Enterprises already using Databricks
Keeps AI development close to existing lakehouse data, permissions, pipelines, and operations.
Independent tool overview
Databricks Mosaic AI at a glance
Databricks Mosaic AI is an enterprise platform for building, evaluating, deploying, governing, and monitoring machine-learning and generative-AI systems on the Databricks lakehouse.
Visit the official Databricks Mosaic AI site ↗
- Product type
- Enterprise AI and machine-learning platform
- Best for
- Governed AI systems built on enterprise data
- Core platform
- Databricks Data Intelligence Platform
- Model access
- Hosted foundation models, custom models, and external providers
- Deployment
- Databricks Apps, batch inference, and Model Serving endpoints
- Evaluation
- MLflow tracing, evaluation, prompt management, and monitoring
- Governance
- Unity Catalog and AI Gateway
- Pricing model
- Usage-based Databricks Units plus applicable cloud costs
Overview
What Databricks Mosaic AI is
Databricks Mosaic AI is best understood as an enterprise AI platform, not a single chatbot or model. It brings conventional machine learning, foundation-model access, retrieval, agents, model serving, evaluation, monitoring, and governance into the same environment as a company's data.
Teams can prototype in notebooks or AI Playground, connect governed tools and data through Unity Catalog, build retrieval-augmented generation systems with Vector Search, and deploy applications through Databricks Apps or real-time Model Serving endpoints. The platform supports Databricks-hosted models as well as external model providers.
MLflow provides prompt management, tracing, evaluation, experiment tracking, and production monitoring. Unity Catalog governs data, models, vector indexes, functions, and other AI assets, while AI Gateway adds centralized endpoint controls such as usage tracking, budgets, rate limits, and guardrails.
The main tradeoff is complexity. Mosaic AI is most compelling when an organization already has substantial data and governance needs on Databricks. Small teams that only need a hosted model API or a simple chatbot builder can usually launch faster with a narrower product.
Use cases
Who Databricks Mosaic AI is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Governed RAG and agent systems
Useful when applications must retrieve private data, call approved tools, and preserve centralized access controls.
Joint data and ML teams
Supports data preparation, feature engineering, classical ML, generative AI, deployment, and monitoring within one platform.
Teams comparing multiple models
Offers hosted, custom, and external models behind consistent development and serving workflows.
Production AI with evaluation requirements
MLflow tracing, evaluation datasets, expert review, and monitoring help teams inspect quality before and after launch.
Capabilities
Core Databricks Mosaic AI features
Foundation Model APIs
Access supported foundation models through OpenAI-compatible REST APIs, Databricks SDKs, SQL functions, and pay-per-token or provisioned endpoints.
Custom model serving
Deploy custom models and agent applications behind managed, autoscaling real-time endpoints.
Agent development
Build tool-calling, retrieval, and multi-agent applications using notebooks, AI Playground, Agent Framework, and Databricks Apps.
Governed tools
Expose Unity Catalog functions, MCP servers, data queries, or external APIs as controlled tools for agents.
Vector Search
Create managed vector indexes for semantic retrieval over enterprise documents and other unstructured data.
Genie and SQL access
Let applications work with governed structured data through natural-language data interfaces and SQL-based tools.
MLflow tracing
Capture prompts, retrieval steps, tool calls, model responses, latency, and other execution details for debugging.
Evaluation and expert review
Evaluate applications against representative datasets and collect feedback from subject-matter experts through Review App workflows.
Prompt management
Version, test, optimize, and deploy prompts with MLflow rather than scattering prompt text through application code.
AI Gateway
Centralize model endpoint governance with usage tracking, budgets, rate limits, guardrails, and inference logging.
Unity Catalog governance
Apply permissions, lineage, discovery, and audit controls to data, models, features, functions, and AI assets.
Classical machine learning
Use notebooks, AutoML, feature engineering, experiment tracking, and common ML libraries alongside generative-AI workflows.
Process
How the Databricks Mosaic AI workflow works
Step 1
Define the use case and risk level
Specify the user, required actions, acceptable failure modes, data sensitivity, latency, and quality targets before choosing technology.
