Bounded Agent Blueprint: Lead‑Qual, Weekly Reporting, and Inbox Triage — SOPs, Schemas, Evals, KPIs

Use this like a Notion playbook. Clone the sections you need, paste the schemas/prompts into your stack, and follow the 14‑day sprint.

14‑Day ship plan:

• Days 1–2: Pick one agent and define its single job + inputs/outputs. Write the ICP and success criteria.
• Days 3–4: Wire the connectors (CRM/Databox/Inbox) and drop in the JSON Schemas + prompts.
• Days 5–6: Build a 50‑row golden set (5–10 edge cases included). Save as CSV.
• Days 7–8: Add acceptance tests + eval harness. Schedule nightly runs.
• Days 9–10: Add error handling: retries/backoff + global error workflow + alerts. Mask PII in logs.
• Day 11: Soft‑launch behind human‑in‑the‑loop (100% review of outputs).
• Day 12: Drop to 20% spot‑check on low‑risk cases; route sensitive cases to human queue.
• Days 13–14: Track KPIs (parse rate, FP rate, FRT, cost/run). If pass, package price + SLOs and ship to first client.

Lisbon Test (pass/fail):

• Survives shaky Wi‑Fi (async job design + retries).
• Handles auth expiry and 429s without human rescue (re‑auth + backoff paths).
• Produces strict JSON/HTML the first time (schemas + validators).
• Escalates edge cases to a human with context attached (HITL + incident SOP).

Single job: qualify inbound interest using a 3–5 question ruleset, score fit, and either auto‑book or hand off. Channel examples: WhatsApp, email reply, web form, chat widget.

Recommended stack:

• Transport: WhatsApp Business API (Twilio/WATI) or web form/email.
• Orchestration: n8n or Make.com.
• LLM: any model with JSON Schema support or strong JSON adherence.
• Calendar/CRM: Google Calendar/Calendly + HubSpot/Pipedrive/Airtable.
• Observability: Langfuse (events + eval traces) + Slack alerts.

n8n/Make module map:

1. Trigger: Webhook (form) or WhatsApp message received.
2. Normalize: Transform to input JSON per schema.
3. LLM: Ask only the allowed questions if missing; otherwise score.
4. Decision: If confidence ≥ 0.75 and nextstep=bookcall → create calendar event + send link; else route to nurture/disqualify or human queue.
5. Log: Post result + payload to Langfuse; emit metrics.
6. Alerts: Confidence < 0.5 or PII/sensitive → Slack channel + assign owner.

Input JSON Schema (v1):
```json
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "https://statelessfounder.com/schemas/lead-qual-input.v1.json",
"type": "object",
"required": ["name", "contact", "source", "answers", "icp"],
"properties": {
"name": {"type": "string", "minLength": 1},
"contact": {"type": "object", "required": ["channel"],
"properties": {
"channel": {"enum": ["email", "whatsapp", "sms", "chat", "webform"]},
"email": {"type": "string", "format": "email"},
"phone": {"type": "string"},
"handle": {"type": "string"}
}
},
"source": {"type": "string"},
"answers": {"type": "object", "additionalProperties": {"type": ["string", "number", "boolean"]}},
"icp": {"type": "object", "description": "Ideal customer profile summary used for scoring.",
"properties": {
"industry": {"type": "string"},
"companysizemin": {"type": "integer"},
"companysizemax": {"type": "integer"},
"geo": {"type": "string"}
},
"required": ["industry"]
}
},
"additionalProperties": false
}
```

Output JSON Schema (v1):
```json
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "https://statelessfounder.com/schemas/lead-qual-output.v1.json",
"type": "object",
"required": ["score", "reasons", "next_step", "confidence"],
"properties": {
"status": {"enum": ["ok", "needmoreinfo"]},
"missing": {"type": "array", "items": {"type": "string"}},
"score": {"type": "integer", "minimum": 0, "maximum": 100},
"reasons": {"type": "array", "items": {"type": "string"}, "minItems": 1},
"nextstep": {"enum": ["bookcall", "nurture", "disqualify"]},
"confidence": {"type": "number", "minimum": 0, "maximum": 1},
"booking_link": {"type": "string", "format": "uri"}
},
"additionalProperties": false
}
```

Prompt scaffold (system):
```
You are a lead qualification assistant. Return STRICT JSON matching lead-qual-output.v1.json.
Use only data in lead-qual-input.v1.json. If required answers are missing, set status:"needmoreinfo" and list missing.
Do not invent data. If nextstep is "bookcall", confidence must be ≥ 0.75.
```

Acceptance tests (ship with these):

• Parse rate: 100 valid JSON outputs over 100 test runs on golden set.
• Business rule: Auto‑fail if nextstep=bookcall AND confidence < 0.75.
• Safety: Redact email/phone in logs; no PII in Slack messages.
• QA: Randomly sample 10% of “book_call” for human review daily.

