL query construction and execution. The SQL Writer Agent builds and runs the query after validating user permissions.
• Real-time Slack feedback. A callback handler updates users on each step of the query process.
• Result delivery. Query results are formatted and posted back in Slack, with options to export to Google Sheets.

Integration with Uber’s Tech Stack

Finch works alongside Uber’s internal data platforms like Presto, IBM Planning Analytics, and Oracle EPM. It uses internal services such as the Generative AI Gateway to meet strict security and scalability requirements while delivering consistent, real-time insights.

How Finch Works: A User Journey

Finch fits naturally into Slack, where finance teams already communicate. Instead of juggling multiple systems, users simply ask questions in plain English. For example, Taya might type, “Show me the Q4 2024 GBs value.”

• The Supervisor Agent recognizes this as a data request and sends it to the SQL Writer Agent.
• The SQL Writer Agent identifies the right tables and columns, applies necessary filters, validates permissions, and builds the SQL query. If errors occur, it revises the query accordingly.
• Within seconds, Finch posts a clear summary and a breakdown table in Slack, e.g., “The Gross Bookings (GBs) value for Q4 2024 is approximately $44,197.28M USD.”
• Taya can ask follow-up questions like “Compare to Q4 2023,” and Finch updates the data, shows trends, and offers export options.

This conversational approach reduces delays and helps finance teams make faster, data-driven decisions.

Performance and Accuracy: Ensuring Reliable Insights

Continuous Evaluation & Benchmarking

• Agent accuracy testing. Sub-agents like the SQL Writer are tested by comparing their outputs to verified "golden" queries to ensure accuracy.
• Supervisor routing accuracy. Ensures the right agent handles each query, avoiding misrouted requests.
• End-to-end validation. Simulated real-world queries test overall reliability.
• Regression testing. Historical queries are rerun to detect any accuracy drops after updates.

This methodical testing ensures Finch consistently delivers correct financial insights.

Optimizing for Speed and Scalability

Finch reduces database load by optimizing SQL queries and running sub-agent tasks in parallel. It also pre-fetches commonly accessed metrics to improve response times. This design supports high query volumes while keeping latency low and results reliable.

How Finch Stands Out from Other AI Finance Tools

Finch is built specifically for Uber’s finance teams, providing live, AI-powered financial insights with security and automation. Unlike generic chatbots or traditional BI tools, Finch offers:

• Real-time financial data queries. It retrieves current data dynamically instead of relying on pre-trained knowledge.
• AI-powered self-querying. Finch automatically crafts the best query strategy without manual SQL input.
• Integration within daily workflows. Embedded in Slack, Finch lets users access data where they work, ideal for finance professionals and executives without SQL skills.
• Enterprise-grade security. It enforces authentication, granular permissions, and query validation to protect sensitive data.

Future Roadmap

Finch will expand its integrations with Uber’s fintech systems to create a unified finance ecosystem supporting automated analysis, reporting, forecasting, and budgeting.

Plans include enhancing the user experience for executives like Uber’s CEO and CFO by adding human-in-the-loop validation on demand. This feature will let leaders request expert review for critical queries, ensuring accuracy when it matters most.

Additionally, Finch will grow by adding more agents and toolkits to cover a wider range of finance use cases, aiming to become a comprehensive tool for Uber’s finance teams.

Conclusion

Finch changes how Uber’s finance teams access data, making financial insights faster, more secure, and easier to get. By combining AI-driven natural language queries with secure, metadata-driven data retrieval, Finch eliminates the delays of traditional methods.

As Finch evolves, it will offer more personalized and reliable financial insights, helping teams focus on strategy rather than data wrangling. The future of financial intelligence at Uber is now—and Finch leads the way.