What Databricks Mosaic AI does
Databricks Mosaic AI combines model development, serving, AI Functions, agents, vector search, evaluation, monitoring, and Unity Catalog governance on the lakehouse.
Databricks Mosaic AI spans the AI lifecycle on the Data Intelligence Platform. Teams can prepare Delta data, call managed or external models through AI Functions, train and track models with MLflow, build retrieval and agent systems, evaluate responses, deploy real-time or batch endpoints, and monitor inference. Unity Catalog governs tables, models, functions, vector indexes, features, and services alongside lineage and audit information.
Cost is not one universal Mosaic AI price. Workspace tier and cloud affect DBU rates; notebooks, jobs, Lakeflow pipelines, SQL warehouses, serverless processing, model serving, foundation-model input and output tokens, AI Functions, vector search, online stores, storage, network, and underlying cloud resources may meter separately. Preview billing can change. Tag workloads, query system billing tables, model concurrency and scale-to-zero behavior, and confirm rate cards and commitments for the chosen cloud and region.
A governed catalog cannot compensate for bad source data or unchecked code. An AI Function may misclassify every row at scale, a Genie Agent can generate an expensive or semantically wrong query, pipeline retries can multiply tokens, and an endpoint can expose sensitive features if grants or credentials are broad. Use Unity Catalog privileges, row filters and masks, isolate compute identities, validate schemas and expectations, sample model outputs, benchmark generated SQL, cap batch sizes and endpoint budgets, inspect lineage, and require engineering review before writes or downstream decisions.
How Databricks Mosaic AI works
Lakehouse data, models, functions, and features are registered as Unity Catalog objects with explicit privileges and lineage. Engineers invoke AI Functions from SQL, notebooks, Lakeflow pipelines, Workflows, or streaming jobs, develop and evaluate models or agents, and expose approved models through serverless serving endpoints. Genie Agents use selected tables, semantic instructions, examples, and SQL warehouse compute to answer natural-language questions.
How to set up Databricks Mosaic AI
Define the governed lakehouse scope
Inventory catalogs, data classes, owners, lineage, quality SLAs, allowed models, regions, identities, serving paths, workloads, expected tokens, DBUs, and cloud costs.
Configure Unity Catalog and isolation
Create separate catalogs and schemas, grant only required SELECT, MODIFY, EXECUTE and compute permissions, apply masks and filters, and use service principals for production.
Build a small evaluated workflow
Pilot one AI Function, model endpoint, or Genie Agent on versioned data with typed outputs, semantic instructions, example SQL, expected failures, and a labeled test set.
Test quality, security, and economics
Measure SQL correctness, model accuracy, drift, latency, retries, prompt injection, data leakage, permission inheritance, token volume, DBUs, warehouse scans, and concurrency.
Deploy with observable controls
Use CI, tests, approvals, inference and quality monitoring, lineage and audit logs, budgets, endpoint limits, rollback, and human review for writes and consequential outputs.
Databricks Mosaic AI FAQs
How much does Databricks Mosaic AI cost?
Costs vary by cloud, tier, DBUs, serverless workloads, model tokens, serving, AI Functions, vector search, SQL compute, storage, network, and infrastructure. Confirm each applicable SKU.
What does Unity Catalog govern?
It governs securable data and AI objects with privileges, ownership, lineage, auditing, and applicable filters or masks. Correct configuration and least privilege remain the customer's responsibility.
Can AI Functions run in data pipelines?
Yes. Official documentation supports SQL, notebooks, Lakeflow pipelines, Workflows, and streaming contexts, with feature status, model, region, and billing limitations.
Does Genie always generate correct SQL?
No. Authors must curate tables, instructions, semantic metadata, and examples, then benchmark generated SQL and results across representative and adversarial questions.
How can a team control batch AI cost?
Limit rows and prompt size, select task-specific functions when suitable, monitor billing tables by job and endpoint, prevent runaway retries, budget compute, and test on samples first.
Listing reviewed 2026-07-15. Product details and pricing can change; verify important terms on the provider's website.
Related Data engineering AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.