Skip to main content

How to Build an AI Compliance Agent with Agent Bricks

Summary

  • Multi-agent compliance: Agent Bricks enables a Supervisor Agent orchestrating between a Knowledge Assistant (for regulatory PDFs) and a Genie Agent (for structured compliance data) to automate regulatory mapping work.
  • Compliance automation: The multi-agent system bridges unstructured regulatory text with structured risk data, transforming weeks of manual mapping of 50-page PDFs into natural-language conversations.
  • Enterprise governance: All agents built on governed data within Unity Catalog, ensuring that compliance officers can identify gaps between regulations and controls while maintaining enterprise-grade security and access controls.

How to Build an AI Compliance Agent with Agent Bricks

Watch: How to Build an AI Compliance Agent with Agent Bricks
Stop the manual compliance scramble. See how AI can turn complex regulations into actionable data.
In the fast-moving world of financial services, staying compliant often feels like manual work. Compliance teams are forced to act as "human glue"—spending weeks manually mapping 50-page regulatory PDFs to internal spreadsheets and audit findings.
In this demo, we show you how to transform the Databricks Lakehouse into an Active Compliance Officer using Agent Bricks. Watch how a multi-agent system bridges the gap between unstructured regulatory text and structured risk data to turn months of work into a natural-language conversation.
In this video, you’ll see:

  • The Multi-Agent Supervisor: How a Supervisor Agent orchestrates between a Knowledge Assistant (for documents) and a Genie Agent (for structured data).
  • Knowledge Assistant and Genie: How Agent Bricks extracts insights from both structured and unstructured data simultaneously.
  • Enterprise-Grade Governance: How all of this is built on a foundation of governed data within Unity Catalog.

Learn more about Agent Bricks:

CHAPTERS:

Chapters

FAQs

How does the multi-agent system work for compliance?

A Supervisor Agent orchestrates between a Knowledge Assistant (processing PDFs) and a Genie Agent (querying compliance data). When asked about regulation alignment, both work together to join unstructured obligations with structured control evidence.

What can compliance officers accomplish?

Officers can summarize obligations under new regulations, identify control themes, compare controls against regulations, and list affected business units. Work that took weeks now takes minutes.

What data does the Knowledge Assistant process?

The Knowledge Assistant processes unstructured regulatory documents, policy mappings, incident reports, and regulatory frameworks. It extracts and synthesizes information from diverse PDFs to answer complex compliance questions.

How does Unity Catalog governance apply?

All agents operate on governed data, so officers only see data they have permission to access. The system maintains audit trails of queries and findings for enterprise security and governance compliance.

Full transcript

0:00 Challenges: The Compliance "Manual Tax"

Hello everyone. In this video, you are going to see how Agent Brick's turns a lake house into a building compliance officer on Databricks. We are going to turn weeks of manual mapping into a quick natural conversation. It's time to stop drowning in regulations and start using your data to stay compliant. Mine is a headache. Instead of starting with the plumbing, let's look at the results. I'll show you the agent at work first and then we'll break down on how including the multi-agent setup and the Unity Catalog foundation.
So, imagine I'm a compliance officer and a brand new regulation just hit my desk. Usually this is a start of a week-long fire drill, but I need to know the results and our status right now. Let's see how the agent helps me to size up our alignment and identify our biggest gaps in just a few minutes. First of all, you can test your agents in a playground. We'll go to a playground and instead of a generic foundation model, we will choose an agent that I've built prior to this demo.

0:42 Real-world Scenario: New Regulation Hits

So, let's start with a quite generic and obvious question. Summarize our current obligation under this new regulation and identify key control themes. The agent summarizes regulatory text and internal policy mapping into a concise cross-reference summary. If you look through it, one important thing to mention here is that um you can see that agent also suggests how these regulatory frameworks relate to a certain incidents we had in our example organization.
It means this is not a generic assistant, but an application tailored to our data. But let's narrow the focus. Given our current incident management controls and the last four quarters of audit findings, what where we are not aligned to this regulation? To be able to answer this question, the agent should join unstructured obligation with a structured control evidence and flag and create misalignment. If you look through the answers of this question and go a little bit up and to see how which agent this multi-agent supervisor actually called, you will see a two components.
You will see um call to a genie agent for structured data and call to a knowledge assistant for unstructured um requirements, regulation, incident reports, and so on. But let's go deeper into impact. For each gap that um agent identify for us, we would like to list affected business units, inherited risk level, and also uh repeated issues from the past audits. The agent uses historical incidents and findings to prioritize gap and highlight patterns.
Look for instance here. From here, we can shift from diagnosis to guidance and ask agent to propose a remediation action for every gaps that uh it identified. The agent produces a draft remediation plan aligned with internal policy language and past successful fixes. To wrap it up, we can ask uh agent to draw an executive summary for the compliance committee describing our readiness to this regulation as well as next steps.
The agent composes a governance ready overview grounded in the same evidence and evaluations. So, how does it work under the hood? If we go back to an agent and see what it consist of, you can see that our agent is actually an agent breaks multi-agent supervisor, which uses all power of Lake House by combining two other agents together. A compliance um compliance assistant and which handles regulatory tax policies and guidance, as well as an Genear agent, which responds to natural language questions over structured data, such as controls, KPIs, and incidents.

3:51 Setup: The Multi-Agent Setup (Multi-Agent Supervisor, Knowledge Assistant & Genie)

A supervisor agent orchestrate both, automatically deciding when to read documents, when to query tables, and how to fuse those results into a single auditable answer. Let's look at both of them, um, one by one. Let's start with Genie. As you can see here, our Genie room contains, uh, a few tables related, for instance, to client KYC transaction table, advise log, and a lot of information about our product, or, for instance, regulatory coverage.
At the same time, if we look at our knowledge assistant, it works with unstructured data. So, you can see here that it has two main sources. One of them is a product governance and marketing guides, and another one is a regulatory and policy library. So, speaking of data, let's look at the data foundation. Let's go to Unity Catalog and select our, catalog and schema. For this demo, the agent uses three main data sources stored and governed in Unity Catalog.
A catalog of a regulatory standards, um, including versions, effective dates, and so on, stored in Unity Catalog volumes. You can see here that those are PDF files, and obviously all of them have different formats. You can also see some here directly in Unity Catalog volumes without leaving the environment. You will also have a bunch of historical incidents and remediation record, which are used to ground recommendations in what has actually worked before in this organization.

5:14 Data Foundation: Protocols & Regulations

As well as we have an operational and risk data tables. This is primarily what we use in, uh, Genie. This data contains information such as KYC profile, transaction monitoring and alerts, control test results, and audit findings. Um we will we will expose them in Genie for a natural language querying. At the end, let's look at the architecture and what we actually built today. So, we have our product governance and marketing guides in a PDF format.
We store them in Unity Catalog volumes. We have our regulatory and policy library, which we also store in Unity Unity Catalog volumes. And we have our risk and compliance database, which are tables in Unity Catalog. For unstructured files, we made them available in Knowledge Assistant. While for tables, we use them in Genie to answer questions to unstructured data. All in all, it was combined in a multi-agent supervisor, which decides which agent to call, which essentially available for a compliance officer to ask question and get consistent answers.

6:24 Architecture Overview

Together, this agent turn complex compliance data into clear, actionable insights to help financial institutions move faster while staying in control. Agent bricks provides the auto optimization, evaluation, and governance needed to run these workloads safely and at enterprise scale. Check it out yourself. >>

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.