Data + AI
Agents that act where work happens, on data teams can trust.
The value of AI shows up when agents reason over governed data inside Snowflake and take the next action in the tools teams already use. We build both, and the interplay between them, with Claude as an Anthropic partner. Proven small first, in production in 8–16 weeks.
How it works
Two planes, one governed loop
Agents run in two places. Inside Snowflake, they reason over governed data without moving it. In the flow of work, Claude takes the next action where teams already are. The point is the interplay: a request crosses both and comes back cited, in policy, and logged.
Inside Snowflake
Reason over governed data, without moving it
- Cortex agents + AISQL, next to the data
- Semantic Views: one shared definition of every metric
- Cortex Analyst + Search for cited answers
- AI Agent Identity: per-agent access, full lineage
In the flow of work
Take the next action where teams already work
- Claude in Slack, docs, and the browser
- Claude Code in the engineering workflow
- Answers + actions back inside the ERP, CRM, and apps
- Custom agents on governed endpoints
One identity
Every agent, inside or out, acts as a governed identity with per-agent permissions.
Governed endpoints
The business exposes data as endpoints agents call, never raw copies they hold.
MCP
Anthropic's open protocol connects Claude to governed data and the tools teams use.
One request, both planes
A rep asks Claude in Slack which accounts are at risk.
Claude calls a governed endpoint; Cortex answers from Semantic Views, with citations.
Claude drafts the outreach in the rep's inbox, inside policy.
The whole exchange is logged against one identity and lineage.
What we ship
Six jobs, running across both planes.
Each one reasons over governed data inside Snowflake and shows up where teams work. Where a real engagement proves it, the numbers are on the card.
Paperwork that reads itself
Claims, invoices, and contracts classified, extracted, and routed in seconds instead of hand-sorted piles. The document backlog becomes a table teams can query.
60→95% accuracy · 4 sec per document
Executives who ask the data directly
Plain-language questions answered with citations against governed definitions, so the Monday meeting starts from the same number, not three versions of it.
Thousands of runaway SKUs, one catalog agent
Churn and failures, flagged early
Models on usage, billing, and sensor data that surface at-risk customers and equipment before the quarter ends, with the reasons attached.
Documents, searched by meaning
Policies, contracts, and wikis answered directly, with sources cited. Retrieval grounded in the organization's own content, so answers stay accurate and current.
Agents that act, inside policy
AI that does the next step, not just describes it: drafting the response, filing the update, kicking off the workflow, each action inside per-agent permissions with a full audit trail.
AI where teams already work
Answers and actions flowing back into the ERP, CRM, and apps the business runs on, so nobody has to visit another dashboard to benefit.
A use case we haven't listed? If the data can carry it, we can ship it. Tell us the job →
Governed by default
One perimeter, one identity, both planes
Inference runs inside the business's Snowflake account, so prompts and results stay within the same perimeter as the data. Agents in the field reach it through governed endpoints, never raw copies, acting as verifiable identities with per-agent permissions. Access, masking, and lineage carry across both planes, and Cortex guardrails screen every call.
- Models run next to governed data: no copies, no export
- Every agent acts as one governed identity, inside or out
- Access, masking, and lineage inherited from Snowflake Horizon
- Cortex Guard and AI guardrails on every call
Anthropic partner
Claude, by default.
As an Anthropic partner, Claude is the reasoning model we reach for first, inside Snowflake and in the flow of work. It is the same model on both planes, so the behavior a team trusts in a demo is the behavior that ships.
- Inside Snowflake
- Claude runs as a Cortex model, so inference happens next to governed data and nothing is copied out.
- In the flow of work
- Claude and Claude Code work in the tools teams already use, from the inbox to the IDE.
- Connected by MCP
- Anthropic's open protocol links agents to governed endpoints and tools, with no bespoke glue.
SELECT AI_COMPLETE(
'claude-opus-4-8', -- our default reasoning model
CONCAT('Classify this claim: ', claim_text)
)
FROM governed.claims;The estate stays open. Every major model is callable in Cortex, and swapping one is a one-line change. We benchmark per use case on quality, cost, and latency, and Claude is where we start for enterprise reasoning.
Credentials
Certified to run AI on enterprise data
Every model runs under Snowflake's independently audited controls, delivered by a SnowPro-certified team that does this every day, and, as an Anthropic partner, builds with Claude.
Anthropic partner, building with Claude
Put the right agent on the work.
Tell us the use case. We'll bring Claude, the governance, and the people to ship it into production with the team, inside Snowflake and in the flow of work.