AI privacy and security for self-hosted support teams
For a self-hosted support team, AI privacy comes down to one question per mailbox: where does the conversation text go when a draft is generated? With ReplyRabbit the answer is always the provider you chose for that mailbox, and only for the features you enabled. With a cloud provider, the text needed for the task is sent to that vendor. With Ollama or LM Studio on the Pro plan, it's processed on your own server and never leaves. Everything else, keys, tags, drafts, the audit log, stays in your FreeScout database.
What leaves the server, and when
| Feature | What's sent to the AI provider | When |
|---|---|---|
| Draft replies | The conversation text needed for the reply, plus company context, response style and examples | When a draft is generated (on demand or by Auto Draft) |
| Next Actions, AI Signals | Conversation text for classification or the checklist | When those features are on (Pro, Team) |
| Semantic search | Chunks of your help articles, docs and products, for embeddings | When knowledge sources are indexed (Pro) |
| Attachment vision | The image or PDF content | When attachment reading is on (Team) |
| Spam and noise filtering | Nothing; filtered mail is skipped before any AI call | Always |
A conversation ReplyRabbit filters out as spam, a newsletter or a receipt is never sent anywhere. Workflow Tags don't make any AI call of their own; they reuse the analysis that already ran.
Local models remove the third party
Choose Ollama (Local) or LM Studio (Local) as a mailbox's provider and every row in that table is processed on the FreeScout host. There's no API key because there's no external account. This is the setting for mailboxes that handle contracts, medical questions, financial details or anything under a data agreement that excludes cloud processing. It's per mailbox, so a general inbox can stay on a cloud model for draft quality.
Controls an admin should set
- Provider per mailbox. Decide which mailboxes may use cloud AI and which must use local, and set them accordingly. Agents can't override it from their side.
- API keys as secrets. Keys are stored encrypted and shown masked after saving. Rotate them when the person who created them leaves.
- Knowledge source boundaries. FreeScout Knowledge Base articles are public and can be quoted in drafts. Outline pages on the Team plan stay internal and are never quoted or linked to customers.
- Tracker and Slack scope. Issue trackers and Slack receive the handoff content, not the whole mailbox. Keep them disabled on mailboxes whose data shouldn't reach those tools.
- Audit log. On the Team plan, every AI action is recorded, so a reviewer can see what was drafted, classified or created, and when.
- Human send. No plan sends a customer reply automatically. The only automatic outward action is issue creation on Team, to your own tracker.
Security that's yours to keep
ReplyRabbit inherits FreeScout's security model. If FreeScout is exposed without HTTPS, weak passwords or stale updates, AI features don't change that risk, and a local model server on the same host needs the same care: restrict it to localhost, keep it updated, and monitor load. A local model is a privacy control, not a security control.
Provider trade-offs are in Choosing an AI provider, local setup in Ollama integration, mailbox configuration in How to set up ReplyRabbit, and the short answers in the FAQ.
Official references
Provider names, model lists, pricing, privacy terms, and legal guidance change. Check the current source before setting a production policy.