Data Analytics

Analysts want a model to summarize a dataset, propose a metric definition, or explain a query. Two risks follow that wish. Personal records leak into a prompt, and a confident paragraph gets treated as a finding. The chain here works on both sides of the model call.

Only document search is permitted to the agent. Requests for a shell, for code execution or for a database write are refused outright. The assistant cannot run a job, write to a table, or reach your warehouse at all. Add a read tool deliberately, and review what that tool exposes before you enable it.

Injection patterns are screened before the prompt leaves the gateway. Key and session identifier shapes are masked next. The words private key, secret and token stop the call outright. Tool call analysis stays local, so query text is not sent to an outside analyzer.

The response floor sits at a combined 0.78, at the low end of the range used across these starters, with accuracy at 0.80. That suits exploratory drafting, where an analyst reads every line anyway. Raise it when the same assistant starts writing text that reaches a customer. Answers below 120 characters or one sentence long are refused.

No role check appears in this chain, so identity remains an assumption you satisfy elsewhere. The oversight step is empty and escalates nothing until configured. Audit marking notes an evaluated request without storing it. Choose the Software as a Service Platforms starter when the assistant answers customers inside your product rather than analysts inside your company.