Platform · Trust & Control

Six layers between your agents and doing the wrong thing.

Per-agent autonomy levels, per-action policies, dynamic risk scoring, contact-level overrides, hard guardrails on every action, and a kill switch that stops everything in under 2 seconds. You don't trust AI by default — you trust it because you can stop it.

AI agents are powerful enough that you need a credible way to say "no, not that one." ScendCore ships six layers of control out of the box — you configure once, the Brain enforces in real time, and every override is audit-logged.

Enforced by the Brain
6 layers, every decision
In this explainer
  1. 01Autonomy Levels
  2. 02Action Policies
  3. 03Dynamic Risk
  4. 04Contact Controls
  5. 05Hard Guardrails
  6. 06Kill Switch
Phase 01

Autonomy Levels · Per-agent dial

Each agent runs at one of four levels: Full Auto (does the work, logs the result), Act + Review (does, you get notified), Suggest Only (proposes, you decide), Off (paused). Per-agent, per-use-case, per-channel.

Most teams start every new agent in Suggest Only, graduate to Act + Review at 30 days, Full Auto at 60-90.

Specification
  • 1.1Four levels: Auto · Review · Suggest · Off
  • 1.2Per-agent + per-use-case + per-channel granularity
  • 1.3Graduation tracking (how long at each level, drift detection)
  • 1.4One-click "back to suggest only" if behavior surprises you
Phase 02

Action Policies · Allow, deny, approve

Per-action rules: "never email a VIP without approval," "cap outbound at 500/day," "pause Alex during renewals," "require 2 approvers for contracts >$50K." Written in RevOps, enforced by the Brain at decision time.

Specification
  • 2.1Allow / deny / require-approval per action type
  • 2.2Volume caps (per day, per week, per agent, per channel)
  • 2.3Multi-approver workflows for high-stakes actions
  • 2.4Policy versioning + diff history
Phase 03

Dynamic Risk · Live re-scoring

Every action carries a risk score computed at decision time: recipient sensitivity, action reversibility, blast radius, dollar value. High-risk actions auto-promote to require approval even if policy normally allows them.

Drift detection too — if an agent suddenly starts producing higher-risk actions, autonomy automatically downgrades until reviewed.

Specification
  • 3.1Per-action risk score (recipient, reversibility, dollars)
  • 3.2Risk-triggered approval promotion
  • 3.3Behavioral-drift detection (sudden risk-profile change)
  • 3.4Auto-downgrade to lower autonomy on anomaly
Phase 04

Contact Controls · Per-account overrides

Account-level + contact-level rules: "Acme is a VIP — Suggest Only," "Jane Doe is on sabbatical — pause until October," "the CEO never gets agent emails." Overrides everything above.

Specification
  • 4.1Per-account autonomy override
  • 4.2Per-contact do-not-contact + pause windows
  • 4.3Role-based contact rules ("CEO + CFO suggest-only")
  • 4.4Industry / region exclusion lists
Phase 05

Hard Guardrails · 13+ always-on

Non-negotiables that fire regardless of policy: never send during off-hours unless flagged urgent, never send to unsubscribed contacts, never repeat the same message verbatim, never escalate without a logged trigger, never act on a stale signal >7 days old. 13+ guardrails shipped; team-level custom ones available.

Specification
  • 5.113+ hard guardrails enabled by default
  • 5.2Custom guardrails (Lua-style policy DSL)
  • 5.3Guardrail violations always alert RevOps
  • 5.4Cannot be disabled by an individual rep — only RevOps admin
Phase 06

Kill Switch · Stop everything, <2s

One button in the dashboard. Hit it, every agent stops mid-action, every queued action is held, every in-flight conversation is handed to a human. Restart is a separate, deliberate action.

Tested monthly. Documented runbook. Audit-logged with reason.

Specification
  • 6.1Global kill switch (all agents, all tenants under your control)
  • 6.2Per-agent kill switch (just Riley, just outbound)
  • 6.3Restart requires admin + reason
  • 6.4Tested monthly, runbook published
FAQs

Frequently asked

Trust the agents because you can stop them.

Six control layers. Configure once. Audit always. Stop in under 2 seconds.

ScendCore

The AI system that runs your customer-facing work. Find, engage, and qualify prospects. Follow up, convert, and support customers — governed end to end.

© 2026 ScendCore. All rights reserved.Sign inPrivacyTermsDPASOC 2
Trust & Control — Autonomy, approvals, kill switch | ScendCore