Databricks AI and Mosaic AI Consulting in 2026

What Mosaic AI does, what it costs, where it wins against Snowflake Cortex, and the Unity Catalog discipline that decides whether agents and model serving survive past the proof of concept. Practitioner notes from inside Databricks AI engagements.

Databricks Mosaic AI is the platform Databricks customers are being asked to evaluate in 2026. Foundation Model APIs deliver pay-per-token inference on Llama, DBRX, Mixtral, and Claude. AI Functions expose LLM calls inside Databricks SQL. Vector Search delivers Unity-Catalog-native managed vector indexes. Model Serving hosts custom models in production. The Mosaic AI Agent Framework orchestrates multi-step agents grounded in lakehouse data. All five sit on top of Unity Catalog, which is what makes the platform different from running models on a separate AI provider with separate governance.

This piece is the practitioner read. We have shipped Mosaic AI deployments alongside Snowflake Cortex builds, we have recommended against Mosaic AI when Cortex or a non-warehouse approach was the better answer, and we have helped Databricks customers ship RAG into production with the Unity Catalog discipline that decides whether the deployment lasts. Here is what we tell buyers on the discovery call.

• 5 surfaces Foundation Model APIs + AI Functions + Vector Search + Model Serving + Agent Framework. All five sit on top of Unity Catalog, which means every AI surface respects the same governance, access control, and lineage as the rest of the lakehouse. The architectural reason Databricks customers buy Mosaic AI instead of running models on a separate provider.

Source: Databricks product documentation, May 2026

What Mosaic AI is in 2026

Five surfaces, all sharing Unity Catalog governance and MLflow lineage:

Foundation Model APIs are pay-per-token inference endpoints for Llama 3.1 and 3.3, DBRX, Mixtral, and Claude (through the Anthropic-on-Bedrock integration). The API is OpenAI-compatible, which means most existing LLM application code ports cleanly. The model catalog refreshes regularly.

AI Functions are SQL-callable LLM operations inside Databricks SQL: ai_classify, ai_extract, ai_fix_grammar, ai_summarize, ai_translate, and the rest. These run inside SQL queries against tables. The model used can be specified per function call.

Vector Search is the Unity-Catalog-native managed vector index. Documents, tables, and structured content can be indexed with Databricks-managed embeddings or custom embedding endpoints, then queried with hybrid retrieval. Indexes inherit Unity Catalog access policies.

Model Serving hosts custom and foundation models on provisioned-throughput endpoints. The serving layer integrates with MLflow for tracking, monitoring, and A/B comparison.

Mosaic AI Agent Framework is the orchestration layer for multi-step agentic workflows. Agents combine retrieval, tool calls, and reasoning steps with MLflow tracing and Unity Catalog access controls. The framework also covers evaluation of agent quality, which is the part most rivals leave to custom code.

Databricks Mosaic AI surfaces in 2026

What each surface does, what it costs, and the dominant use case.

Source: Databricks product documentation, May 2026

What Mosaic AI costs in 2026

DBU consumption with separate rates per surface. Four main consumption levers:

Foundation Model APIs are pay-per-token at model-specific rates. Llama 3 family runs cheaper than Claude. Long-context workloads cost more than short-context ones. Most starting workloads land at $1K to $20K monthly in API spend.

Vector Search is metered per query unit and per index storage GB. The reindex cost is the part that surprises buyers. Document corpora that change daily cost meaningfully more than static reference corpora.

Model Serving is metered per serving DBU at provisioned-throughput tiers. Custom model serving for production workloads typically lands at $5K to $40K monthly per endpoint depending on throughput.

AI Functions consumption rolls up into the SQL warehouse DBU bill. Heavy AI Functions usage can double SQL warehouse spend, so cost attribution and per-team budgets matter.

The two cost mistakes we see most. First, sizing on demo data: ai_classify on 1,000 rows runs $5; the same call on 100M rows runs $5,000 and surprises the team. Second, treating Vector Search as a one-time index build: indexes refresh on data changes, and high-churn corpora consume index DBUs continuously.

