In the Weeds: Human-in-the-Loop AI with gotoHuman MCP

The Trust Problem

Here's the scenario that keeps AI engineers up at night: an AI agent processes a customer refund request, decides it's legitimate, and issues the refund. Except the request was fraudulent, the refund was $50,000, and nobody reviewed it because the automation was supposed to handle everything.

Fully autonomous AI is efficient. It's also terrifying for high-stakes operations. The solution isn't to remove AI — it's to build systems where AI does the work but humans make the critical decisions. This is called "human-in-the-loop" (HITL), and ggotoHuman MCP is a purpose-built tool for implementing it.

What gotoHuman Does

gotoHuman is an approval workflow platform designed specifically for AI systems. When your AI agent encounters a decision that exceeds its authority — a large transaction, a sensitive communication, a destructive action — it pauses and sends the decision to a human reviewer. The human approves or rejects. The AI continues accordingly.

The MCP server makes this available to your AI assistant. Instead of building custom approval logic for every workflow, you get a standardized pattern: AI works → hits a decision gate → human reviews → AI continues.

The Core Pattern

Every gotoHuman interaction follows this flow:

AI Agent doing work
    ↓
Decision point reached (exceeds AI authority)
    ↓
ggotoHuman MCP sends approval request
    ↓
Human reviewer receives notification
    ↓
Human reviews context and decides (approve/reject/modify)
    ↓
Decision returned to AI agent
    ↓
AI continues based on human decision

The key innovation is the context package. When gotoHuman sends an approval request, it includes everything the human needs to make a good decision: what the AI is trying to do, why, what data informed the recommendation, and what the consequences are. The reviewer doesn't need to reconstruct the entire workflow — the context is delivered.

Implementation Patterns

Financial Operations

This is the most obvious use case. Any AI system that touches money should have human gates:

Invoice approval:
- AI receives an invoice, validates it against purchase orders and contracts
- AI recommends approval or flags discrepancies
- gotoHuman sends the recommendation to a finance manager
- Finance manager reviews and approves/rejects
- AI processes the payment or flags for further review

Refund processing:
- AI evaluates a refund request against policy criteria
- AI recommends refund amount and reason code
- For refunds above threshold (say, $500), gotoHuman triggers
- Manager reviews the specifics and decides
- AI processes the approved refund through PPayPal MCP or your payment processor

Subscription changes:
- Customer requests enterprise plan upgrade through RRevenueCat MCP
- AI validates eligibility and prepares the change
- gotoHuman routes to sales for approval (ensuring custom pricing is honored)
- Sales confirms, AI executes the upgrade

Content and Communication

AI-generated communications carry brand risk. Human review for important communications:

Customer-facing emails:
- AI drafts response to a customer complaint
- For VIP customers or sensitive topics, gotoHuman triggers
- Customer success manager reviews the draft
- Approved draft is sent; rejected draft is revised

Social media posts:
- AI generates post content based on content calendar
- gotoHuman sends to marketing manager for approval
- Manager approves, modifies, or kills the post
- Approved content is published through automation

Public statements:
- Any AI-drafted content that goes on your website, in a press release, or to investors
- Always goes through human review
- No exceptions

Infrastructure and DevOps

AI agents managing infrastructure need hard limits:

Deployment approvals:
- AI prepares a production deployment based on CI/CD pipeline results
- gotoHuman sends deployment plan to on-call engineer
- Engineer reviews changes, test results, and rollback plan
- Approved deployment proceeds; rejected deployment is held

Scaling decisions:
- AI recommends scaling up infrastructure based on load patterns
- For changes above cost threshold, gotoHuman triggers
- SRE reviews the recommendation and approves
- AI executes the scaling action

Data operations:
- Any AI operation that modifies or deletes production data
- Always gates through human approval
- The context package includes exactly what will change and why

Designing Good Approval Workflows

The art of HITL design is knowing where to put the gates. Too few gates and you have unchecked AI. Too many gates and you've recreated the manual process with extra steps.

Principles for Gate Placement

Irreversibility. If an action can't be easily undone, gate it. Sending an email can't be unsent. Deleting data can't be undeleted. Processing a payment is hard to reverse. These get gates.

Magnitude. Small actions can be automated. Large actions need review. A $5 refund? Auto-approve. A $5,000 refund? Human review. The threshold depends on your risk tolerance.

Novelty. AI handles common patterns well. Unusual situations — edge cases, first-time events, anomalies — should be routed to humans. "This refund request doesn't match any common pattern" → human review.

Audience sensitivity. Internal draft? Maybe auto-approve. Email to your biggest client? Human review. Public statement? Absolutely human review.

Context Packages

The approval request is only as good as the context it provides. A good context package includes:

• What: The specific action being proposed
• Why: The reasoning behind the recommendation
• Data: The relevant information that informed the decision
• Risk: What could go wrong
• Alternatives: Other options considered
• Urgency: How time-sensitive the decision is

Bad context: "Approve refund of $2,000?"
Good context: "Customer Jane Doe (5-year customer, $50K lifetime value) requested refund for Order #12345 ($2,000). Reason: product defective. Order was 15 days ago, within 30-day policy. Customer has 2 previous refund requests in 5 years (both legitimate). AI recommendation: approve full refund. Risk: low (consistent with policy, loyal customer)."

The second version lets the reviewer make an informed decision in 30 seconds. The first version requires them to go dig up the context themselves, defeating the purpose.

Combining with Other MCP Servers

gotoHuman is a horizontal capability that enhances any AI workflow:

• Nn8n MCP: Build n8n workflows with gotoHuman gates. The automation handles the routine; humans handle the critical.
• KKeboola MCP: Data pipeline changes that affect production analytics? Human review before execution.
• SSlack MCP: Route approval requests to a dedicated SSlack channel. Reviewers approve directly from Slack.
• DDrand MCP: For selection processes, use drand for verifiable randomness + gotoHuman for final human sign-off on results.

The Philosophy

The goal isn't to distrust AI. It's to build systems that are trustworthy. Trust comes from appropriate oversight, not blind faith.

The best AI systems in production today are human-in-the-loop by design. Not because the AI isn't capable, but because accountability requires a human being willing to say "yes, this is the right decision." Until we solve the accountability problem (legally, ethically, practically), human-in-the-loop isn't a compromise — it's the correct architecture.

ggotoHuman MCP makes this architecture easy to implement. Build the automation. Place the gates. Let AI handle the volume. Let humans handle the judgment. That's the system that works.