👤 Human Review Gates

BaC Principle: Humans Direct, Agents Execute

In a Business as Code workflow, humans are not removed — they are repositioned. Instead of spending 10 hours/week writing emails and researching leads, you spend 45 minutes/week reviewing what agents produced and approving what’s next.

Manifest Stages: S3, S5 → See 01-Process-Manifest

There are two human gates in this process. Both are deliberate checkpoints — not exceptions or escalations. They are where your judgment shapes the entire batch.

---

🚦 Gate Overview

Gate	Stage	Trigger	Time Required	Decision
[[#Gate 1 — Lead Batch Review|Gate #1]]	S3	After 03-Lead-Sourcing-Agent + 04-Research-Agent complete	~15 min	Approve / Prune / Flag
[[#Gate 2 — Email Template Review|Gate #2]]	S5	After 06-Personalization-Agent completes	~10 min	Approve sample / Edit / Pause

---

🔵 Gate 1 — Lead Batch Review

Triggered by: 03-Lead-Sourcing-Agent + 04-Research-Agent completion Frequency: Twice per week (Mon + Thu) Time required: ~15 min Outcome feeds: 06-Personalization-Agent

What You Review

You receive a sorted list of ~50 leads with:

• Score (0–100), score tier (hot/warm/cold)
• Key ICP match signals
• Research summary (top 2 signals from 04-Research-Agent)
• Disqualification flags

Your Decision Per Lead

✅ APPROVE    → include in this batch, proceed to personalization
❌ REJECT     → remove, log reason (feeds [[09-Metrics-and-Self-Improvement]])
⏸️ DEFER      → hold for next batch (wrong timing, not enough context)
🚩 FLAG       → something the agent shouldn't have included — note why

Decision Heuristics

What to look for in 15 minutes

You’re not re-researching every lead. You’re pattern-matching for things agents miss:

• Competitor — you know your space, agents don’t always catch this
• Bad-fit company culture — sometimes firmographics match but you know it won’t work
• Personal connection — someone you know should be handled manually
• Recent context — you heard about this company at an event or on a podcast

Approval Thresholds

gate_1_rules:
  minimum_approval_rate: 0.65          # if you approve <65% of a batch, trigger ICP review
  maximum_rejection_rate: 0.35         # >35% rejection = [[02-ICP-Definition]] needs update
  auto_escalate_if:
    - batch_size < 20                  # sourcing agent produced too few leads
    - avg_score < 60                   # quality dropped significantly
  action_on_escalate: pause_and_alert_COO

If You Reject >35% Consistently

This is a signal that 02-ICP-Definition is miscalibrated, not that agents are failing. Trigger an ICP review via the improvement loop.

---

🟢 Gate 2 — Email Template Review

Triggered by: 06-Personalization-Agent completion on approved batch Frequency: Twice per week (Mon + Thu, 2–4 hours after Gate #1) Time required: ~10 min Outcome feeds: 05-Email-Sequence-Engine

What You Review

You see a 10% random sample of personalized emails, plus any auto-flagged emails:

• Full email (subject A + B, body)
• Hook type used
• Confidence score
• Lead context summary (why this hook was chosen)

Your Decision

✅ APPROVE SAMPLE    → release entire batch to sequence
✏️ EDIT + RESAMPLE   → fix template/prompt, re-generate 10% and review again
⏸️ PAUSE BATCH       → something systemic is wrong, investigate before sending
🚩 FLAG EMAIL        → this specific email should not send (remove from batch)

What “Good” Looks Like

Approve if all of these are true:

• Hook references a real, specific signal (not a guess or assumption)
• Email reads like a human wrote it (no AI tells: bullet lists, excessive politeness)
• The ask is clear and single (not multiple CTAs)
• Subject line is curious, not salesy
• Word count is 80–120 words

Edit or Pause if any of these are true:

• Hook is generic despite agent claiming “high confidence”
• Email is >140 words
• Subject line is a question with a ”?” (lower deliverability)
• More than 30% of sample uses the fallback hook template

Template Change Protocol

If you edit an email manually during Gate #2:

template_change_flow:
  1: Edit the specific email → approve manually
  2: If it reflects a systemic issue:
     - Note the issue in [[09-Metrics-and-Self-Improvement]]
     - Propose a prompt update to [[06-Personalization-Agent]]
     - Tag for A/B test before deploying to full batch
  3: Never change the prompt directly — always go through [[09-Metrics-and-Self-Improvement]] loop

---

🔁 Feedback Loop to Agents

Every rejection and flag at both gates produces structured feedback:

{
  "gate": "gate_1",
  "lead_id": "lead_042",
  "decision": "REJECT",
  "reason": "competitor",
  "agent_source": "LeadSourcingAgent",
  "feedback_category": "ICP_gap",
  "timestamp": "2026-04-01T10:15:00Z"
}

This data flows to the improvement loop and accumulates into:

• ICP filter updates
• Lead scoring recalibrations
• Personalization prompt improvements

---

📅 Weekly Human Time Budget

Monday:
  09:00 — Gate #1: Lead Batch Review     (~15 min)
  13:00 — Gate #2: Email Template Review (~10 min)

Thursday:
  09:00 — Gate #1: Lead Batch Review     (~15 min)
  13:00 — Gate #2: Email Template Review (~10 min)

Friday:
  Weekly metrics review                  (~20 min)
  → [[09-Metrics-and-Self-Improvement]]

Total: ~70 min/week (incl. metrics review)

This is the entire human involvement in the process.

Everything else runs autonomously.

---

📎 Related Files

• 01-Process-Manifest — Stages S3 + S5 definition
• 03-Lead-Sourcing-Agent — Output reviewed at Gate #1
• 04-Research-Agent — Output reviewed at Gate #1
• 06-Personalization-Agent — Output reviewed at Gate #2
• 02-ICP-Definition — Updated when Gate #1 rejection rate is high
• 09-Metrics-and-Self-Improvement — Receives all gate feedback