Escalation matrix:

• Confidence < 0.5 → Human review queue.
• “High deal value” keywords or competitor mentions → Tag owner and escalate immediately.
• Three consecutive failures in 10 minutes (same node) → open incident, notify on‑call, pause auto‑booking.

Core KPIs:

• Contact→Qualified% (7‑day rolling), Auto‑book rate, False‑positive rate (human reversals), First‑response time (FRT).

Single job: pull weekly KPIs from trusted sources, validate math, summarize to Slack and an HTML email. No dashboards to click, no hallucinated numbers.

Recommended stack:

• Data: Databox MCP (to GA4/Ads/CRM) or native connectors.
• Orchestration: n8n (CRON Monday 09:00 local) or Make scheduler.
• LLM: model capable of strict JSON + HTML generation.
• Output: Slack channel + HTML email to client list.

n8n/Make module map:

1. Trigger: Schedule.
2. Fetch: GA4/Ads/CRM metrics via Databox MCP.
3. Validate: Check totals = sum(children); percentages in [0,100]. If missing data, set value="unknown".
4. LLM: Produce JSON with {slack[], html}.
5. HTML validator: fail fast on malformed tags/styles; if fail → human QA + retry.
6. Send: Slack + email. Log artifacts + costs to Langfuse.

Output JSON Schema (v1):
```json
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "https://statelessfounder.com/schemas/weekly-report-output.v1.json",
"type": "object",
"required": ["slack", "html"],
"properties": {
"slack": {
"type": "array",
"items": {"type": "string", "maxLength": 140},
"minItems": 3, "maxItems": 12
},
"html": {"type": "string", "minLength": 200}
},
"additionalProperties": false
}
```

Prompt scaffold (system):
```
Return STRICT JSON per weekly-report-output.v1.json.

• Slack: ≤12 bullet lines, each < 140 chars.
• HTML: table-based layout, inline CSS only, no external assets. If any metric source is missing, write "unknown" and explain briefly at the bottom.

Validate numeric sums; if inconsistency, emit a "consistency_note" in an HTML footer.
```

Acceptance tests:

• HTML passes validator; no unclosed tags.
• Numeric checks recompute within ±0.1% tolerance.
• If any dataset is missing, HTML includes an “unknown” callout and no fabricated values.
• Slack array contains 3–12 lines; each < 140 characters.

Core KPIs:

• On‑time delivery (% Mondays by 09:15), Parse/validation pass rate, Client replies within 24h, Support tickets opened from report (should be low).

Single job: classify incoming messages, set priority, and route to the right owner/folder. Drafting replies is optional and should require human send.

Recommended stack:

• Inbox: Gmail/Outlook or Helpdesk (Help Scout/Zendesk).
• Orchestration: n8n/Make with per‑module error handling.
• LLM: classification‑focused with deterministic labels.
• Storage: Airtable/DB table for message log + outcomes.

Allowed labels: {support, sales, billing, spam, personal}
Sensitive intents: {legal, refund, escalation}

Input JSON (expect):
```json
{
"message_id": "str",
"thread_id": "str",
"subject": "str",
"from": "email",
"body": "str",
"attachments": [{"filename": "str", "mime": "str"}],
"received_at": "ISO-8601"
}
```

Output JSON Schema (v1):
```json
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "https://statelessfounder.com/schemas/inbox-triage-output.v1.json",
"type": "object",
"required": ["label", "priority", "reason", "sensitive"],
"properties": {
"label": {"enum": ["support", "sales", "billing", "spam", "personal"]},
"priority": {"type": "integer", "minimum": 1, "maximum": 5},
"reason": {"type": "string"},
"sensitive": {"type": "boolean"},
"route_to": {"type": "string", "description": "Mailbox/folder/owner ID"}
},
"additionalProperties": false
}
```

Prompt scaffold (system):
```
Classify the message into one allowed label. Return STRICT JSON per inbox-triage-output.v1.json.
Never write to customers. If refund/legal/escalation is detected, set sensitive=true and raise priority (1=highest).
```

Acceptance tests:

• Labels limited to the allowed set; any other value auto‑fails.
• PII masking in logs (email/phone regex) before storage/alerts.
• Sensitive=true always routes to human queue; never send auto‑replies.
• Duplicates: same thread_id within 10 minutes → dedupe and skip.

Core KPIs:

• First response time (FRT), Deflection rate (auto‑resolved without human), Sensitive‑case catch rate, Manual review hit rate.

