Security & governance
Autonomous where safe.
Gated where it matters.
The core principle: the system must remain safe even when an AI agent is wrong. Safety comes from technical boundaries - permissions, approvals, limits, audit and rollback - not from trusting agent instructions.
The approval model
P0-P4: what runs, what waits for a human
Risk tiers are assigned by policy, not user convenience.
| Tier | Default rule | Examples |
|---|---|---|
| P0 | Autonomous - internal, reversible, low risk | Monitoring, analysis, research, summaries, issue creation |
| P1 | Autonomous inside thresholds | Safe retries, routing, hygiene, bounded metadata changes |
| P2 | Controlled external action | Routine publishing or communication under approved content, audience and policy |
| P3 | Human approval required | New strategy, audience or offer; meaningful budget, targeting or production changes |
| P4 | Executive / legal / security authority | Contract or legal matters, sensitive claims, privacy or security incidents, destructive actions |
Risk tiers are assigned by policy, not user convenience - the UI cannot downgrade risk, and P3/P4 actions cannot execute without a valid, unexpired, action-bound approval.
Controls
How control is enforced
Client ownership and environment separation
One client, one environment. Your accounts, your data, your assets. Client environments and Teamulate's own environment never share memory or credentials.
Least-privilege access
Every integration gets the minimum scopes the signed scope requires - visible in the dashboard, revocable by you.
Secrets stay outside prompts
Credentials live in a managed secrets facility, referenced by connection ID. They never appear in prompts, agent memory, forms, logs or analytics.
Action-specific, expiring approvals
An approval authorizes exactly one action, bound to its content, audience, budget and target. Change anything and the approval is invalid. Every approval expires.
Spend and action limits
Budget caps, batch limits, publish counts and allowed audiences are enforced by the platform - independently of what any agent requests.
Independent Guardian and Metric
QA and measurement are separate agents that do not report to execution. Guardian can block a release; Metric reports what actually happened, with confidence stated.
Append-only audit trail
Who acted, on what, under which policy and approval, with before/after evidence and outcomes. Exports are themselves audited.
Rollback and kill switch
Documented rollback or compensation per workflow where technically possible; pause switches per workflow, per integration, per tenant - and a global external-write stop.
Operations
Incidents and offboarding
Incidents
Incidents follow an explicit lifecycle - open, triaged, contained, investigating, recovering, monitoring, resolved - with severity levels, named owners and audited changes. Security-critical notifications reach named contacts immediately and cannot be silently disabled.
Offboarding
A documented 12-step process: stop workflows, revoke approvals and connectors, export agreed data and evidence, transfer asset ownership, remove access, apply retention policy and obtain sign-off. Because everything runs in your accounts, you keep everything.
Disclaimer
Security controls are configured per client, integration, contract and risk profile. This page describes the Teamulate operating model and target architecture; it does not represent a certification or replace a client security review.
Have a security questionnaire?
We will walk through your requirements, the access model and the approval configuration for your stack.