What runs where, what it reads, and what leaves.
Keeping raw events in your own cloud account removes one class of exposure. It does not remove every one. This page sets out each flow so your engineering and security reviewers can check it against what they see. Anything that is still open is marked as open.
Deployment boundary
In the customer-cloud deployment, the DepSmith application runs inside your cloud account and queries your event store where it already sits. We do not run an environment that holds a copy of your events. A hosted option is being defined separately and will have its own boundary; it is described on the Deployment page once it exists.
Every data flow
This table describes the customer-cloud deployment. "Open" means the answer is being finalised with our first design partners, and we will show it to you before you connect anything.
| Data | Where it goes | Who can access it | Retention | Optional? |
|---|---|---|---|---|
| Raw event tables | Stay in your warehouse. DepSmith queries them read-only. | Your warehouse access model | Yours — we hold no copy | Required |
| Query results and aggregates | Computed in your account; shown to DepSmith users in your deployment | Users you grant access | Open: result caching in your account | Required |
| Catalog definitions | Stored in your deployment as readable text | Users you grant access; approvers you name | Until you delete them | Required |
| Questions to the AI analyst, relevant definitions, aggregated results | Sent to a language model to produce an answer | Open: model provider and processing region | Open: provider retention and no-training terms | Only when AI features are on |
| Answers sent to Slack Planned | Your Slack workspace | Members of the channels you choose | Your Slack retention | Off unless you connect it |
| Answers over MCP Planned | The AI client your user connects (e.g. Claude, Cursor) | That user, under the client's terms | The client's retention | Off unless you enable it |
| Product telemetry | Open: whether any usage telemetry leaves your account, and what it contains | — | — | Open |
| Support material (logs, screenshots) | Shared with us only when you choose to send it | The DepSmith people working on your case | Open: support retention period | Your choice per case |
A small aggregate can still be sensitive. If a question's result would reveal something your policy forbids sending to a model, turn AI features off for that data. How to do that per dataset is being designed and is not available today.
The AI model boundary
Raw tables are not sent to a model. To answer a question, the model sees the question, the catalog definitions it needs, and the aggregated results it explains. Before you connect anything, we will tell you:
- which provider runs the model, and in which region it processes data;
- whether you can use your own API key or your own model endpoint — noting that your own key does not, by itself, decide where processing happens or how long data is kept;
- the provider's retention terms and whether your data may be used for training;
- a sanitised example of a real request payload.
All four are open today. We will not describe this boundary as "no egress": an AI answer needs a model, and the model has to see what it explains.
Access and identity
DepSmith reads your warehouse through credentials that stay inside your account. One service account doesn't automatically mirror each person's warehouse permissions, so how DepSmith users map to warehouse roles, row policies and projects matters. We document that mapping for each deployment and will show it in a two-user test during a pilot.
Open: whether per-user warehouse roles can be passed through, or whether access is enforced in DepSmith alone.
What this does not protect against
Raw rows staying inside your perimeter does not remove software supply-chain risk, compromised credentials, malicious or careless queries, or disclosure through results someone shares. We'd rather say that than imply otherwise.
Certifications and compliance
We display no compliance badges because we hold none to display. Keeping raw events in your account simplifies data-residency questions under India's DPDP Act. That is an architectural fact, not a compliance claim, and it does not cover consent, purpose limitation or rights requests, which stay your responsibility as the data fiduciary.
A more detailed security pack — architecture diagram, least-privilege policy example, operating runbook summary — is available on request during evaluation, as each item is ready.
Questions this page doesn't answer are the ones we want to hear. Ask on a call, or write to contact@depsmith.com.
Request a demo