ActionGuard AI

Inspiration

AI agents can call purchasing, vendor, customer-data and operations APIs in seconds, but most approval software was designed for people filling forms and waiting in ticket queues. ActionGuard AI closes that gap by checking an agent's intended action before it becomes a real-world consequence.

What it does

An agent sends its intended business action to ActionGuard before calling the target API. A deterministic engine evaluates tenant, identity, permissions, amount, approved-provider status, record volume and sensitive fields, then returns one of three explicit outcomes:

• ALLOW — execute the synthetic effect exactly once and record the evidence.
• REVIEW — pause with no side effect until an authorized human approves or rejects it.
• DENY — block the action and explain the matched rule.

The demo proves the complete control model with three synthetic scenarios: a USD 120 approved software renewal is allowed; a USD 7,500 payment to a new vendor containing bank data requires human review; and an unauthorized export of 10,000 customers is denied.

Policy Studio turns plain-language governance into an editable policy draft. AI is advisory: it cannot publish a policy, approve an action, override DENY, or execute an effect. If AI is unavailable, the deterministic fallback keeps the product usable and honest.

How we built it

The operations console uses React, TypeScript and Vite. Xano is the central backend, not a decorative proxy: it hosts tenant-scoped data, authentication, deterministic policy evaluation, idempotent action processing, approval workflows, sensitive-field masking and persisted SHA-256 audit evidence.

The repository also includes a dependency-light Node reference server, OpenAPI specification, synthetic dataset, XanoScript source, architecture and security documentation, plus unit, contract and integration tests. Browser authentication stores the Xano bearer token only in session storage and fails closed if the session is missing, expired or rejected.

Challenges we ran into

Safety controls can fail through duplicate requests, approval races, cross-tenant reads, sensitive logging, incorrect timestamp serialization, or unavailable AI. We addressed those risks with unique tenant-qualified idempotency keys, transactional single approval, authenticated-user tenant lookup, masking before persistence, evaluated execution timestamps, deterministic fallback and independent hash-chain recomputation.

Accomplishments that we're proud of

• One reusable API produces three visibly different safety outcomes.
• The USD 7,500 payment cannot execute before human approval.
• The unauthorized 10,000-customer export never executes.
• Repeated idempotency keys return the original action and execution.
• Cross-tenant evidence access is rejected.
• Bank fields are redacted before display and persistence.
• Every lifecycle transition is linked with SHA-256 evidence.
• The full verification command passes 25 automated tests and validates all 15 versioned XanoScript artifacts.

What we learned

Business AI safety is less about generating better text and more about controlling state transitions. Deterministic policy, identity, idempotency, human judgment and tamper-evident evidence must work together.

What's next for ActionGuard AI

Signed policy bundles, procurement and CRM connectors, configurable separation of duties, external audit anchoring and policy simulation before activation.

Try it out

• Live demo
• Source and technical evidence

Built With

• codex
• crypto
• node.js
• openapi
• react
• typescript
• vite
• vitest
• web
• xano
• xanoscript