Pre-execution control for AI agents

Stop unsafe agent actions before they happen.

Preflight AI sits between an AI agent and irreversible business operations. It simulates impact, applies policy, requires human approval when needed, and records an auditable execution trail.

✓ Pre-execution✓ Human-in-the-loop✓ Audit-ready✓ Sandbox-safe
Control flow
01Propose
02Simulate
03Gate
04Execute
agent_action = { "operation": "apply_discount", "segment": "paid_customers", "value": 20 } preflight → BLOCK risk_score → 92/100 reason → "High-value customer impact exceeds policy threshold"
Interactive sandbox

See the gate make a decision

Choose an agent action. The demo calculates blast radius, sensitivity and policy outcome before any execution step is allowed.

Proposed agent action
Connected to the live Xano API. If the API is temporarily unavailable, the demo falls back to deterministic sandbox logic so the investor flow remains testable.
Decision

Ready to simulate

WAITING

Choose a scenario and run Preflight to see the policy decision.

Security controls

One narrow job, done before execution

The wedge is deliberately simple: evaluate the action before it reaches production.

◌

Impact simulation

Estimate affected records, target segment and expected changes before execution.

◇

Risk gate

Combine blast radius, data sensitivity and action type into a policy decision.

✓

Human approval

Require an accountable checkpoint for actions that cross a configurable threshold.

≡

Audit trail

Preserve proposed action, decision, approver and execution state for later review.

Integration surface

A simple API contract

Designed to sit between an agent orchestration layer and the system that performs the real business action.

Request

POST /simulate { "agent_id": "billing-agent-04", "operation": "apply_discount", "target": {"segment":"paid"}, "params": {"percent":20}, "requested_by": "agent" }

Decision

{ "decision": "block", "risk_score": 92, "affected_count": 50, "policy": "high_value_change", "requires_approval": false, "reason": "Impact exceeds allowed threshold" }
What the MVP proves

Enough to validate the control point

3 outcomes

Allow, require approval, or block before execution.

5 endpoints

Seed, customers, simulate, approve and execute in the current Xano MVP.

1 audit chain

Proposal → simulation → decision → approval → controlled execution.