Product descriptions:
Vale, one of the world’s largest mining companies, operates across nineteen countries and employs more than 60,000 people. As part of a broader digital transformation, Vale’s finance organization set out to become truly data-driven — moving beyond traditional reporting to advanced analytics and AI. Yet many of its most critical processes relied on dense, unstructured documents and manual review, slowing decision-making and limiting insight. To change this, the Finance Data Office team adopted the Databricks Data Intelligence Platform and Agent Bricks to build governed AI assistants. In just days, the team deployed reusable agent frameworks that help finance teams analyze complex documents, boost productivity and unlock new forms of financial analysis that were previously impossible.
Turning finance into a data-driven organization
The Vale Financial Department manages complex financial, tax, insurance and governance processes across multiple regions and regulatory regimes. While core transactional data resides in SAP and related systems, many of the organization’s most essential finance workflows rely on unstructured PDFs — including hundreds of insurance policies, internal control documents and thousands of pages of government tax regulations. Historically, answering even a single question could mean searching through dozens of files, manually comparing clauses and coordinating across multiple teams.
To accelerate decisions and modernize finance, Vale embedded a dedicated data and analytics team directly inside the finance organization. This Finance Data Office combines formal IT and data backgrounds with a deep understanding of financial rules, enabling the team to bridge the gap between business and IT, speak both languages and drive data initiatives aligned with governance. “Our mission is to make finance more data-driven, with advanced analytics and democratized insight,” explained Marcelo Baltar, Manager of the Finance Data Office at Vale. Demand quickly grew, with leaders in treasury and tax requesting AI-powered automation to support “mini digital transformations” in their respective areas.
Early attempts to build AI agents, however, proved difficult. The team collaborated with a dedicated data scientist to develop an agent framework from scratch. Early demos showed strong potential, but translating them into an initial version that users could test and use proved challenging. “We struggled with governance, repeatability and operationalization,” said Henrique Vasconcelos, Finance Data Engineer, at Vale. “We had unstructured data, new LLM tools, and no common framework Vale Finance team needed a way to move from experimentation to scalable, governed agents that could be reused across finance use cases.
Building governed AI assistants with Databricks Agent Bricks
The Finance Data Office turned to Databricks to establish a standardized, production-ready foundation for agentic AI. Using the Databricks Data Intelligence Platform and Agent Bricks, the team assembled AI assistants without building an agent framework from scratch. In a three-day intensive build with Databricks experts, they created a reusable ingestion and RAG pipeline and a working v0 agent that finance users could immediately test — a breakthrough after four months of experimentation with no production outcomes.
At the core of the solution is Agent Bricks Knowledge Assistant, which lets finance users ask natural-language questions grounded in Vale’s internal documents. To handle dense PDFs, the team built an ingestion framework in Databricks that processes each document once, stores it in an efficient binary format, parses it into structured text and writes the results to Delta tables. These parsed outputs feed both Databricks Vector Search indexes for semantic retrieval and structured tables that support downstream analysis. “Now we process each document once and reuse it everywhere,” Henrique noted. “That alone saves us from rework every time an agent changes.”
The Multi-Agent Supervisor orchestrates how each request is handled, deciding when to call the Knowledge Assistant, when to query structured tables and when to combine both for a richer answer. Behind the scenes, the Finance team uses Anthropic Claude Sonnet as the primary model for retrieval-augmented generation, given its strong performance on long, complex documents while retaining the flexibility to switch to other models, such as OpenAI or Llama, as needs evolve. Evaluation workflows and user review loops enable finance teams to validate responses, refine prompts and continually improve quality. “In three days with Agent Bricks, we built a reusable framework and a working agent,” said Henrique. “Now, any new document-based use case can be stood up in a week or two by reusing the same pattern.”
Improving productivity and unlocking new financial insights
With the new framework in place, the Finance team is applying Agent Bricks across several high-impact finance use cases. One production deployment supports insurance analysis, where the team manages more than 430 insurance policies across various regions, each with distinct coverage terms. Previously, answering a question such as whether a flood in a specific country was covered required manually opening and reading through many policies. With Agent Bricks Knowledge Assistant and Multi-Agent Supervisor, insurance teams can now query all 430+ policies at once (80% reduction in time spent reviewing insurance policies), identify relevant coverage, compare overlapping policies and receive suggestions on potential gaps or redundancies — enabling analyses that were either too slow or impossible before.
Another initiative focuses on tax and legal workflows. Governments regularly issue new tax laws, and the tax squad has a backlog of more than 10,000 regulations that must be checked for financial impact and potential penalties. Historically, a single expert read each law, determined whether it affected and then communicated the findings to around 100 people across tax and legal. Vale Finance Data Office is building an AI assistant that consolidates these 10,000+ laws into a single, searchable knowledge base, enabling roughly 100 finance and tax professionals to assess their impact independently rather than waiting for manual distribution, saving approximately 16 hours per month. “The productivity gain here is huge,” said Marcelo. “Instead of one person being a bottleneck, we give everyone self-service access to the same governed knowledge.”
A third use case centers on internal controls and governance policies, where employees across the company often need clarification on internal rules. Today, those questions can require meetings and manual interpretation from control teams. With Agent Bricks, Vale is creating a self-service assistant that surfaces relevant internal policies on demand, reducing ad hoc requests and freeing control teams to focus on higher-value work. In parallel, the finance data team continues to deliver additional AI-driven initiatives — including a depreciation project that has already identified roughly 33 million USD in recoverable value — all underpinned by Databricks as the unified platform.
While some of these agents are still scaling, the impact is already clear. The Finance team has reduced AI agent development time to just a few days for an initial version and roughly one to two weeks to deploy new use cases based on the shared framework. With around 100 finance and tax users now supported by AI assistants, finance teams gain faster access to previously buried insights and the risk of missing critical regulatory or coverage details is reduced. “Databricks has empowered our finance team to accelerate data-driven decisions by providing a unified platform for advanced analytics, machine learning and scalable data processing,” said Marcelo. “It turned AI from isolated experiments into a core part of how finance operates.”