Mosaic AI first-year cost composition (% of total)

Foundation Model APIs and Vector Search dominate the bill in RAG-heavy deployments; Model Serving dominates when production custom-model inference is the core use case.

• Foundation Model APIs (token-based)
• Vector Search (query + index storage)
• Model Serving (custom model throughput)
• AI Functions (SQL warehouse rollup)
• Agent Framework (combined tools)

Source: Thinklytics Modern Data Platform Practice, Mosaic AI deployment cost analysis, 2024 to 2026

When Mosaic AI is the right answer

Three conditions, any one of which can carry the decision.

Your data already lives in a Databricks lakehouse. If the workload runs on Unity Catalog, the inheritance of access policies, lineage, and audit trail is the deciding architectural fact. Building the equivalent on top of Snowflake Cortex plus a Databricks replication pipeline is real engineering work that Mosaic AI removes.

The team is notebook-first. Mosaic AI surfaces are designed for ML engineers and data scientists who live in notebooks. If the team writes Python and uses MLflow, Mosaic AI is the cleaner path.

The workload requires training, fine-tuning, or custom production serving. Mosaic AI Training and Model Serving are the platform's strongest surfaces against alternatives. For these workloads Snowflake Cortex is not a competitor; the right comparison is to AWS SageMaker, Vertex AI, or a fully self-hosted stack.

When Mosaic AI is not the right answer

Three scenarios where we have recommended against Mosaic AI.

The data does not live in Databricks. If the primary data layer is Snowflake, BigQuery, or a transactional database without a clean Databricks pipeline, integrating Mosaic AI requires replication that erodes the cost and latency advantages.

The team is SQL-first and the workload is BI-augmenting. For natural-language Q and A on dashboards, semantic search on documents, or analytical assistant use cases, Snowflake Cortex on a Snowflake-resident dataset is usually the cleaner path.

The use case is small and simple. Spinning up Unity Catalog, MLflow, Model Serving, and Vector Search for a single-document RAG demo is overkill. For sub-100-document RAG with no compliance bar, a managed RAG service or a simpler stack ships in days, not weeks.

Databricks Mosaic AI vs Snowflake Cortex

• Databricks Mosaic AI. Notebook-first. Wins for training, fine-tuning, custom model serving, and mixed Python + Spark pipelines. Best for ML engineers and data scientists who live in notebooks and use MLflow. Unity Catalog governance is the architectural advantage.
• Snowflake Cortex. SQL-first. Wins for warehouse-grounded RAG, analytical Q and A, in-SQL classification and extraction. Best for analytics engineers and SQL-first teams. Native RBAC and row-access inheritance is the architectural advantage.

Most modern data platform decisions are not Databricks-or-Snowflake anymore; they are which workload goes where. Most organizations with serious AI investments end up with both, one workload-type at a time.

Source: Thinklytics Modern Data Platform Practice, multi-platform engagement portfolio, 2024 to 2026

The implementation pattern that ships

Five phases.

Decision support. Two weeks. Independent assessment of whether Mosaic AI is the right call, Unity Catalog readiness review, DBU consumption projection sampled from real workload patterns, written go or no-go.

Unity Catalog and MLflow foundation. Three to six weeks. Unity Catalog enabled across target workspaces, sensitivity tagging in place, MLflow tracking wired into every model call, DBU budget alerts configured. The phase nobody asks for and every successful Mosaic AI rollout requires.

Wave one build. Four to eight weeks. One use case (RAG application on one corpus, Model Serving for one inference workload, OR AI Functions on one table). Production deployment with monitoring, evaluation, and parallel-run validation.

Wave two and beyond. Six to twelve weeks per wave. Additional use cases, additional surfaces, Mosaic AI Agents when foundation is in place. Each wave is its own SOW.

Mosaic AI Agents (optional). Six to ten weeks. Agent design, tool catalog configuration, guardrails, MLflow tracing, Agent Evaluation framework wired up, escalation patterns when the agent disagrees with the source of truth. Most agent rollbacks happen at month two when agents read across roles; this phase prevents that.

The Mosaic AI implementation pattern that ships

Five phases. Skip phase one or two and you ship a project that gets rolled back at month two.

