Human-in-the-Loop Automation

Operations teams run on inbound volume — support tickets, sales leads, public reviews, new-account onboarding, staff policy questions. Most of it is routine and could be handled in seconds. A small slice is genuinely high-stakes: a refund dispute, a safety incident, an enterprise contract, a one-star review naming a guest.

The naive fix — let AI answer everything — is unsafe, because a confident wrong answer lands exactly on that high-stakes slice. The opposite — make a person read everything — spends the scarce resource on cases that never needed it. These kits are built around that tension: absorb the routine volume automatically, and guarantee the high-stakes cases reach a human.

One rule holds across all five: AI drafts, a human approves anything risky — never the reverse. And the escalation decision is not left to the model's discretion alone. Each workflow runs a two-layer gate. First the model returns its own requires_human_approval flag inside strict JSON. Then deterministic code re-checks the case against kit-specific rules and can force escalation regardless of what the model said.

The model is allowed to be cautious; it is never the only thing standing between an automated action and a guest, a dollar, or a safety event.

Every kit is the same runnable spine — five n8n nodes, no cloud API keys, executable on a laptop:

• Trigger — a sample case enters the workflow.
• Load Inputs— the case is paired with the kit's system prompt.
• Local LLM — Ollama (phi4:14b) returns strict JSON (json mode, temperature 0.2).
• Parse + HIL Gate — the JSON is parsed, then the kit-specific rule decides auto-resolve vs. human approval.
• Result — a structured object plus the human-approval flag.

The same five kits also ship a design-exact production graph that swaps the local model for real channels and models — covered below.

Five kits, one spine. Each classifies, scores, or drafts in its domain and carries its own escalation rule. Every example below is a verified sample run from the local build.

Alongside each runnable kit is a design-exact production scaffold — an importable n8n graph that is not yet run with live data. They share a common shape:

• Multi-channel intake — webhook, email/IMAP, schedule, or Telegram.
• Normalize — dedup, canonical schema, and consent / lawful-basis checks.
• Cheap classifier → premium drafter — cost-tiered models, with citation-enforced retrieval (RAG) where there is a knowledge base.
• Force-HIL validation — the deterministic gate, in code.
• Mandatory approval — a send-and-wait step in Slack or email, with an approve / redraft / escalate router; execution only fires after approval.
• Audit + recovery — an immutable log with cost tracking, a dedicated error sub-workflow, and feedback capture for continuous improvement.

Productionizing a kit means swapping the local-Ollama node for its real channels and model, adding credentials, and feeding sanitized data.