Skip to content
Optima docs
AI assistants (MCP)

Permissions and privacy

Which roles unlock which assistant tools, and how purposes, the sensitive data level and the audit log apply to an assistant.

An assistant connected to Optima works through one API key, so everything that limits the key limits the assistant. This page explains how to choose the key's settings and how to see what an assistant did.

Roles decide which tools appear

The assistant is shown only the tools its key's roles allow; a tool the key cannot use is not offered at all. When you change the key's roles, the assistant's tool list follows within about a minute. Start a new chat so the assistant picks up the new list.

Use caseRolesWhat the assistant can do
Analyst (read only)data_viewer, analytics_viewer, campaign_viewerRead the data model, profiles, events, reports, metrics, dashboards, segments, campaigns, journeys and their performance. No tool that changes anything appears.
Campaign drafterthe analyst roles plus audience_editor and campaign_editorAlso create and edit segments, create draft campaigns, edit drafts, run pre-flight checks, draft journeys (saved inactive) and request approval.
Report builderadd analytics_editorSave reports and metrics. A metric with an alert rule can send alert emails.
Data writeradd data_editor (and privacy_officer to record consent)Write events, set profile attributes and record consent. Use only when the assistant really has to write customer data.

A campaign drafter cannot activate, approve or send a campaign through the assistant, even though campaign_editor can do so in the console. Those tools do not exist. See Tools for the complete list of what is never available.

Purposes

When your environment requires a purpose for every read of customer data, the assistant declares one with each such request, and it can only choose from the purposes granted to its key. Grant the key the purposes its job needs and no more, from the key's Purpose grants action on API keys. A request under a purpose the key does not hold is refused. See Consent and purposes.

Sensitive data level

Set the key's PII level (Set PII level) to the lowest level that still answers the questions you expect. Attributes above that level are withheld from every answer, and the assistant is told which attributes were withheld, so it can say so instead of guessing. An analyst assistant rarely needs names, email addresses or phone numbers.

Writes ask first

Tools that change something are marked as such, and assistants ask you to confirm before they use one. Even then, nothing an assistant does is destructive:

  • Campaigns are created in draft, and only drafts can be edited.
  • Journeys are saved inactive.
  • Next-best-action decisions are always a dry run: no touch is logged and no frequency cap is used up.
  • Recording a withdrawal on a purpose whose policy erases data on withdrawal is refused; do that in the console.
  • Events written by an assistant cannot be undone through the assistant, and editing a segment changes the audience of every campaign that uses it. Grant those roles with care.

What the audit log shows

Every read of personal data is recorded in the privacy audit log with the key as the Actor, the purpose and the attributes returned. For a chat connection (Claude, ChatGPT) the entry also says "via AI connection" with a short id, so you can tell which connection read a profile when one key has several. Review it under Settings → Workspace → Privacy, tab Audit log. See Privacy.

Large answers

Answers have a size limit. When a result is too large, the assistant receives part of it with a note that rows were left out, and should ask for a smaller page or a narrower filter. Ask it to do so if it does not.

On this page