Step 2
Confirm cloud and region support
Check that required models, serving modes, and AI features are available in the workspace's cloud and region.
Step 3
Prepare governed data and tools
Register data, models, functions, vector indexes, and tool permissions in Unity Catalog using least-privilege access.
Step 4
Build a representative prototype
Use realistic documents, queries, APIs, and user tasks rather than a polished demo with hand-picked examples.
Step 5
Compare models and retrieval designs
Measure quality, latency, and cost across model choices, prompts, chunking strategies, and tool configurations.
Step 6
Trace the complete application
Instrument prompts, retrieval, tool calls, outputs, errors, token use, and latency with MLflow.
Step 7
Create an evaluation set
Combine curated cases, production-like examples, edge cases, and subject-matter expert judgments.
Step 8
Harden the preproduction system
Test permissions, prompt injection resistance, sensitive-data controls, budgets, rate limits, fallbacks, and human review paths.
Step 9
Deploy and monitor continuously
Track quality, feedback, latency, errors, token consumption, endpoint utilization, and billing after launch.
Cost
Databricks Mosaic AI pricing and free plan
Databricks does not sell Mosaic AI as one flat monthly subscription. Workloads consume Databricks Units, and the dollar cost varies by cloud, region, workspace agreement, compute type, model, tokens, endpoint capacity, and supporting data workloads. Use the Databricks pricing calculator and billing system tables for an actual estimate.
Foundation models: pay per token
Usage based
Designed for development, exploration, and variable traffic.
- Billed according to model input and output usage
- Model availability and rates vary by cloud and region
- Associated application and data workloads may add separate usage
Reserved provisioned throughput
Committed capacity
Reserved model capacity for predictable, business-critical workloads.
- Available in one- or three-month reservations for supported models
- Reserved capacity is billed for the full term whether fully used or not
- Spillover can add pay-per-token charges where supported
Custom model and agent serving
DBUs per active compute
Managed CPU or GPU endpoints for custom models and agent applications.
- Cost depends on selected workload size and active serving time
- Autoscaling and scale-to-zero settings affect spend and cold-start behavior
- Route-optimized serving and regional availability should be checked before production
Vector Search
Capacity based
Managed vector endpoints and indexes for retrieval workloads.
- Cost depends on endpoint and index capacity
- Embedding generation and source data processing can add usage
- Retrieval load and index design materially affect total cost
AI Gateway and evaluation
Feature and request usage
Governance, logging, guardrails, and model-assisted evaluation can create additional usage.
- Inference tables store request and response logs
- Guardrails and judge-model evaluations may invoke extra model work
- Retention and analytics over logs also consume storage or compute
Supporting Databricks workloads
Additional usage
Data engineering, SQL, notebooks, Apps, storage, networking, and other platform work are priced separately.
- Total cost is broader than the model endpoint alone
- Contracted Databricks Unit rates determine the final dollar amount
- System billing tables, budgets, and usage dashboards should be configured before launch
Pricing checked . Check current pricing at the source ↗
Assessment
Databricks Mosaic AI strengths and limitations
Where it stands out
- Combines enterprise data, AI development, deployment, monitoring, and governance in one platform
- Supports classical machine learning and generative AI rather than forcing teams into separate stacks
- Offers multiple model paths, including hosted foundation models, custom models, and external providers
- Unity Catalog provides a consistent permission and governance layer across data and AI assets
- MLflow tracing and evaluation expose application behavior beyond the final model response
- Managed serving, vector retrieval, Apps, and AI Gateway reduce infrastructure assembly work
- SQL, Python, REST, SDK, and OpenAI-compatible interfaces support different teams and workloads
- Billing system tables and usage controls make production cost analysis possible at workspace scale
What to consider
- It is a broad enterprise platform with a steeper learning curve than a standalone model API or no-code chatbot builder
- There is no single Mosaic AI monthly price; accurate budgeting requires modeling several usage components
- Features, supported models, and serving capabilities vary by cloud and region
- Teams need Databricks, data engineering, ML, and production operations expertise to use the platform well
- Reserved provisioned throughput creates committed spend even when traffic is below capacity
- Autoscaling and scale-to-zero can reduce idle cost but may introduce cold-start latency
- Evaluation scores and model judges do not replace expert review for high-stakes decisions
- Guardrails reduce risk but cannot guarantee that agents will never reveal sensitive data or take an unsafe action
- Tool-calling agents require careful permission boundaries because a model can invoke real systems
- Retrieval quality depends on source quality, chunking, metadata, embeddings, access controls, and evaluation
- Using external model providers can introduce separate data-processing terms, availability limits, and costs
- Consolidating workloads on Databricks can increase platform coupling and make migrations more involved
- Request and response logging requires deliberate privacy, retention, and access policies
- It is usually excessive for small teams that only need a simple hosted inference endpoint
Compare
Databricks Mosaic AI alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Together AI
A more focused cloud for foundation-model inference, fine-tuning, and dedicated endpoints without adopting the full Databricks data platform.
