---
title: "Pioneer AI data and privacy"
description: "What Pioneer AI sends to a model provider, what it never sends, where conversations are stored, how long they are kept, and who can read them."
url: "https://support.outpostplatform.com/pioneer-ai/data-and-privacy/"
product: "depot"
type: "concept"
status: "draft"
ai_summary: "Pioneer AI sends the conversation, page context, and tool results to the model provider only when a technician sends a message, and never sends vault credentials, recovery or BitLocker keys, or session recordings. Remote control tasks send screenshots, never stored. Conversations are stored encrypted in the Depot for 0 to 365 days (30 by default), readable by the technician and ai:review holders."
availability:
  plans: ["home-lab", "essentials", "enterprise", "msp"]
  preview: false
source: "https://gitlab.com/outpostplatform/docs/-/edit/main/src/content/docs/pioneer-ai/data-and-privacy.md"
license: "CC BY 4.0"
---

# Pioneer AI data and privacy

This page is for the people who answer for your data: what Pioneer AI sends
to a model provider, what it never sends, where its conversations are kept,
and who can read them. The rules are the same on every plan. Where the answer
depends on the source your organization chose, the page says so.

## When data is sent

Pioneer AI sends data only when a technician acts: when they send a message,
and while the reply to that message is being written. Nothing is sent in the
background, on a schedule, or without a person in the conversation.

## What is sent

Each step of a reply sends the model:

- Fixed instructions that describe Pioneer AI, with the technician's display
  name and which console they are in.
- Your organization's [instructions](/pioneer-ai/instructions/), and the
  client's when the conversation is about one.
- Pinned [memories](/pioneer-ai/memories/) in scope: the technician's own, the
  client's, and the organization's. Pioneer AI can look up other memories
  when they seem relevant.
- The conversation so far: the technician's messages, Pioneer AI's replies,
  and the tools it called with their results. A long conversation sends a
  summary of its earlier messages instead, and older tool results are
  shortened.
- With each message, the page the technician was on: its address, its title,
  and the item it showed, such as an endpoint or a ticket.
- What the tools Pioneer AI reads return, limited to what the technician's
  permissions allow.

Whatever a technician types is sent, so a message must never contain
passwords, keys, or other secrets.

A few smaller requests use the same conversation: writing its title,
summarizing it when it grows long, and writing the note for
[Save to ticket](/pioneer-ai/use-pioneer-ai/#save-a-conversation-to-a-ticket).
[Sub-agents](/pioneer-ai/use-pioneer-ai/#sub-agents) receive their task and
the same instructions and memories. The
[automatic approval](/pioneer-ai/automatic-approval/) review sends less: the
technician's own recent messages and the summary of the proposed action, plus
any safety review policy and team memories that a person saved.

## Web search

Unless an owner turns [web search](/pioneer-ai/web-search/) off, Pioneer AI
can send search queries to Brave, the search provider named in Schedule B, and
read public web pages it found or that a technician linked. A query or page
address is refused before it is sent if it contains a client name, an endpoint
hostname, an email address, an IP address, or your console's address. Brave
receives only the text of each query. Conversations keep only the title and
address of each result, never the result text or the page text.

## Outpost documentation

Unless an owner turns [Outpost documentation](/pioneer-ai/outpost-docs/) off,
Pioneer AI can send search queries to this documentation site and read its
pages. A query is refused before it is sent under the same rules as a web
search. Conversations keep only the title and address of each page, never the
page text.

## Screenshots of a remote session

When a technician approves a task in which Pioneer AI
[controls a remote session](/remote-access/pioneer-control/), each step sends
a screenshot of that session's screen, as the technician's viewer shows it, to
the AI model, through the same path as every other request. Nothing is sent
outside an approved task, and the `remote_access.ai_control` policy setting
can turn the feature off for any endpoint.

