19. Data Handling & Safety¶
This section states plainly where your information lives, what leaves your machine, and which controls protect you. Precision matters here, so the claims are deliberately narrow.
Storage is local. Inference is not automatic.¶
Local-first describes storage. The working record — your notes, documents, analyses, and BOB's outputs — is held as plain files on hardware you control. There is no supplier-hosted copy of your vault and no requirement to upload your document library to anyone's platform.
Inference is a separate question. When BOB performs an analysis, the specific material needed for that task — your request plus the relevant extracts — is transmitted to the AI provider you have configured and processed under your own agreement with that provider. BOB operates bring-your-own-key: the supplier of BOB is not in the data path and holds none of your material.
| What | Where it lives / goes |
|---|---|
| Your vault (all business data) | Your machine, as files you control |
| Task prompts + relevant extracts | Transmitted to your configured AI provider, per task |
| BOB's outputs | Written back to your vault as files |
| Nothing | Held by BOB's supplier |
What your provider terms govern¶
On paid business or API terms, the major providers state that customer content is not used to train their models and is retained only briefly. Paid terms change what the provider may do with transmitted content; they do not change the fact of transmission. If your requirement is that nothing leaves your premises at all, the answer is a locally hosted model (BOB supports this) — with the trade-off that local models are less capable than the frontier commercial ones.
The drafts-first rule¶
Anything hard to undo — sending an email, publishing a post, deleting files, changing external systems — waits for your explicit approval. BOB drafts; you approve. This is not a setting you can forget to enable; it is how the workflows are written.
Review discipline: AI output is a draft, not a determination¶
Language models are probabilistic and can err — including by omission. BOB's controls are designed to make errors detectable, not to pretend they're impossible:
- You review before relying. Business-critical output gets read by a qualified human — you — before it's sent, filed, or acted on.
- Grounded work cites its sources. Analytical deliverables reference the vault files they drew from, so any claim can be traced back and checked.
- Runs are logged. Playbook and agent runs record what was invoked and what files were produced, so you can retrace what happened.
Practical safety habits¶
| Habit | Why |
|---|---|
| Keep credentials in configuration, never in chat | Chat content goes to the AI provider |
| Back up the vault like any business-critical folder | It's plain files — any backup tool works |
| Review outward-facing drafts word by word | Your name is on what gets sent |
| Record which provider, account, and terms you use | The one-time check that makes the confidentiality position auditable |