Compliance

One compliance posture, set by your admins, that every agent follows whoever built it and however it's set up.

Many people build and change agents in Mindset. Compliance isn't left to each of them. Your admins set the rules once, for the whole org, and Mindset applies them to every agent.

One posture for every agent#

An agent's setup covers its instructions, its model and the actions it may take. The compliance rules sit above that, in the org's settings. Changing them takes an org admin, in the Studio or through a coding agent or AI client signed in as that admin. Your agents, the people they talk to, members and API keys are refused.

A new agent starts under the same rules as every other, and changing an agent never changes them.

What your admins decide#

Rule	What it does	Default
Personal data protection	Swaps personal details for placeholders before a message reaches the AI provider	Off
What counts as personal data	Email addresses, phone numbers, card numbers, IBANs, national IDs and IP addresses, plus patterns of your own	All on, once protection is on
Approving actions that change other systems	Whether each new one is switched on automatically or waits for a person to approve it	Switched on automatically
Audit retention	How long audit records keep identifying details, from 6 to 120 months	6 months
Conversation retention	When conversations expire	Kept, except test runs after 30 days
AI provider account	Which provider account your agents' AI model calls go through, so the provider bills you directly	Your org's own key
Telemetry export	Sends traces, metrics and costs to your own monitoring	Not set up
Data residency	The region your org's data lives in	Fixed when the org is created

Each setting and where to find it is on Org settings.

What holds without any setup#

These rules apply to every agent in every org, and no setting turns them off.

Agents reach your systems only through named actions. Mindset calls them operations. There's no general web access, so even a misled agent can do only what its operations allow. See Connections and operations.
No agent can approve its own actions. A person, or the org setting above, approves each action that changes another system.
Passwords and keys never reach an agent or the AI model. They're kept in a secret store and added to each call on the way out.
Untested actions don't go live. A change that uses an action that hasn't run successfully is refused unless an admin overrides it, and the override is recorded.
Every action is recorded. No agent can skip the audit record, nobody can edit or delete it, and each entry is chained to the one before, so tampering shows. Re-verify now checks the whole chain.
Test and production are kept apart. Each environment has its own agents, connections and keys.

The detail is on Security and approval.

Evidence for your auditor#

Audit → Obligations maps the record to seven frameworks: EU AI Act, GDPR, SOC 2, NIST 800-53 AU, ISO 42001, HIPAA and SEC 17a-4. Most controls are met by how Mindset is built. The rest need a person to vouch for them: an admin attests, and the attestation goes into the record with their name. Generate pack exports the evidence for one framework, with the chain re-verified.

See Audit and personal data.

Defaults to decide on#

What it doesn't cover#

Mindset's own building agents. Orca and the builders, which your team uses to make agents, don't replace personal data.
An AI model connection can switch personal data protection off for calls made through it, such as a function's. Agent conversations always follow the org setting.
Detection is best effort. Placeholders can be turned back into real values, so the data is pseudonymised, not anonymised.
The record shows tampering. It isn't certified as a legal archive. HIPAA and SEC 17a-4 controls need an admin's attestation.
Turning automatic approval on or off isn't written to the audit record.

For the questions a security or procurement team asks, see Mindset for your security review.