What Is Mosaic AI? Databricks' GenAI and Agent Platform
Mosaic AI is the GenAI and machine learning platform that lives inside Databricks. If you already run data engineering, analytics, or BI on the Databricks lakehouse, Mosaic AI is the layer that lets the same teams serve models, fine-tune open-source LLMs, build agents, and run retrieval over enterprise data, all governed by the same access controls you already use for your tables. It is the part of Databricks aimed squarely at the move from dashboards to agents.
This breakdown is written for practitioners deciding whether Mosaic AI belongs in their stack. We cover where it came from, the six components you will actually touch, the governance foundation that holds it together, and the customer ROI numbers Databricks publishes, which we label plainly as vendor-reported so you can weigh them honestly.
What Is Mosaic AI?
Mosaic AI is Databricks' answer to a single question: how do you build, serve, and govern AI systems on the same platform that already holds your data? Rather than ship your data out to a separate model vendor and a separate vector database and a separate evaluation tool, Mosaic AI puts those capabilities next to your lakehouse tables. The model serving, the training jobs, the agents, and the retrieval index all read from and write to data your team already controls.
The pitch matters because the alternative is a sprawl of disconnected tools. Most teams that start building with LLMs end up stitching together a model API, an orchestration framework, a vector store, a tracing tool, and a governance layer bolted on after the fact. Mosaic AI collapses that into one platform where lineage and access control are not an afterthought. Databricks goes further and markets it as the only unified platform for agent systems, which is a vendor claim worth testing against your own requirements rather than taking at face value.
Practitioner note: Mosaic AI is not a model you call. It is the platform you build on. The value is consolidation: one governance perimeter, one place to trace a bad answer back through retrieval and prompts, and one set of credentials. The tradeoff is that you are committing to the Databricks ecosystem to get it.
From MosaicML to Mosaic AI
Mosaic AI did not start inside Databricks. It traces directly to the $1.4 billion acquisition of MosaicML in June 2023. MosaicML had built a reputation for efficient large-model training, the unglamorous engineering work of pretraining and fine-tuning models without burning a fortune in compute. Databricks bought that capability and folded it into its broader data platform.
The result is a platform that did not bolt generative AI onto a warehouse as a marketing exercise. The training muscle came from a team whose entire product was making model training cheaper and more reliable. That heritage shows up in the Mosaic AI Training component, where pretraining a custom LLM and fine-tuning an open-source one are first-class workflows rather than tutorials.
Understanding this origin helps you read the platform correctly. Mosaic AI leans toward teams that want to own their models, not just call someone else's API. If your goal is to fine-tune on proprietary data and keep that data inside your own governance perimeter, the MosaicML lineage is why Mosaic AI handles that path well.
The Six Components
Mosaic AI is best understood as six components that cover the full lifecycle of an AI system, from training a model to deploying an agent and proving it works. You rarely use all six at once, but most production agent projects touch four or five of them. Here is what each one does.
The two pieces practitioners tend to undervalue are Agent Evaluation and the Unity AI Gateway. Evaluation with AI judges is what separates a demo from a system you can put in front of customers, because it catches the silent quality regressions that creep in when you change a prompt or swap a model. The Unity AI Gateway is an LLM-gateway-style control plane that routes and governs every model and MCP call, so policy and rate limits are enforced in one place instead of scattered across application code.
Built on MLflow and Unity Catalog
Mosaic AI does not invent its governance from scratch. It rests on two open foundations that Databricks also maintains in the open-source community, which is part of why the platform feels coherent rather than bolted together.
Why does this matter to a practitioner? Because the governance is the same whether you are querying a table, serving a fine-tuned model, or running an agent. The access control that protects your customer data also protects the model that reads it and the agent that acts on it. You do not configure three separate permission systems and hope they agree.
MLflow is also the reason traceability is built in rather than retrofitted. When an agent gives a bad answer, the model registry and experiment tracking give you a path back to which model version, which prompt, and which retrieval results produced it. For anyone who has tried to debug an LLM system without that lineage, this is the difference between an afternoon and a week.
Practitioner note: Both foundations are open source, which lowers lock-in at the governance layer specifically. Unity Catalog being able to govern external models means you are not forced to serve everything inside Databricks to keep a single audit trail. Treat that as a real architectural advantage when you compare platforms.
Customer ROI, Vendor-Reported
Databricks publishes several customer outcomes for Mosaic AI. We are repeating them here because they are part of how the platform is positioned, but read them with the right frame: these are vendor-reported customer claims, not independently audited results. They tell you what is possible for a well-resourced customer, not what you should budget for.
The honest way to use numbers like these is as evidence that the platform can support serious production workloads, paired with healthy skepticism about whether your team, your data, and your use case will land in the same range. A 10x cost reduction usually compares a tuned Mosaic AI deployment against a previous architecture that had its own inefficiencies. Your baseline is different, so your result will be too. Treat these as directional, run your own pilot, and measure against the system you actually run today.
Who Mosaic AI Is For
Mosaic AI is not the right starting point for everyone. It pays off when the data is already on Databricks and the team needs to govern AI the way it governs data. Here is who tends to get the most out of it.
What to Watch Out For
No platform is free of tradeoffs. Mosaic AI's strengths come with commitments you should weigh before you build on it.
Frequently Asked Questions
Is Mosaic AI a separate product from Databricks?
No. Mosaic AI is the GenAI and machine learning platform inside Databricks, not a standalone purchase. It uses the same lakehouse data, the same Unity Catalog governance, and the same workspace as the rest of Databricks. You reach it through Databricks rather than as a separate service.
Where did Mosaic AI come from?
It originates from the $1.4 billion MosaicML acquisition Databricks completed in June 2023. MosaicML specialized in efficient large-model training, and that capability became the Mosaic AI Training component. The rest of the platform grew around it.
What is the difference between Agent Bricks and the Agent Framework?
Both are part of Mosaic AI's agent tooling. Agent Bricks and the Agent Framework let you build, deploy, and evaluate agents grounded in enterprise data, and they include Genie Code. Treat them as the agent-building surface of Mosaic AI rather than two competing products.
How does Mosaic AI handle retrieval-augmented generation?
Through AI and Vector Search, a vector database with real-time sync. You embed your enterprise data, keep the index synced as the underlying data changes, and let agents retrieve relevant context at query time. Because it is part of Mosaic AI, the same governance applies to the retrieval index as to your tables.
Can Mosaic AI govern models hosted outside Databricks?
Yes. Unity Catalog, which underpins Mosaic AI governance, can govern models hosted outside Databricks, and the Unity AI Gateway governs every LLM and MCP endpoint. That means you can keep a single audit trail and policy layer even when some models live elsewhere.