Databricks Mosaic AI
The AI layer inside Databricks: Agent Bricks, model serving, vector search and MLflow, all resolving back to Unity Catalog governed data.
Overview
Mosaic AI is the name Databricks uses for everything AI-shaped sitting on top of the lakehouse. In practice that means half a dozen products stitched into one control plane: Agent Bricks for building agents, Model Serving for hosting them, Mosaic AI Vector Search for retrieval, Unity AI Gateway for policy and rate limiting, Document Intelligence for parsing files at scale and Managed MLflow for tracing and evaluation. The design bet is that governance and data locality matter more than which model you pick, so every component resolves back to Unity Catalog.
Agent Bricks is the piece that changed the workflow. You describe the task in plain language, point it at your tables, and it generates synthetic evaluation data, selects an approach and tunes toward a cost and quality target. It covers the four patterns most enterprise agents actually fall into: structured extraction from documents, knowledge assistants over internal corpora, custom text transformation and multi-agent orchestration. Faster than assembling the equivalent from LangGraph plus a hand-built eval set. Harder to debug when the answer is wrong.
Who should skip this: anyone whose data lives in Snowflake or BigQuery. Mosaic AI is a reason to be on Databricks rather than a product that stands alone, and the pricing assumes you are already committed. Teams that are committed get one lineage graph covering the raw table, the vector index, the model call and the eval result. For a manufacturing or logistics operation with governed data and a real audit requirement, that single property tends to decide the purchase.
Key Features
- ✓ Agent Bricks takes a plain-language task description plus a pointer at your tables, then generates synthetic evaluation data and tunes the agent against a cost and quality target you set
- ✓ Unity AI Gateway applies rate limits, PII guardrails and usage logging in front of every model call, including external OpenAI and Anthropic endpoints
- ✓ Mosaic AI Vector Search keeps indexes synced to Delta tables, so RAG retrieval does not silently drift from the source of truth
- ✓ Managed MLflow 3 handles prompt versioning, trace capture and LLM-as-judge scorers in the same workspace as the training runs
- ✓ Model Serving hosts custom fine-tunes on hourly GPU endpoints alongside pay-per-token Foundation Model APIs
- ✓ Document Intelligence parses PDFs and scanned files at warehouse scale instead of routing through a separate OCR service
Where it holds
- • Governance is not bolted on afterwards. Unity Catalog lineage covers the raw table, the vector index, the model call and the eval result in one graph, which is what actually gets an agent through an enterprise security review.
- • Agent Bricks removes most of the eval-harness plumbing that eats the first month of any serious agent project.
- • Runs on AWS, Azure and GCP with the same control plane, which matters for anyone with data residency constraints across regions.
- • MLflow is open source, so the tracing format stays portable if you ever leave.
Where it breaks
- • Cost is the standing complaint and the DBU rate is only half of it. Cloud compute lands on top, and teams routinely discover five figures a month going to one agent workflow.
- • Nothing here is worth adopting on its own. If your warehouse is Snowflake or BigQuery, this is a reason to migrate, not a product you can buy.
- • Agent Bricks hides a lot of the pipeline. When output quality drops there is less to grab onto than with a hand-written DSPy or LangGraph loop.
- • GPU serving capacity varies by region, which bites when your data residency is fixed and the GPUs are somewhere else.
My Take
Around $0.07 per DBU, software only, with AWS or Azure compute billed separately: that arithmetic is how teams end up in the $5,000 to $15,000 a month range for a single production agent workflow. What the money buys is something the standalone eval vendors cannot sell at any price, namely agent traces, vector indexes and model calls sitting inside one Unity Catalog lineage graph, so a security reviewer can see exactly what touched what. Agent Bricks is the real shift since the 2025 Data + AI Summit, and describing a task in English while the platform generates the eval set beats hand-rolling DSPy. The tradeoff is opacity when quality drops.
Quick Info
- Pricing:
- paid
- Openness:
- Proprietary
- Starting at:
- Consumption-based on DBUs. Mid-2026 list rates put Mosaic AI compute around $0.07/DBU and Model Training around $0.65/DBU, with GPU Model Serving roughly 10 to 630 DBU/hr and Vector Search 4.00 or 18.29 DBU/hr depending on tier. Foundation Model APIs bill in DBUs per million tokens. Critical detail: the DBU is the software charge only, the underlying AWS, Azure or GCP compute is billed separately on top. A production agent workflow commonly lands between $5,000 and $15,000 a month.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Databricks Mosaic AI holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Databricks Mosaic AI
Huawei Cloud ModelArts
paidHuawei Cloud's full-lifecycle AI platform. Training and serving on Ascend NPUs through CANN, with Lite Server, Lite Cluster and a MaaS layer. Proprietary and quote-priced.
Weights & Biases
freemiumExperiment tracking and model registry for ML training, plus Weave for tracing and evaluating production LLM agents. CoreWeave-owned since May 2025.