What agents can see

Data sources

Settings → Integrations → Data sources lists every connected origin of data and when it last synced. That list is the honest answer to "what does Taizen know about us".

Four kinds of record arrive from those sources:

If an agent's output looks generic, check here first. An integration that hasn't synced is the most common reason an agent reasons in generalities — it has nothing specific to reason about.

Scoping

Everything an agent reads is scoped to your workspace. Within that, scope is governed by:

• Integration scopes. What you granted when connecting. An integration connected read-only can't write back, and the run will say so rather than silently skipping (integrations overview).
• The agent's own instructions. Which team, which region, which segment, which call types. Most "the agent saw too much" problems are missing exclusions rather than a permissions issue (build an agent).
• Member roles. What a person can see and change in the UI — Admin, Editor, Viewer (roles).

Where your data lives

• Your data stays in your own region.
• It is not used to train models.
• Interview transcripts are encrypted.
• Taizen is SOC 2 Type II and ISO 27001 certified.

Retention and deletion

You can delete an interview at any time, and deletions cascade to derived artifacts — battlecards, roleplays, coaching notes. Something built from a deleted interview doesn't survive quoting it.

That cascade is the property to lean on when a buyer asks to be removed: you don't have to hunt down every artifact that mentioned them. See consent and compliance.

For documents, archive rather than delete when the material is superseded but historical sessions and scores still reference it — archiving keeps old results interpretable.

Secrets

Never put a secret in an agent's instructions or a skill: run logs record inputs and steps, so a key pasted into a prompt ends up in the run history where anyone who can read the Timeline can see it. Connect the system under Settings → Integrations instead, so the agent reaches it through the connection rather than through its prompt.

Auditing what happened

Every run logs its inputs, its steps and its output (runs and debugging). When someone asks "how did the agent know that", the answer is in the run, not in a guess about the model.

A short checklist for a security review

• What's connected — Settings → Integrations, and the scopes on each
• Who has Admin — Settings → Organization
• What's shared externally — Settings → Shared links
• Which MCP servers are connected — Settings → MCPs
• What the audit trail shows — Agents → Timeline