The core problem

Permission is not
the same thing as safe execution.

An agent having access does not mean every action should execute under every context. InterAI adds a decision boundary where a mistake becomes consequential; it does not replace existing permissions.

✓ Authorized

but unusually large

Proposed action

send_vendor_payment

$250,000production

Risk indicatorAmount exceeds configured limit
Policy review required

✓ Authorized

but irreversible

Proposed action

delete_customer_data

production

Risk indicatorIrreversible action detected
No recovery path

✓ Authorized

but wrong context

Proposed action

modify_pricing_rules

staging

Risk indicatorContext mismatch
Staging vs. production

Agent Has access

The tool is available.

Proposed action Agent proposes an action

Amount, environment, confirmation, tool identity, and policy context.

Policy checkpointInterAI evaluates context before execution

EXECUTE

Safe to run

REVIEW

Human-in-the-loop

BLOCK

Risk policy triggered

An example decision

A contract-aligned illustration of the checkpoint before an external side effect occurs.

AI

vendor_payment_agent Illustrative request

Actionrelease_vendor_payment

Amount$125

Environmentproduction

◔ REVIEW REQUIRED

Human review required before execution

External side effect
User confirmation required
Production environment

Risk score

0.42

example receipt tr_01JZPUBLICEXAMPLE

Public contract fields

Where it fits

PAY

Payments

Transfers, payouts, vendor payments.

WEB

Browser actions

Web navigation, form submissions.

InterAI Checkpoint

Policy • Review • Control

CRM

Email / CRM

Send emails, update records, create tasks.

DB

Database writes

Insert, update, delete records.

MCP

MCP tools

Tool calls with external impact.

Operating principles Agent-controlled Policy-first Verifiable receipts

Developer integration

One request between
intent and execution

Verify is the primary decision endpoint of InterAI Risk Oracle. Call it before execution, act on recommended_action, and store the returned receipt. Use HTTP, the registry-published MCP remote, or your own client layer while preserving the same public contract.

Supported interfaces

HTTPexample below MCPremote server SDKclient layer A2Aagent discovery

HTTP requestPOST /verify

POST https://ai-risk-oracle.fly.dev/verify
Authorization: Bearer {YOUR_API_KEY}
X-Idempotency-Key: vendor-payment-001
Content-Type: application/json

{
  "use_case": "agent-before-payment",
  "action": {
    "type": "payment",
    "name": "release_vendor_payment",
    "description": "Release payment to a vendor agent",
    "amount_usd": 125,
    "currency": "USD",
    "irreversible": false,
    "external_side_effect": true
  },
  "context": {
    "environment": "production",
    "counterparty_id": "vendor_agent_456",
    "user_confirmation": false
  },
  "policy": {
    "max_risk_level": "medium",
    "require_trust_receipt": true,
    "amount_usd_limit": 500
  }
}

Response excerptJSON

{
  "decision_id": "tr_01JZPUBLICEXAMPLE",
  "request_contract": "autonomous_execution",
  "score": 0.42,
  "risk_level": "medium",
  "signals": {
    "has_external_side_effect": true,
    "is_irreversible": false,
    "involves_money": true,
    "amount_usd": 125,
    "requires_user_confirmation": true,
    "environment": "production"
  },
  "recommended_action": "review_required",
  "policy_result": "review_required",
  "policy_violations": [],
  "trust_receipt_id": "tr_01JZPUBLICEXAMPLE"
}

DecisionRouting

ALLOW

Action permitted.

REVIEW

Human-in-the-loop review required.

BLOCK

Action blocked. Risk policy triggered.

30-day design partner pilot

ONE ACTION.

Choose one consequential boundary. We’ll wire it with you. Keep the first pilot narrow, measurable, and useful.

Start a design-partner pilot View docs

Remote MCP A2A discovery Signed receipts Live API

Permission is notthe same thing as safe execution.

but unusually large

but irreversible

but wrong context

An example decision

One request betweenintent and execution

ONE ACTION.

Permission is not
the same thing as safe execution.

One request between
intent and execution