AI Assistant

Lakemeter includes a built-in AI assistant powered by Claude that can answer pricing questions, configure workloads, and generate entire estimates from a natural language description.

The AI Assistant guide — two modes, five specialized tools, and conversation examples.

The five tools the AI assistant uses to propose workloads, analyze estimates, and answer questions.

Animated: asking the AI assistant to add a workload — it proposes a configuration you can confirm, edit, or cancel.

How It Works

The assistant appears as a chat panel on the right side of the screen. It has access to Lakemeter's full pricing database and can create workloads directly in your estimate — no manual form-filling required.

Behind the scenes, the assistant uses five specialized tools:

Tool	What it does
Propose Workload	Creates a fully configured workload (any of the 12 types) and adds it to your estimate after you confirm
Ask Clarifying Questions	Asks follow-up questions when your request is ambiguous before proposing anything
Get Estimate Summary	Reads your current estimate to understand what you already have
Analyze Estimate	Reviews your workloads and suggests cost optimizations or configuration improvements
Propose GenAI Architecture	Suggests a complete multi-workload stack for common AI patterns (RAG, document processing, etc.)

Two Modes

Home Mode (Q&A Only)

When you're on the home page with no estimate open, the assistant operates in question-and-answer mode. It can answer general Databricks pricing questions but cannot create or modify workloads.

Good questions for Home Mode:

• "What's the difference between Classic and Serverless compute?"
• "How does DLT pricing work across editions?"
• "Which GPU types are available for Model Serving on AWS?"

Estimate Mode (Full Capabilities)

When you have an estimate open, the assistant gains full context — it can see your cloud provider, region, pricing tier, and all existing workloads. This unlocks workload creation, optimization analysis, and architecture suggestions.

Conversation Examples

Example 1: Creating a Single Workload

You: I need a Spark ETL job that runs 5 times a day, takes about 45 minutes each run, with 8 i3.xlarge workers on AWS. Use Photon.

Assistant: I'll set that up for you. Here's what I'm proposing:

Setting	Value
Type	Jobs Compute
Mode	Classic
Photon	Enabled
Driver	i3.xlarge
Workers	8 x i3.xlarge
Runs/Day	5
Avg Runtime	45 min
Days/Month	22 (business days)

Estimated monthly cost: ~$X,XXX

[Confirm] [Edit] [Cancel]

When you click Confirm, the workload is created in your estimate with all fields populated. Click Edit to adjust any values before confirming.

Example 2: Generating a Complete Estimate

You: I'm building a data platform for a mid-size company. We need daily ETL pipelines, a BI analytics warehouse for 20 analysts, and a real-time ML inference endpoint. We're on AWS us-east-1, Premium tier.

Assistant: I'll design a complete platform for you. Let me propose three workloads:

1. Daily ETL Pipeline — Jobs Serverless with Photon for your ingestion and transformation
2. BI Analytics Warehouse — DBSQL Serverless (Medium) sized for 20 concurrent analysts
3. ML Inference Endpoint — Model Serving with GPU for real-time predictions

[The assistant proposes each workload one at a time for your confirmation]

Example 3: Optimizing an Existing Estimate

You: How can I reduce costs on this estimate?

Assistant: Looking at your current estimate, I see a few opportunities:

1. Jobs workload "Nightly Ingest" — You're using on-demand Classic compute. Switching to Serverless would eliminate VM costs and you'd only pay for actual compute time, saving ~30% for bursty workloads.
2. DBSQL warehouse "Analyst Pool" — Your Large warehouse runs 730 hours/month (24/7). If analysts only work business hours, reducing to ~200 hours/month saves ~70%.
3. Model Serving "Recommender" — Consider whether you need GPU Medium or if GPU Small would handle your throughput.

Example 4: Getting a GenAI Architecture

You: I want to build a RAG chatbot for our internal documentation.

Assistant: Here's a recommended GenAI stack for a RAG chatbot:

1. Vector Search (Storage Optimized) — Store and query document embeddings
2. Model Serving (GPU) — Host the embedding model for document ingestion
3. FMAPI (Claude or GPT) — Generate responses using retrieved context
4. Jobs — Scheduled pipeline to re-index new documents

Shall I create all four workloads? I'll configure each one with sensible defaults for a documentation chatbot use case.

Applying AI-Generated Workloads

When the assistant proposes a workload:

1. Review — The proposed configuration appears in the chat with all settings visible
2. Confirm — Click to accept and create the workload in your estimate
3. Edit — Modify any values before confirming (opens the workload form pre-filled)
4. Cancel — Reject the proposal without creating anything

For multi-workload proposals (like a full platform estimate), the assistant proposes each workload sequentially. You confirm or reject each one individually.

Conversation Management

• The assistant maintains conversation history within a session. When the history grows beyond 25 messages, it is trimmed to the 15 most recent messages to keep responses fast and relevant.
• Click Clear to reset the conversation and start fresh
• History does not persist across browser sessions

tip

Be specific in your requests. Instead of "add a warehouse," say "add a DBSQL Serverless Medium warehouse running 8 hours per business day for our analytics team." The more detail you provide, the more accurate the proposed configuration will be.

caution

AI responses are generated by a language model and may contain inaccuracies. Always verify pricing details against the official Databricks pricing page before making procurement decisions. The assistant does not perform live pricing lookups — it uses Lakemeter's bundled pricing data.