• Decision support (2 weeks). Independent assessment, Unity Catalog readiness review, DBU consumption projection from real workload samples, written go or no-go.
• Unity Catalog + MLflow foundation (3 to 6 weeks). Unity Catalog enabled across workspaces, sensitivity tagging, MLflow tracking, DBU budget alerts.
• Wave one build (4 to 8 weeks). One use case end to end (RAG, Model Serving, or AI Functions) with monitoring, evaluation, and parallel-run validation.
• Wave two and beyond (6 to 12 weeks per wave). Additional surfaces and use cases. Each wave is its own SOW with its own go or no-go gate.
• Mosaic AI Agents (optional, 6 to 10 weeks). Agent design, tool catalog, guardrails, MLflow tracing, Agent Evaluation framework, escalation patterns.

Source: Thinklytics Modern Data Platform Practice, Mosaic AI delivery model, 2024 to 2026

What good Mosaic AI consulting looks like

Five attributes that separate the firms that ship from the firms that bill.

Unity Catalog first by default. Good firms refuse to start Mosaic AI rollout until Unity Catalog is enabled and properly configured. Bad firms ship features on top of Hive metastore and inherit governance gaps.

DBU-modeling discipline. Good firms sample real workload patterns and write DBU projections into the SOW. Bad firms recommend Mosaic AI without modeling the cost.

Certified team. Databricks Certified ML Engineer Associate and Generative AI Engineer are the relevant credentials. A team without those on the proposed engagement is selling intent, not capability.

Multi-platform reference book. Good firms have shipped Mosaic AI AND Snowflake Cortex AND non-warehouse builds. The recommendation is decided per engagement.

Wave-based delivery with named go or no-go gates. Mosaic AI deployments are decision-heavy at every wave. Good firms ship in waves with explicit checkpoints; bad firms sell a 12-month fixed-bid with no off-ramp.

What good Mosaic AI consulting looks like

Five attributes that separate the firms that ship from the firms that bill.

• Unity Catalog first by default. Mosaic AI rollout starts after Unity Catalog is enabled and configured. No shipping on top of Hive metastore.
• DBU-modeling discipline. DBU projections are sampled from real workload patterns and written into the SOW.
• Databricks Certified team. Certified ML Engineer Associate and Generative AI Engineer credentials on the proposed team.
• Multi-platform reference book. Shipped Mosaic AI, Cortex, and non-warehouse AI builds. Recommendation decided per engagement.
• Wave-based delivery with named gates. Each wave is its own SOW with explicit checkpoints. No 12-month fixed-bid with no off-ramp.

Source: Thinklytics Modern Data Platform Practice, Mosaic AI engagement portfolio, 2024 to 2026

What we do

Thinklytics ships Databricks Mosaic AI consulting as part of the broader Modern Data Platform practice. Our reference book includes Mosaic AI deployments alongside Snowflake Cortex builds and non-warehouse AI architectures. We do not take Databricks, Snowflake, Microsoft, or AWS commissions, so the recommendation is decided per engagement. Most engagements start with a 2-week decision-support phase that produces a written go or no-go, DBU-sized projections, and a Unity Catalog readiness checklist.

If Mosaic AI is on the roadmap, the companion read is our Snowflake Cortex consulting piece. Most modern data platform decisions are not "Snowflake or Databricks" anymore; they are "which workload goes where." Reading both pieces side by side gives the buyer the workload-by-workload decision framework.

Frequently asked questions

What is Databricks Mosaic AI and what does it do?

Mosaic AI is the AI and ML platform inside Databricks, branded after the 2023 acquisition of MosaicML and steadily rebuilt around Unity Catalog. It ships as five surfaces: Foundation Model APIs (pay-per-token Llama, DBRX, Mixtral, and Claude inference), AI Functions (SQL-callable LLM functions inside Databricks SQL), Vector Search (Unity-Catalog-native vector index), Model Serving (production endpoint hosting for custom and foundation models), and Mosaic AI Agent Framework (orchestrated multi-step agents grounded in Unity Catalog). The platform's defining trait is Unity Catalog inheritance: every AI surface respects the same governance, access control, and lineage as the rest of the lakehouse.

What does Mosaic AI cost in 2026?

