AI agents make promises
your company has to keep.

GreenLightz evaluates every AI-generated commitment against your policies in real time — before it reaches the customer.

Monetary commitment
AI Agent"Issue $150 refund"
GreenLightzPolicy evaluation
ResponseAMBER — reduce to $75
Non-monetary commitment
AI Agent"365-day uptime SLA"
GreenLightzPolicy evaluation
ResponseAMBER — max 180 days

The problem with autonomous AI

AI agents are making financial commitments on your behalf. Without governance, every interaction is an uncontrolled risk.

Unauthorized Commitments

Your AI agent promised a full refund, free shipping, and an extended warranty — all in one conversation. Nobody approved it.

Revenue Leakage

Discounts and credits pile up silently. By the time finance notices, thousands have leaked through unapproved agent actions.

Compliance Risk

Regulators ask: 'Who authorized this?' With no audit trail, you can't answer. Every untracked commitment is a liability.

What changes when governance is real-time

Your AI agents stay fast. Your finance team stays confident. Your compliance team gets evidence they can actually use.

Block unauthorized exposure

Every agent action is evaluated against financial ceilings, velocity limits, and aggregate patterns before it reaches your customer.

Zero approval bottlenecks

Routine actions auto-approve in under a millisecond. Borderline actions auto-adjust within policy bounds. Only hard violations need a human.

Audit-ready from day one

Every verdict produces a signed evidence packet — what was evaluated, which rules fired, and what the agent did next. Immutable and queryable.

Policy changes, not code changes

Swap a YAML file to change governance behavior. Different thresholds per tenant, per action type, per approval mode — no deployment required.

Three steps to governed AI

No agent framework dependency. No training data required. Pure policy evaluation at sub-millisecond latency.

Step 01

Integrate

One POST request before each agent action. No SDK, no sidecar, no agent framework lock-in. Any language that speaks HTTP can integrate in minutes.

POST /gateway/evaluate
{
  "action_type": "credit_or_refund.issue",
  "tenant_id": "acme_corp",
  "actor_id": "agent-7b",
  "target_id": "customer-4492",
  "amount_cents": 15000,
  "currency": "USD",
  "reason": "Product arrived damaged",
  "correlation_id": "conv-8812-msg-3"
}
Step 02

Define Policies

Set per-action-type amount ceilings, pre-authorization thresholds, and approval modes in YAML. Policy packs are version-controlled, tenant-scoped, and hot-swappable without downtime.

policy.yml
# Per-tenant policy configuration
max_amount_cents:              # Per-action-type ceilings
  credit_or_refund.issue:      # Refund ceiling
  discount.offer:              # Discount ceiling
  delivery_promise.commit:     # Shipping commitment ceiling

pre_authorized_limit:          # Auto-approve threshold
approval_mode:                 # manual | pre_authorized | block

enabled_dimensions:            # Active governance dimensions
  - financial_impact
  - policy_compliance
  - behavioral_pattern

policy_version:                # Version tag for drift detection
webhook_url:                   # Approval notification endpoint
webhook_secret:                # env-ref (secrets never inline)
Step 03

Govern in Real-Time

Every action receives a deterministic verdict with a signed evidence packet. Amber verdicts include an intervention plan — the engine tells your agent exactly how to modify the action and retry within policy bounds, with no human in the loop.

200 OK — Response
{
  "verdict": "REQUIRE_APPROVAL",
  "reasons": [
    "Amount $150 exceeds pre-authorized limit $75",
    "30-day customer aggregate: $420"
  ],
  "evidence_ref": "ev-8a3f...",
  "evidence_hash": "sha256:b94d...",
  "signed": true,
  "intervention_plan": {
    "action": "modify_and_retry",
    "band": "amber",
    "safe_degrade_actions": ["lower_amount"],
    "retry_guidance": {
      "max_retries": 3,
      "stop_condition": "action_allowed_or_max_retries_exhausted"
    }
  }
}

Judgment, not just thresholds

Every action is evaluated across four dimensions simultaneously. Each dimension is a separate veto gate — and borderline actions get modification guidance, not a wall.

Financial thresholds

Per-action-type ceilings, pre-authorization limits, and daily caps.

Velocity patterns

Per-agent and per-customer frequency tracking across rolling windows.

Aggregate exposure

Cumulative value tracking per customer, per agent, and per agent-customer pair.

Behavioral analysis

Stacking detection, escalation patterns, and precedent tracking across turns.

Three possible outcomes

Green
ALLOW

Within policy. Proceed immediately, zero delay.

Amber
REQUIRE_APPROVAL

Exceeds a threshold. Engine returns a concrete modification — agent auto-adjusts and retries.

Red
BLOCK

Hard violation. Signed evidence documents the exact rule. Requires human override.

Most actions resolve as green or auto-adjusted amber. Only hard violations reach red.

