AI adoption was accelerating Snowflake burn
More people were using AI on the firm’s data, and the credit budget was coming under pressure. AI-related workloads had grown from roughly 4% to about half of consumption in the audited account. Cortex Agents consumed credits for inference, with warehouse compute billed on top. Agents connecting directly through MCP also ran live queries to discover schemas, sample data and retry SQL.
An audit with IT and Data established the opportunity
Rig started with the IT and Data teams. Together, they reviewed account metering, Cortex request charges and warehouse idle time, then tested the Rig approach in a proof of concept. The audit validated the cost-saving opportunity against observed query costs. Rig built context across the entire warehouse, bringing table relationships, joins and business definitions into one place for agents to use.
Rig became the harness for the data warehouse
Rig wraps the warehouse with the context, tools and access controls an agent needs to work with company data. The agent can understand the available data, use agreed definitions and check its approach before it sends SQL to Snowflake. The warehouse continues to serve live data. Rig handles the surrounding exploration and preparation, so each new question does not require the agent to rediscover the same relationships through paid queries.
The in-process context sandbox does the preparation
The sandbox runs inside Rig’s agent process, alongside the warehouse context. Agents can inspect that context, test query logic against available samples and shape results there before executing live SQL. These steps avoid warehouse execution when the context or samples are sufficient. The final query goes to Snowflake through governed tools. A shared serving warehouse also reduces separate cold starts, which matter because Snowflake charges a 60-second minimum each time a warehouse starts.
Less Snowflake compute per question
The audit put Cortex Agents at about $0.50 per request in token fees, before SQL compute. Rig’s estimated Snowflake serving cost was $0.01 to $0.07 per question, informed by observed account query costs and Rig’s session benchmark. In that multi-tenant benchmark, more than 60% of tool calls handled context or shaping without executing warehouse queries, and 22% of sessions needed no live query. Moving preparation into the sandbox reduces warehouse hits and the compute time spent on exploratory work.
A slower credit burndown at the same adoption
The cost model kept the base workload and AI adoption curve the same on both paths. Serving AI through Rig projected 84% lower AI-serving spend and about 29% lower total Snowflake spend over the modelled term. The credit pool lasted roughly seven to eight months longer in the scenario. That gives the firm more room to expand AI access before buying additional Snowflake capacity.
Shared definitions support more accurate answers
The same implementation addressed answer quality. Warehouse-wide context gives agents the joins and business definitions behind a question, and the sandbox lets them test query logic before live execution. IT and Data retain control of definitions and access as more people use AI on the warehouse.
Measurement
The comparison combines the account audit with Rig’s multi-tenant session benchmark. Savings and credit runway are projections at equal adoption and the audit’s credit price. Per-question estimates cover Snowflake serving consumption; Rig fees, external model costs and context builds are separate. The model uses a hosted Rig benchmark for its AI-serving assumption.