Sidecar Agents is built for businesses that cannot hand their documents to a public chatbot. This page sets out how data is hosted, isolated, permitted and audited, in the terms we are prepared to be held to.
Every client environment runs on Amazon Web Services. Your documents, indexes, chat history and audit logs are stored in the Sydney region. AI answers are processed within Australia, in Sydney or Melbourne. This has been verified on the running system service by service, not just designed that way.
The AI models run on Amazon Bedrock. Sidecar does not train models on your content, and Amazon Bedrock does not use customer content to train the models it hosts.
Each client has its own storage, queues, indexes and analytics. There is no shared state of any kind between clients.
Isolation is asserted by automated tests that run on every change, and a failure is treated as a breach rather than a bug.
Credentials issued by Google, Microsoft or HubSpot live only in an encrypted secrets store, never in code, configuration or logs.
| Provider | Permission | What it is for |
|---|---|---|
| Google Drive | drive.readonly | Reading file content so a document can be classified and answered from |
| Google Drive | drive.metadata.readonly | Listing drives, folders and sharing so the estate can be mapped |
| Microsoft 365 | Files.Read.All, Sites.Read.All (delegated) | Reading files and sites the consenting administrator can already open |
| Microsoft 365 | Sites.Selected (application, optional) | Archie filing documents, only in libraries granted one at a time. Grants nothing on its own |
| Microsoft 365 | offline_access, openid | Keeping the connection alive, and identifying which directory it belongs to |
| HubSpot | Read scopes for companies, contacts, deals, tickets and owners | Customer context for the Brain. Write scopes are requested only if the retention agent is enabled |
Reading is delegated on purpose: the agents see exactly what the consenting account sees and nothing more. Inside the Brain, a second layer applies. Documents carry one of four sensitivity levels, and each role has a ceiling.
No agent has a delete operation. Only a signed-in person can delete, and disposing of a file always needs approval.
Archie will not give more people access to anything, and no configuration changes that.
File moves, permission changes and CRM updates are proposals until the file owner, document controller or named approver confirms. A request that times out is not approved.
Before any permission change, the current state is saved. A restore returns people to the access they had.
After each change Archie re-reads the file to confirm it happened, and stops at the first thing it cannot verify.
Each run can change only a set number of items. Access is added before it is removed, and each step is approved separately.
Archie keeps a permanent journal of every change it makes, with the full state before and after. That journal is what reversals and audits are built from. The other agents keep a retrieval ledger: every lookup is logged before and after it runs, including failed and empty ones, and findings cite those entries.
In the reporting and retention agents, the action log, grounding record and model-call log cannot be edited or deleted, even by the agent’s own database role.
We do not claim certifications we do not hold. Ask us directly about the compliance requirements that apply to your business. Full details are in the Privacy Policy and Terms of Service.
Send them over. We would rather answer them before you connect anything than after.