HUMANLATCH by VerdictLayer

Human approval control plane for AI actions from VerdictLayer

Safe autonomy for AI systems

Let AI agents operate without giving them unsafe autonomy.

HumanLatch is a control plane that sits between AI and real-world action. Agents propose capabilities, policies decide the route, humans approve when needed, and every decision lands in an audit trail.

Agents propose actions before they execute

Policies auto-approve, require approval, or block

Humans approve from dashboard, API, and future channels

Every decision is written to an audit trail

Approval runtime

HumanLatch control plane

Step 1

Agent proposes a capability

A bot, copilot, robot controller, or machine optimizer sends a proposed action before it executes.

Step 2

Policy decides the route

HumanLatch evaluates risk and policy to auto-approve, require approval, or block the request.

Step 3

Humans approve only when needed

Approvers review the risky capability in a dashboard today, with Slack or API channels layered on later.

Decision example

manufacturing.line.override

AI requested a safety-threshold override on line 3.

RouteNeeds approval

What HumanLatch Does

A control plane for AI capabilities, not just another dashboard.

Developers integrate HumanLatch into their own apps and systems. Before a capability executes, they propose it to HumanLatch. HumanLatch evaluates policy, determines whether a human is required, and records the decision.

Developers

Integrate once

Ship the control plane behind your own chatbot, robot, industrial controller, or cloud agent.

Operators

Review with confidence

Understand what the AI wants to do, why policy escalated it, and whether to approve or deny.

Compliance

Record every decision

Keep an audit trail of proposed actions, policy outcomes, and human decisions for replay and review.

Industries

Useful anywhere AI can trigger a sensitive capability.

HumanLatch is domain-agnostic. The same control plane works for digital actions, customer operations, robotics, and physical industrial systems.

Cloud and DevOps

Stop infra bots from having blind autonomy

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

Customer Support and Chatbots

Let AI help customers without letting it overreach

Route refunds, account changes, escalations, or sensitive messages through policy and human review when thresholds are crossed.

Robotics and Physical Systems

Add human checkpoints before real-world motion

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

Manufacturing and Industrial AI

Control production line changes with policy

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

Developer Flow

Works with self-hosted CE or a remote HumanLatch API.

Developers keep their own runtime and execution engine. HumanLatch becomes the decision service that sits in front of execution.

Self-hosted CE

Run the dashboard and API in your own Docker environment for internal bots, private factory networks, or regulated environments.

Remote hosted API

Point your app at a managed HumanLatch endpoint and use the same integration model without hosting the control plane yourself.

Example integration

const decision = await humanlatch.proposeAction({
  action_type: "manufacturing.line.override",
  target: "line-3",
  summary: "Override conveyor safety threshold",
  payload: { threshold: 0.92, duration_seconds: 180 },
  context: {
    environment: "production",
    requested_by: "agent:line-optimizer",
    domain: "manufacturing"
  }
})

if (decision.status === "approved_auto") {
  await executeOverride()
} else if (decision.status === "pending_approval") {
  await waitForApproval(decision.id)
} else {
  throw new Error("Blocked by policy")
}

Next step

Give your agents a safe path to act in the real world.

Use the hosted control plane, read the docs, or start with self-hosted CE. The product model stays the same: propose, route, approve if needed, then execute.