Mosaic AI is metered on DBU consumption with separate rates for each surface. Foundation Model APIs are pay-per-token at model-specific rates (Llama 3 family runs cheaper than Claude). Model Serving for custom models is metered per serving DBU at provisioned-throughput rates. Vector Search is metered per query unit and per index storage GB. AI Functions roll up into the SQL warehouse DBU bill. Most mid-market Mosaic AI deployments land at $80K to $300K annualized in DBU spend during the first production year. Enterprise deployments with heavy custom model serving and many agents regularly exceed $1M annualized. Customers under-budget Vector Search reindex cost most often.

Databricks Mosaic AI vs Snowflake Cortex, when does Databricks win?

Databricks wins when the team is notebook-first and the workload requires training, fine-tuning, custom model serving, or mixed Python + Spark pipelines. Snowflake Cortex wins when the team is SQL-first and the workload is warehouse-grounded retrieval or analytical Q and A. For RAG on enterprise documents Databricks Vector Search is competitive with Cortex Search; the choice usually follows where the data lives. For training custom foundation models or running production ML serving at scale Databricks remains the deeper platform. Many organizations end up with both. The decision is per workload, not per vendor.

How long does a Mosaic AI implementation take?

Six to ten weeks for a focused first-wave deployment on one use case (typically a RAG application over a Unity-Catalog-governed document corpus, or a Model Serving endpoint for a single inference workload). Four to six months for a multi-surface rollout that includes AI Functions in SQL, Vector Search in production, and one Mosaic AI Agent. Twelve months and longer for enterprise rollouts that include custom foundation-model fine-tuning, multi-tenant agent governance, and integration with downstream applications. The biggest predictor of duration is Unity Catalog readiness. Workspaces still on Hive metastore add 6 to 12 weeks to every engagement.

What governance has to be in place before Mosaic AI rolls out?

Four things. First, Unity Catalog must be enabled and properly configured across the workspaces where Mosaic AI surfaces will run, otherwise lineage, access control, and audit trail are all degraded. Second, MLflow tracking enabled so every model call and Agent step is logged. Third, DBU budget alerts because Foundation Model APIs and Model Serving consume DBUs at much higher rates than typical SQL workloads. Fourth, data classification and sensitivity tagging in Unity Catalog so agents and RAG applications respect access policies. All four are table stakes; ship without any of them and the rollback conversation starts at month two.

Should we use Foundation Model APIs or self-host a model on Model Serving?

Foundation Model APIs for almost every starting workload in 2026. Pay-per-token consumption matches the demand pattern of most AI applications, Databricks manages model lifecycle and security, and the model catalog is broad enough to cover most needs (Llama 3.1 and 3.3, DBRX, Mixtral, Claude on Anthropic-hosted Bedrock). Self-hosting on Model Serving makes sense when the workload has very high sustained QPS where provisioned throughput is cheaper, when latency requirements are stricter than the API can guarantee, or when fine-tuned model weights cannot leave the Databricks workspace for compliance reasons. The crossover for most enterprises is around 100K to 500K daily inference calls sustained.

Do you take Databricks commissions on Mosaic AI deployments?

No. Thinklytics is a Databricks-fluent consulting firm that does not take licensing commissions from Databricks, Snowflake, Microsoft, AWS, or any other vendor. That means we have recommended against Databricks Mosaic AI in cases where Snowflake Cortex or a non-warehouse approach was the better answer, and recommended for Mosaic AI in cases where the lakehouse-grounded training, fine-tuning, and notebook-first workflow dominated. The recommendation is decided per engagement.

What are red flags when evaluating Mosaic AI consulting firms?

Five show up consistently. (1) The proposal recommends Mosaic AI in week one without auditing Unity Catalog readiness. (2) DBU consumption is estimated without sampling real workload patterns. (3) Vector Search is described as plug-and-play with no detail on chunking strategy or reindex cadence. (4) Mosaic AI Agents are scoped before guardrails, MLflow tracking, and Unity Catalog access policies are in place. (5) The proposed team has no Databricks Certified ML Engineer or Generative AI credentials. Any two of these together is a near-certainty for overrun.