Why existing approaches fall short

Common approaches each solve part of the problem. None address all dimensions at once.

Simple rule engines

Single-threshold checks. Amount > $100? Block.

  • No cross-action memory — can't see $420 aggregate this month
  • No velocity tracking — 5 refunds in an hour looks the same as 5 in a year
  • Binary outcome — block or allow, nothing in between

LLM-based guardrails

Ask another LLM whether the action is safe.

  • Non-deterministic by design — same input may produce different verdicts
  • Model inference adds latency to every evaluation in the agent's critical path
  • Verdicts are opinions, not signed evidence — harder to audit at scale

Manual approval queues

Route every flagged action to a human reviewer.

  • Doesn't scale — 100 agents generating 1,000 actions/hour
  • Introduces hours of delay into real-time conversations
  • Reviewer fatigue increases risk of inconsistent decisions over time

GreenLightz is a different category

Multi-dimensional: financial + velocity + aggregate + behavioral — evaluated simultaneously
Deterministic: same input, same verdict, every time. Signed and auditable
Accelerating: amber verdicts include modification guidance — agents auto-adjust, no human needed
Sub-millisecond: policy evaluation, not model inference. Zero latency impact on agent conversations
Framework-agnostic: one HTTP call. No SDK, no sidecar, no agent framework lock-in

Every commitment type, governed

GreenLightz ships with 9 built-in commitment types. Each one is evaluated against your policy pack with full audit trail.

Refund

Full or partial monetary return. Evaluated against per-agent limits and aggregate thresholds.

Discount

Percentage or fixed-amount reduction. Policy-checked for stacking and cumulative exposure.

Delivery Promise

Shipping and delivery date commitments. Velocity-limited and deadline-bounded per policy.

Subscription Extension

Extension of subscription periods. Duration caps and cumulative exposure tracking.

SLA Commitment

Service level agreements on uptime, response time, or resolution targets. Duration and scope governed.

Data Handling

Data retention and processing commitments. Retention period ceilings and compliance-bounded.

Support Response

Response time commitments for support tickets. Floor-bounded to prevent impossible promises.

Policy Exception

Override grants and policy exceptions. Scope-limited with mandatory justification tracking.

Goodwill Remedy

Replacement, credit, or gesture of goodwill. Value-capped and precedent-tracked.

6 built-in guarantees. Always enforced.

These aren't features you toggle on. They're architectural guarantees baked into every code path.

Fail-Closed

Any error, timeout, or ambiguity results in BLOCK — never silent pass-through. The default state is denial.

Escalate-Only

Verdicts only move toward stricter enforcement. A green can become amber or red mid-evaluation, never the reverse.

Deterministic

Identical inputs produce identical verdicts across every run. No sampling, no temperature, no stochastic paths.

Offline-First

The core evaluation engine has zero external dependencies. LLM enrichment is optional and non-blocking.

Privacy-First

Every identifier is HMAC-hashed with per-tenant keys before storage. Zero PII in logs, evidence, or API responses.

Tamper-Evident

Every verdict produces a signed evidence packet with a deterministic content hash. Mutations are cryptographically detectable.

Built to be trusted

A deterministic governance engine with cryptographic evidence trails, built through 93 hardening iterations.

< 1ms
Evaluation latency
P99 under 1 millisecond
6
Architectural invariants
Always enforced, never toggleable
7,900+
Governance test cases
Deterministic, CI-gated on every deploy
0
External dependencies
Core engine works fully offline

What a pilot looks like

2-4 weeks from first API call to production decision. Start in shadow mode with zero risk to live operations.

Day 0

Setup

1-2 hours

Configure your policy thresholds, connect your AI agent with one API call, and validate with a test evaluation.

Week 1

Shadow mode

Observe

GreenLightz evaluates in parallel but doesn't block your agents. See what would be caught — zero impact on live operations.

Week 2-3

Live governance

Govern

Enable real-time governance. Tune thresholds based on Week 1 observations. Monitor approval queue and aggregate patterns.

Week 4

Review

Decide

Structured review: blocked exposure, approval patterns, false positive rate. Decide on production rollout together.

Enterprise security, not enterprise complexity

Every security property is verified by automated tests on every deploy. Not a checklist — a runtime guarantee.

End-to-End Encryption
HTTPS + HMAC-signed evidence
Tamper-Evident Audit Trail
Cryptographically signed verdicts
Deterministic Verdicts
Same input, same output — always
Privacy-First Architecture
All IDs hashed, zero PII in logs
Fail-Closed by Default
On error, actions are blocked — not allowed
Offline-First Engine
No LLM dependency for core governance

Stop your AI from writing checks
you can't cash.

See how GreenLightz brings governance to your AI agents. 30-minute demo, no commitment.

Book a Demo