Databricks Mosaic AI is a unified artificial intelligence platform for enterprises launched by Databricks, with the core goal of helping enterprises build, deploy, evaluate, and govern AI agent systems based on their own data.
Main Uses
This platform is used to build and operate AI agent systems. A so-called agent system means that an enterprise can build and run one or more agents to perform complex tasks by combining multiple interacting components. It is not limited to a single model, but instead integrates large language models, classic machine learning models, enterprise data, and tools to efficiently achieve specific goals, and has built-in evaluation techniques and governance mechanisms to ensure that the system is delivered with high quality according to established goals and that all components are in a governed state.
Main Capabilities
- Build agents based on enterprise data: Quickly develop agents across multiple systems and based on enterprise data, using any model from classic machine learning to generative AI.
- Custom evaluation: Provides built-in agent evaluation capabilities, supporting any AI model. Measure agent output quality through AI judges, evaluate remediation effectiveness, and quickly redeploy; can identify production issues, analyze root causes, and take corrective actions.
- End-to-end governance: Provides end-to-end governance for agents, implements guardrails for all models, automated access control, sets rate limits, and tracks data lineage across the entire workflow.
- Toolset: Includes Agent Bricks (build agents based on enterprise data, using synthetic data, custom evaluation, and automatic tuning to optimize quality and cost), Document Intelligence (parse, extract, and classify enterprise documents at scale), AI Gateway (access, security, and cost control plane across models, agents, MCP, tools, and coding agents), AI Search (a high-performance vector database with real-time synchronization of source data), Agent Framework and Agent Evaluation (build production-grade agents and ensure output quality), Model Serving (unified deployment of agents, generative AI, and classic machine learning models), Model Training (fine-tune open-source LLMs, pretrain custom LLMs, or build classic ML models), Databricks Notebooks (real-time collaboration), Managed MLflow (enterprise-grade MLOps/LLMOps), Data Quality Monitoring (detect anomalies, track freshness).
How to use
- Data preparation: Supports ingesting any data type and orchestrating jobs to prepare data for GenAI or ML applications; simplifies feature engineering with built-in governance, and uses Databricks AI Search to create vector indexes for RAG.
- Model management: Manage models such as OpenAI GPT-4, Meta Llama 3, and Anthropic Claude 3 through Agent Bricks AI Gateway, and can further fine-tune or build custom models; can write, publish, and share AI tools to retrieve unstructured data, execute code, or communicate with remote services.
- Deployment: Deploy agents, chains, GenAI models, and classic ML models through Model Serving; MLflow covers the entire AI lifecycle from training to deployment.
- Evaluation and improvement: Use MLflow's built-in evaluation, with AI-assisted judging to score responses, and allow human experts to provide peer feedback; after discovering issues, you can trace root causes, evaluate fixes, and quickly redeploy.
- Governance: Implement governance and guardrails for all AI models through Unity Catalog (including models hosted outside Databricks), automatically enforce access control, rate limiting, prevent harmful content, and track lineage from data to deployment.
Applicable scenarios
Databricks Mosaic AI is aimed at enterprises that need to build agent systems based on their own data rather than general-purpose AI models, and is suitable for multiple industries such as finance, technology, and media. Cases mentioned in the materials include: FactSet built a text-to-code knowledge agent to answer natural language questions for non-technical users, improving accuracy by 44%; Block built an AI agent system to automate seller operations (such as generating customized menus), achieving $10 million in productivity gains; Intercontinental Exchange (ICE) built an agent system to securely use proprietary financial data to answer customer questions, achieving 96% accuracy; Comcast built personalized viewing experiences, increased engagement through intelligent voice commands, and reduced ML costs by 10x.
Limitations and fees
The materials show that Databricks offers a 14-day free trial, up to $400 in free credits, and free GenAI training courses and an introductory guide to AI agents. The page mentions that pricing details can be viewed on the Databricks pricing page, which covers foundation model serving pricing; pricing for specific use cases can be obtained by contacting the sales team. Apart from the above information, the materials do not provide further details about specific charges and limitations.