Outpost does not store the screenshots. Your Outpost holds each one,
encrypted, only until the task reads it, and for at most 5 minutes, then
deletes it. While the task runs, the model receives the last 3 screenshots
with each step. Screenshots are never written to the
conversation, the logs, or the [automatic approval](/pioneer-ai/automatic-approval/)
review. The conversation keeps each step as text, such as `Click (412, 233)`.
With Outpost Cloud AI, the providers' terms in Schedule B apply to the
screenshots as to any other request.

## What is never sent

No Pioneer AI request sends vault credentials, managed administrator
passwords, recovery keys, BitLocker keys, or recordings of remote sessions.
Pioneer AI cannot reveal them. See
[Actions that are never allowed](/pioneer-ai/approve-actions/#actions-that-are-never-allowed).

## Where it goes

| | Outpost Cloud AI | Your own endpoint |
| --- | --- | --- |
| Path | Your Outpost sends the request to the Outpost service, which forwards it to an AI service provider. | Your Outpost sends the request straight to your endpoint. |
| Providers | The AI service providers named in Schedule B of the Outpost Terms of Service: OpenAI, Anthropic, Google (the paid Gemini API), and Fireworks AI. | The provider you configured. It is your subprocessor, not Outpost's. |
| Training | Each provider is under a contract that prohibits training models on the data Outpost sends. | Decided by your provider's terms. |
| Provider retention | Set by each provider's terms, which Schedule B summarizes. | Decided by your provider's terms. |

The Outpost service does not store the text of a prompt or a reply. For each
request it records usage: the model, token counts, timing, the outcome, and
references to your account, the console user, the conversation, and on the
MSP plan the client. Outpost Cloud AI is billed from that record.

## Where conversations are stored

Pioneer AI conversations are stored in your Outpost. Their titles,
messages, tool results, large results kept for filtering, page context, and
summaries are encrypted. The model's **Thinking** is shown while a reply is
written and is never stored. They are deleted with the conversation.

[Memories](/pioneer-ai/memories/) are stored encrypted in your Outpost too.
They are kept until someone deletes them, a person's personal memories are
deleted with their console account, and a client's memories are deleted when
the client is offboarded.

Conversations are never used to train models, by Outpost or by the providers
in Schedule B.

## How long they are kept

An owner sets the retention period on **Settings** > **AI**, in **Keep
conversations for (days)**:

- From 1 to 365 days, counted from a conversation's last message. The default
  is 30 days.
- `0` deletes a conversation after a day without activity.

Deletion runs automatically every hour and removes the whole conversation. A
technician can delete one of their own conversations at any time from **Your
conversations**. A deleted conversation cannot be recovered. Usage records,
which hold no text, are kept for billing and reporting.

## Who can read them

| Who | What they can do |
| --- | --- |
| The technician who started a conversation | Read it, continue it, rename it, and delete it. |
| People with `ai:review` | Read every technician's conversations, read-only. Each time one opens a conversation, the [audit log](/depot/audit-log/) records it. |
| Everyone else | Nothing. A conversation that is not yours does not open. |

The owner holds `ai:review` by default. On an MSP's practice console, every
role that holds `ai:manage` holds it too. See
[Review a technician's conversation](/pioneer-ai/usage/#review-a-technicians-conversation).

## What is logged

- Every Pioneer AI tool call, including refused ones, under **Recent
  activity** on **Settings** > **AI**.
- Every change Pioneer AI makes, in the [audit log](/depot/audit-log/), as the
  technician's own change, marked as made through Pioneer AI.
- Every model call, on **Settings** > **AI** > **Usage**. See
  [Review Pioneer AI usage](/pioneer-ai/usage/).

## Related

- [Pioneer AI](/pioneer-ai/)
- [Approve Pioneer AI actions](/pioneer-ai/approve-actions/)
- [Turn on automatic approval](/pioneer-ai/automatic-approval/)
- [Set up Pioneer AI](/pioneer-ai/set-up/)