Explore Together AI →Coding
Replicate
A simpler developer experience for running and deploying individual AI models through an API.
Explore Replicate →Consumer
Amazon Nova
An AWS-native model family and agent ecosystem to consider when workloads and governance are already centered on Amazon Bedrock.
Explore Amazon Nova →Questions
Databricks Mosaic AI FAQs
What is Databricks Mosaic AI?
It is Databricks' collection of capabilities for building, evaluating, deploying, governing, and monitoring machine-learning and generative-AI applications on enterprise data.
Is Mosaic AI a single AI model?
No. It is a platform layer that can use Databricks-hosted foundation models, custom models, and supported external model providers.
Did Databricks discontinue Mosaic AI?
No. Mosaic AI remains an umbrella product name, although current documentation often uses shorter capability names such as Model Serving, AI Search, Agent Framework, MLflow, and AI Gateway.
Can Mosaic AI build AI agents?
Yes. Teams can build retrieval, tool-calling, and multi-agent systems, connect governed functions or MCP tools, trace them with MLflow, and deploy them through Apps or Model Serving.
Can it work with private company data?
Yes. That is a primary use case. Unity Catalog permissions, governed retrieval, data pipelines, and audit controls help teams restrict which data and tools an application can access.
Does it support retrieval-augmented generation?
Yes. Vector Search, governed data, model endpoints, tools, and MLflow evaluation can be combined into production RAG systems.
How much does Databricks Mosaic AI cost?
There is no universal flat fee. Costs depend on contracted Databricks Unit rates and usage across models, serving, vector search, evaluation, data processing, storage, and other platform services.
What is the difference between pay-per-token and provisioned throughput?
Pay-per-token is flexible for exploration or variable traffic. Provisioned throughput reserves capacity for predictable production demand and is billed for the commitment term.
Can it use OpenAI-compatible APIs?
Yes. Databricks Foundation Model APIs expose an OpenAI-compatible request format for supported pay-per-token and provisioned endpoints.
What does MLflow do in Mosaic AI?
MLflow supports prompt management, application tracing, experiment tracking, evaluation, expert review workflows, and monitoring across development and production.
What is Unity Catalog's role?
Unity Catalog governs access to data, models, functions, features, vector indexes, and other AI assets so applications inherit centralized permissions and auditability.
Who should choose Mosaic AI?
It fits organizations with substantial data, governance, deployment, and evaluation needs—especially teams already operating on Databricks. A smaller API platform is often easier for a narrow inference use case.
Bottom line
Our Databricks Mosaic AI verdict
Databricks Mosaic AI is a strong choice for organizations that want governed AI applications to sit directly on top of their lakehouse data and production operations. Its combination of model choice, agents, retrieval, MLflow evaluation, Unity Catalog, and managed serving is unusually complete. The payoff is strongest at enterprise scale; smaller projects should weigh that integration against the platform's cost complexity and operational learning curve.
Visit Databricks Mosaic AI website ↗Similar Data Analysis Tools
View all Data Analysis AI tools →
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.