At Uber, financial analysts and accountants sift through vast data sets to support real-time decision-making. Traditional data access—often involving complex SQL queries across multiple platforms—creates delays that can impact critical financial insights. 

To address this, our team built Finch, an AI agent integrated directly into Slack^®. Finch transforms natural language queries into structured data retrieval, delivering secure, real-time financial intelligence. This solution accelerates analysis and empowers finance teams to focus on strategy rather than troubleshooting data pipelines. This need to eliminate friction in data retrieval was the driving force behind Finch’s development.

We developed Finch to simplify and optimize the process of data retrieval for financial teams at Uber. During the development of Finch, the Uber FinTech team evaluated various technologies, and ultimately, we selected a combination of generative AI, RAG, and self-querying agents. Key factors that influenced this decision included ensuring data security, achieving seamless integration with existing systems, and maintaining scalability for future growth. By using these technologies, Finch empowers finance teams to access real-time, secure, and accurate financial insights directly within their Slack environment.

At Uber, financial analysts rely on data to make critical decisions. Yet, retrieving that data has historically been slow and inefficient. Financial Analysts like Taya often face multiple challenges when preparing reports.

By the time her report is complete, valuable time has been lost, impacting decision-making and potentially delaying key business insights. This friction in traditional data access made it clear that a better solution was needed—one that could simplify data retrieval without compromising security or accuracy.

We developed Finch, an AI agent that streamlines financial data retrieval to address these challenges. Integrated directly into Slack, Finch allows finance teams to query data using natural language, removing the need for manual SQL queries or data requests. 

Finch is designed with several key capabilities that address common pain points:

By combining natural language understanding, metadata-driven query generation, and secure data access, Finch reduces the complexity of financial data retrieval—giving analysts faster access to the information they need.

Finch’s architecture is built on a robust, modular framework designed to deliver financial insights with high accuracy. The solution uses a mix of internal and third-party technologies that integrate into Uber’s tech ecosystem.

What sets Finch apart is our approach to data retrieval: curated domain-specific data marts and a semantic layer on metadata. Finch relies on a curated set of single-table data marts that consolidate frequently accessed financial and operational metrics to maximize efficiency and accuracy. These datasets are optimized for simplicity. They’re structured to minimize query complexity and retrieval time, making it easier for the LLM to write a query and quickly get feedback, like error messages, so it can take action and correct its mistakes. The datasets are also structured and governed. Controlled access ensures only authorized users can query specific financial dimensions.

Finch dramatically improves the accuracy of WHERE clause filters by storing natural language aliases for both SQL table columns and their values in our OpenSearch^™ index. This is a significant enhancement over traditional methods used by LLM-powered SQL agents that rely solely on table schemas and a few sample rows.

Finch accesses a wide array of self-hosted and third-party large language models via Uber’s Generative AI Gateway. This approach allows us to swap out models easily as the underlying AI technology advances.

We use LangChain Langgraph^™ to construct and orchestrate our agents. It coordinates the flow of queries between specialized agents (like the SQL Writer Agent and Supervisor Agent) and ensures that each task is executed in the proper sequence.

Finch uses an OpenSearch index to store dataset metadata, including SQL table columns, values, and natural language aliases. This novel approach allows for fuzzy matching of terms extracted by the LLM. Consequently, Finch achieves significantly improved reliability in generating accurate WHERE clause filters compared to conventional methods.

Using the Slack SDK, Finch connects directly to Slack via API. A dedicated callback handler in our Langgraph application updates Slack status messages in real-time so users can track what the agent is doing at every stage of the query process. We also leverage the recently released Slack AI Assistant APIs, which enable suggested questions at the start of a new conversation, the ability to pin Finch within the Slack application, and an always-open split pane to talk with Finch. 

Here’s the flow of the Finch data agent. 

Finch is designed to work in harmony with Uber’s broader tech stack. It integrates with internal data platforms (such as Presto, IBM Planning Analytics, and Oracle EPM) and leverages internal services like the Generative AI Gateway. This integration ensures that Finch meets stringent data security and scalability requirements as well as delivers consistent, real-time insights to finance teams across the organization.

Finch is designed to integrate seamlessly into Slack, where Uber’s finance teams already collaborate. Instead of logging into multiple systems, users can ask questions in plain English, just like messaging a colleague.

For example, while preparing a report, Taya types in Slack, “Show me the Q4 2024 GBs value.”

As shown in ‌Figure 3, Finch performs a number of steps, which can vary from question to question.

As shown in Figure 4, Finch updates the query, fetching the data from Q4 2023 and summarizing the trend (increase), presents it in Slack, and offers an option to export to Google Sheets.

Finch revolutionizes how Uber’s finance teams access data through conversational retrieval, leading to less friction, fewer delays, and faster data-driven decisions.

Finch is built to deliver fast, accurate, and reliable financial insights at scale. To achieve this, it undergoes rigorous performance testing and continuous optimization.

To maintain high accuracy and efficiency, Finch employs:

This iterative approach ensures that Finch delivers correct financial insights and maintains consistency as new improvements are introduced.

Finch is designed to handle high-demand queries efficiently. Finch minimizes database load by optimizing SQL queries. Sub-agents execute tasks simultaneously, reducing response time. Lastly, Finch pre-fetching frequently accessed metrics enhances performance.

Finch continuously improves its accuracy and speed, ensuring that finance teams receive real-time insights with minimal latency and maximum reliability.

Finch is specifically designed to meet the needs of Uber’s finance teams, offering real-time, AI-powered financial insights with built-in security, automation, and scalability. As shown in Figure 5, unlike generic AI chatbots or traditional BI tools, Finch combines structured data retrieval, intelligent self-querying, and enterprise-grade access controls.

Finch’s key differentiators are: 

Finch continuously evolves to provide deeper insights and improved efficiency for Uber’s finance teams. 

One area of planned improvement is to expand integrations with Uber FinTech systems. This will offer a more unified and interactive finance ecosystem for automated financial analysis, reporting, forecasting, and budgeting.

Another area of improvement is to support the executive leadership user experience. Executive leaders like Dara (Uber CEO) and Prashanth (Uber CFO) need fast and accurate data for decision-making. While Finch is good enough for day-to-day operations for most Finance teams at Uber, like all generative AI technologies at this point, it’s not 100% reliable and is prone to hallucination. Beyond our ongoing focus on improving accuracy, we also want to integrate human-in-the-loop validation upon request for our leaders. For example, Finch will have a button next to its response for “Request validation,” which will send the context and Finch-generated response to one of our subject matter experts to verify or correct the response. Once that expert ensures the answer is correct, Finch will inform the leader. This “on-demand” human-in-the-loop will allow leaders to use Finch freely without including a human for less critical questions. Still, having a human verify accuracy will allow them to use Finch confidently for critical questions.

We also want to update Finch to handle more user intents with more agents and toolkits. Our goal is to make Finch a one-stop-shop for Finance users at Uber, so we’re continuing to build more agents and agent toolkits to address those needs. 

Finch is transforming how Uber’s finance teams interact with data, making financial insights more accessible, secure, and real-time. By combining generative AI, self-querying agents, and enterprise-grade security, Finch eliminates the inefficiencies of traditional data retrieval, empowering teams to focus on strategy rather than SQL.

As Finch continues to evolve, its capabilities will expand to deliver even more personalized financial insights. The future of financial intelligence at Uber is here, and Finch is at the forefront.

With Finch, we aren’t just answering questions, we’re reshaping financial decision-making one query at a time.

Stay up to date with the latest from Uber Engineering—follow us on LinkedIn for our newest blog posts and insights.