The operator's memory — how your business runs, its patterns, your processes, your corrections — lives inside your own Microsoft 365 or Google environment. Not on an APGC server. Not in a shared system. We access it the same way any authorized app does: a standard OAuth connection you grant once and can revoke any time.
When the operator processes something, it runs through the Claude API (enterprise-grade, SOC-2 compliant, contractually barred from training on your data) and a private automation layer we operate. Processed in the moment. Not kept. Nothing about your business feeds into a shared or learning model.
Every client is fully isolated. You own the operator's memory. If you leave, you keep it — or we delete it. Your call.
The operator is read-first. It reads across your tools to surface what matters — then acts only when you say so, in the scope we've agreed to. Nothing moves without approval. We request only the permissions the operator needs to do its job — no broad admin access, no blanket read of your file system.
| What we connect to | Why | Access level |
|---|---|---|
| Microsoft 365 or Google Workspace | Email, calendar, and document context — the operator's memory lives here | Read + limited write (scoped) |
| CRM (if applicable) | Pipeline and client data — surfaces lead-speed and reliability gaps | Read only |
| Accounting / invoicing (if applicable) | Invoice and collections data — surfaces collections gap | Read only |
| Scheduling tool (if applicable) | Appointment data — surfaces reliability gap, powers reminders | Read + write (scoped) |
The exact scope is agreed in writing before the build begins. You see every connection before it goes live.
When the operator runs a task, it reads what it needs, processes it through the Claude API and our private automation layer (n8n, self-hosted), and acts or reports. The data is in transit — not stored on APGC infrastructure. Logs record that a task ran and what it did; they do not store the underlying content of your emails, invoices, or client records.
The Claude API is operated by Anthropic. It is SOC 2 Type II certified. Anthropic's enterprise terms contractually prohibit using your data for model training. If you'd like to review Anthropic's privacy documentation directly, it's at anthropic.com/privacy.
If something goes wrong — an unexpected access, a behavior outside scope, a security concern — we follow a defined response. We log it, kill the relevant connection immediately, and notify you the same day. We walk through what happened and fix it before anything goes live again. You won't be left guessing.
You can revoke any OAuth connection at any time from your own Microsoft or Google account settings, instantly cutting the operator's access. No call required, no waiting on us.