Architecture
Human Approval Gateway
AI that shows you exactly what it plans to do and waits for your approval before acting.
The Business Problem
"The AI posted WHAT on our company social media?" "Who approved sending this email to 50,000 customers?" "Why did the system deploy to production without anyone reviewing the changes?"
Every enterprise that's deployed AI for real-world actions has a horror story. The AI did something it shouldn't have -- published content, sent communications, executed transactions, or modified infrastructure -- and no one saw it coming because no one saw it at all.
The problem isn't that the AI is wrong most of the time -- it's that when it's wrong that one time, the consequences are severe. A single bad social media post can go viral. A single wrong email to your customer list can't be unsent.
How It Solves It
Human Approval Gateway sandboxes every AI action, produces a detailed preview, and gates execution on explicit human approval.
Simplified Flow
Generate Action
Execute in Sandbox
Display Preview
Human Approves/Rejects
Execute or Log
The AI generates a candidate action (a post, an email, a deployment, a transaction). It runs in sandbox mode -- fully rendered but not live. The reviewer sees exactly what will happen: the content, the recipients, the affected systems, the expected outcome. If approved, it executes. If rejected, it logs the decision and exits cleanly.
Key Capabilities
Full sandbox preview
Every action is fully rendered in sandbox mode before execution, showing exactly what will happen
Human approval gate
Nothing executes until a human explicitly approves
Rejection logging
Rejected actions are logged with reviewer reasoning, building a training set for future improvement
Configurable review rules
Route different action types to different reviewers (e.g., content to marketing, deploys to SRE)
Batch approval
Review and approve multiple actions at once when appropriate
Audit trail
Complete record of what was proposed, who reviewed it, and what the decision was
Industry Applications
Media & Publishing — Content Publishing
AI generates social media posts, blog articles, or press releases. Editors preview the exact content -- including formatting, images, and hashtags -- before any content goes live.
Financial Services — Transaction Approval
AI prepares wire transfers, trades, or account modifications. Officers preview the full transaction details and approve or reject. Transactions above thresholds require additional approvers.
Technology & SaaS — Code Deployment
AI generates deployment plans with infrastructure changes, database migrations, and code updates. SRE reviews the deployment diff, canary plan, and rollback strategy before execution.
Legal — Document Filing
AI prepares court filings, regulatory submissions, or contract amendments. Partners preview the exact documents, check for accuracy and completeness, and approve before submission.
Ideal For
- • Any action with real-world consequences that can't easily be undone
- • Publishing, communications, and public-facing AI outputs
- • Financial transactions above configurable thresholds
- • Regulated industries where human oversight is required
Consider Alternatives When
- • The action is low-risk and high-volume -- human approval becomes a bottleneck
- • The AI has demonstrated sufficient reliability for full automation
- • You need the AI to self-assess risk rather than requiring human review of everything (use Self-Aware Safety Agent)
- • The task is purely analytical with no real-world actions (no approval gate needed)
Human Approval Gateway vs. Self-Aware Safety Agent
Approval Gateway requires human review of every action. Safety Agent decides autonomously which actions need human review. Think of Approval Gateway as a mandatory sign-off policy, and Safety Agent as a senior employee who knows when to escalate.
| Human Approval Gateway | Self-Aware Safety Agent | |
|---|---|---|
| Human involvement | Always -- every action reviewed | Selective -- only uncertain/risky actions |
| Throughput | Limited by human reviewer capacity | High -- routine actions are autonomous |
| Safety guarantee | Maximum -- nothing bypasses review | High -- depends on confidence calibration |
| Best alone for | High-consequence actions | High-volume mixed-risk queries |
Implementation Overview
Typical Deployment
2-4 weeks
Integration Points
Action execution systems (publishing platforms, trading systems, deployment pipelines), reviewer notification channels
Data Requirements
Action type definitions, reviewer routing rules, approval threshold configurations
Configuration
Sandbox rendering, reviewer assignments, escalation rules, batch approval settings
Infrastructure
Sandbox environment for action preview; notification system for reviewers
Get Started