VERDICTLAYER

Human oversight for AI

Infrastructure for responsible AI

The human oversight
layer for AI systems

As AI agents take on more autonomous capabilities, the question isn't whether to use AI — it's how to keep humans meaningfully in control. VerdictLayer builds the infrastructure that makes that possible.

Start with HumanLatch Open source on GitHub

Risk scoringDeterministic

Policy engineGlob-pattern rules

Audit trailEvery decision

Self-hostedApache 2.0

Why VerdictLayer

AI is getting faster.
Human oversight can't be an afterthought.

Today's AI agents can write code, execute infrastructure changes, process refunds, control physical systems, and make decisions that affect real people — all in milliseconds. The systems built around them weren't designed for that speed.

VerdictLayer builds the infrastructure layer between AI capability and real-world execution. Not to slow AI down — but to give humans the control points they need to stay meaningfully in the loop.

Outcomes per action

Auto-approve, require human review, or block — driven by policy.

100

Point risk scale

Deterministic scoring across environment, action type, and target.

∞

Domain support

Cloud ops, customer support, robotics, manufacturing, and any AI system.

Missed decisions

Every approval, rejection, and auto-decision written to an audit trail.

Products

Built for the agentic era

We start where the risk is highest: approving or blocking actions before AI agents execute them.

HumanLatch

Now available

Human approval control plane for AI agents

HumanLatch intercepts capability proposals from AI agents before they execute, evaluates risk and workspace policy, routes to humans when needed, and writes an immutable audit record of every decision.

⚡

One API call

Propose an action, get a verdict in milliseconds.

🎯

Policy engine

Glob-pattern rules auto-approve, require approval, or block.

🧑‍💻

Approval dashboard

Review, approve, or reject pending actions from any device.

📋

Audit trail

Every decision logged with who, what, when, and why.

Get started free Learn more Self-host

agent.ts

// 1. Propose the action
const verdict = await humanlatch
  .proposeAction({
    action_type: "aws.iam.policy_change",
    target: "arn:aws:iam::123:role/prod",
    summary: "Attach AdministratorAccess",
    context: {
      environment: "production",
      requested_by: "agent:infra-bot"
    }
  })

// 2. Act on the verdict
if (verdict.status === "approved_auto") {
  await applyIAMChange()
} else if (verdict.status === "pending_approval") {
  await waitForHuman(verdict.id)
} else {
  throw new Error("Blocked by policy")
}

Coming 2025

VerdictLayer Signals

Real-time behavioral anomaly detection across agent fleets. Know when an agent deviates from expected patterns before it causes damage.

Coming 2025

VerdictLayer Policies

Organization-wide policy as code. Define what AI is and isn't allowed to do across all systems, tools, and teams in one place.

In planning

VerdictLayer Audit

Compliance-grade audit exports, AI decision timelines, and forensic tooling for understanding what your agents did and why.

Use cases

Wherever AI acts, humans can decide

☁️

Cloud & DevOps

Stop infra bots from having blind autonomy

Require human approval before IAM changes, destructive Terraform applies, credential rotation, and production-impacting operations.

aws.iam.policy_change

💬

Customer Support & AI

Let AI help customers without letting it overreach

Route high-value refunds, account changes, sensitive escalations through policy and human review when thresholds are crossed.

support.refund.issue

🦾

Robotics

Add human checkpoints before real-world motion

Pause robot motions, safety overrides, or maintenance commands until an operator explicitly approves the request.

robot.motion.execute

🏭

Manufacturing

Control production line changes with policy

Escalate machine-state overrides, quality threshold changes, and line-control decisions before AI can act on them.

manufacturing.line.override

How we build

Principles behind the layer

Non-invasive by design

HumanLatch doesn't replace your execution engine, orchestrator, or existing tools. It adds a single decision gate — one function call — that sits between intent and action.

Deterministic, not probabilistic

Risk scores are computed from explicit rules, not model outputs. You can read the scoring logic, understand why a verdict was reached, and audit it six months later.

Audit-first architecture

Every decision — auto-approved, approved by a human, rejected, or blocked — is written to an immutable audit record with actor, timestamp, and decision note.

Open source at the core

HumanLatch Community Edition is Apache 2.0 licensed. Self-host it in a regulated environment, air-gapped network, or private cloud with no telemetry.

Policy as structured data

Workspace policies are first-class objects — not inline code, regex hacks, or prompt instructions. They're versioned, testable, and auditable.

Human in the loop, not in the way

Low-risk actions auto-approve immediately. Only the decisions that actually need a human reach the approval queue. The bottleneck is risk, not policy.

Pricing

Start free, scale as you grow

HumanLatch Community Edition is free and open source. Upgrade to managed hosting when you need teams, notifications, or compliance-grade retention.

Community

Free

Self-hosted forever

The full HumanLatch core, self-hosted in your own infrastructure. Apache 2.0 licensed.

Full approval control plane
Policy engine (glob-pattern rules)
Audit trail (14-day retention)
API keys & dashboard
Single workspace
Docker Compose deploy

Self-host on GitHub

Keep humans in the loop

Start with HumanLatch. Self-host the community edition free, or get on cloud in minutes. Have an enterprise use case? Get in touch.

Start free Self-host

Or email us directly at hello@verdictlayer.com

Talk to us

Enterprise pricing, custom deployments, or general questions.

The human oversightlayer for AI systems

AI is getting faster.Human oversight can't be an afterthought.