Understand the connection you are using before deciding which information to give an agent. Skybits can connect people and AI through several routes, and each route has its own data path.
The document's sharing audience answers who can collaborate in Skybits. It does not fully answer which model service, runtime, or operational system may process information needed for an AI task.
The cloud agent works through hosted services
When you ask a hosted agent to summarize or edit work, the task can require your instructions, document content, comments, generated output, and related context. The published Skybits privacy policy also describes stored interaction history and operational processing.
A persistent conversation can be useful because you can return to earlier work. Its existence also matters when deciding what to include in the conversation. Avoid pasting secrets or unrelated personal data just because they are available.
For the active provider list and current commitments, read the published privacy policy and ask about any requirement that needs contractual confirmation. Provider identity, account tier, processing location, retention, and training terms are all relevant.
An external assistant has its own provider relationship
When you connect an assistant you already use, that assistant may receive the Skybits content it reads as part of its task. The provider's terms and your account configuration matter alongside Skybits permissions.
A read-only connection can still disclose the content returned by a read to that authorized assistant. Read-only describes mutation authority; it does not mean that no information leaves the document service.
Review the selected account's terms rather than assuming a familiar brand behaves the same across consumer, team, enterprise, and API arrangements.
Agent Bridge keeps some things local while connecting the conversation
The current Guide says Agent Bridge does not send your model-provider credential to Skybits. It also explains that chat messages, replies, tool transcripts, task status, and usage information can cross the network.
The working folder is not automatically uploaded as a whole. However, content quoted in chat or read by an approved tool can become part of the task. The local runtime may also use a remote model provider.
Use the precise statement relevant to your concern: credentials stay with the local runtime; authorized content may still be transmitted. Calling a workflow local does not settle its privacy behavior.
Separate the information stores
| Store or recipient | Question to ask |
|---|---|
| Skybits item and history | What is stored and who can access it? |
| Conversation context | What information persists across turns? |
| Model provider | What terms govern training, retention, and location? |
| Operational telemetry | What can enter logs, debugging, or support systems? |
| Export or collaborator copy | Who controls it after distribution? |
Encryption, training restrictions, and deletion address different parts of this picture. Encryption does not prevent an authorized service from processing content. A no-training term does not necessarily forbid temporary retention. Deleting an item does not necessarily retract earlier copies.
Draw one data path for your workflow
For a first pilot, write down the source, agent connection, model service, expected output, and intended recipients. Add any external tools the assistant can call. Check each transition against your organization's requirements.
For example: approved project notes → selected agent connection → permitted model service → draft client update → named reviewer → approved recipients. If you cannot identify a service or a retention condition that matters, resolve it before adding sensitive information.
Next step: Review the Privacy Policy and the Skybits Help. Ask the team for current provider terms when your requirements go beyond the published explanation.