Enterprise use case · Snowflake + OpenAI + Microsoft 365
Questions over operational data
Business users wait for an analyst to translate a question, while informal AI answers can omit the metric definition or invent a number. Authorized users receive answers grounded in a visible, validated Snowflake query and its actual results.
Systems involved
What each system contributes
The model step runs against the AI provider you choose, including Anthropic, Google Gemini or a model you host yourself. Anything outside a pre-built action runs through an API, protocol or database call, or as Rules + Code in Python, Bash, Go, Java, .NET, PHP, PowerShell or JavaScript/TypeScript, with the same approvals and audit.
Architecture
One governed run
How it starts: An authorized user asks an operational-data question through the designated Microsoft 365 interface.
Step by step
How the process runs
- 01
Resolve the user's scope
Apply the requester's permitted data access and the approved Snowflake views and metric definitions.
- 02
Translate the question
The selected model proposes a read-only query and explains its filters, period and interpretation of the requested metric.
- 03
Validate before execution
Rules check allowed objects, statement type and query limits, requesting clarification when the business meaning is ambiguous.
- 04
Run the permitted query
A separate Snowflake step executes only the validated query under appropriately scoped credentials.
- 05
Answer from the result
The model summarizes the returned data through Microsoft 365 with the query, scope and limitations available for inspection.
Control
Human decision and exception path
Human decision
A named data owner approves new views, metric definitions or access scopes before they are made available. Sensitive or exceptional queries pause for the required reviewer; ordinary permitted reads do not authorize any write, access expansion or financial action.
Exception path
- Ambiguous periods, entities or metric definitions prompt a clarification rather than an assumed query.
- Disallowed SQL, objects or data scopes are blocked instead of executed with broader credentials.
- Empty, failed or truncated results are reported as such; the model does not substitute a plausible numerical answer.
Evidence
What gets audited
- Business question and authorized data scope.
- Proposed query, metric definition and validation results.
- Executed query reference and permitted result context.
- Final answer, limitations and any data-owner approval.
Nexus runs on premises, in private or sovereign cloud, or air-gapped. Credentials stay in the vault, roles decide who can approve, and the execution record stays inside your deployment boundary.
Sovereignty and deploymentRelated
Patterns, connectors and next steps
- Snowflake-SAP integration
Integration pattern · SAP data reaches Snowflake as validated, repeatable loads with a record of completeness.
- Snowflake connector
35 pre-built actions
- OpenAI connector
25 pre-built actions
- Microsoft 365 connector
45 pre-built actions
- Governed self-service
Requests with policy and approval built in.
- Order to cash across CRM, ERP and billing
Another enterprise use case.
FAQ
Questions about this use case
- Does Nexus replace any of these systems?
- No. Each system keeps its own records. Nexus runs the process between them and records each decision.
- How does the run start?
- An authorized user asks an operational-data question through the designated Microsoft 365 interface.
- What if a step is not covered by a pre-built action?
- The model step runs against the AI provider you choose, including Anthropic, Google Gemini or a model you host yourself. Anything outside a pre-built action runs through an API, protocol or database call, or as Rules + Code in Python, Bash, Go, Java, .NET, PHP, PowerShell or JavaScript/TypeScript, with the same approvals and audit.
- Can the model run arbitrary SQL?
- No; the proposed statement passes read-only, object-scope and execution-limit checks before a separate step runs it.
- Can two users receive different answers to the same question?
- Yes, if their permitted data scopes differ; the answer must show the scope and filters rather than imply a company-wide total.
Use case review
Discuss this use case
Bring your version of the process. We will map the systems, approvals and evidence it needs.