Own your reliability. The eval harness keeps you from shipping regressions and gives you numbers for clients.

Golden set (per agent):

• Size: 50 rows minimum (10 edge cases: missing fields, OOO auto‑replies, malformed HTML, rate‑limit simulation, duplicate thread, high‑value exceptions).
• Format: CSV with columns: inputjson, expectedoutput_json, notes, tags.

Sample CSV (3 rows shown):
```csv
inputjson,expectedoutput_json,notes,tags
"{\"name\":\"Ada\",\"contact\":{\"channel\":\"webform\",\"email\":\"ada@example.com\"},\"source\":\"LP\",\"answers\":{\"budget\":\"2k\"},\"icp\":{\"industry\":\"coaching\"}}","{\"status\":\"needmoreinfo\",\"missing\":[\"timeline\"],\"score\":0,\"reasons\":[\"Missing timeline\"],\"next_step\":\"nurture\",\"confidence\":0.4}","Lead-qual missing timeline","edge,lead-qual"
"{\"slack\":[],\"html\":\"<html><body><table><tr><td>T</td></tr></table></body></html>\"}","SCHEMA_FAIL","Reporting with empty Slack block","edge,reporting"
"{\"messageid\":\"1\",\"threadid\":\"A\",\"subject\":\"Refund\",\"from\":\"c@example.com\",\"body\":\"I want a refund\"}","{\"label\":\"billing\",\"priority\":1,\"reason\":\"Refund keyword\",\"sensitive\":true,\"routeto\":\"billingqueue\"}","Inbox refund path","edge,inbox"
```

Metrics to track:

• Schema parse rate = validoutputs / totalruns
• Business rule pass rate = passesbusinessrules / total_runs
• False‑positive rate (agent booked call but human reversed) = fp / totalbookcalls
• Deflection rate (triage only) = autoresolved / totaltickets
• Manual review hit rate = senttohuman / total_runs
• Latency p95 and cost/run (sum of LLM + platform costs)

Harness setup (Langfuse/OpenAI‑evals friendly):

• Store golden rows as dataset. Nightly job runs agent on dataset; compare outputs; upload traces + costs.
• Gate deploys: if parse rate or business rule pass rate drops by >2% vs last green, auto‑block and alert.
• Weekly: sample 5 outputs per agent for human rubric scoring (1–5) and trend.

Command sketch (pseudocode):
```
make eval AGENT=lead-qual DATA=golden/leadqual.csv OUT=reports/leadqual_$(date).json
```

Governance defaults:

• Keep 90 days of eval artifacts.
• Tag runs by model+prompt version.
• Treat model changes like code changes (review + rollback path).

Most failures aren’t “AI problems.” They’re ops: timeouts, 429s, auth expiry, brittle HTML. Put guardrails in writing.

Global error workflow (copy‑ready SOP):

1. Detection: Route all n8n/Make errors to a global Error Workflow. Capture: workflowid, node, message, executionurl, retryOf.
2. Classify:

• HTTP 5xx/timeout/429 → Retry path.
• 4xx auth (401/403) → Re‑auth path.
• Schema/validation fail → Human QA path.

1. Retry: 3 attempts with exponential backoff (1m, 5m, 15m). Jitter recommended.
2. Escalation: Still failing after retries or confidence < threshold → open ticket, notify on‑call (Slack/PagerDuty). Attach execution_url and a redacted payload snippet.
3. Post‑incident: Create a postmortem card. Add a regression case to the golden set. Tag the run with a new incident ID.

Operational safeguards:

• PII: Mask emails/phones in logs and alerts.
• Rate limits: Centralize backoff settings; never parallelize beyond provider quotas.
• Auth: Monitor token age; auto‑rotate credentials with least privilege; alert before expiry.
• Zapier caveat: Keep trigger work under 30s; push long calls to async callbacks/queues.

Expected outcomes:

• Incidents resolved within 30 minutes during business hours.
• <1 false‑positive booking per 100 auto‑booked calls.
• Weekly reporting failure rate <2% (auto‑retry recovers most).

Track the numbers that pay the bills and price with a margin, not vibes.

KPI sheet (minimum set):

• Parse rate (schema): validjson / totalruns
• Business rule pass rate: passesrules / totalruns
• False‑positive rate (lead‑qual): fp / totalbookcalls
• FRT (inbox/lead‑qual): p50 and p95 in minutes
• On‑time delivery (reporting): deliveriesbefore0915 / totalweeks
• Cost/run: llmcost + platformcost + infra_cost

Pricing calculator (logic):
Inputs:

• setup_hours
• hourly_rate
• platform_monthly (n8n/Make/Zapier tier)
• avgrunsper_month
• llmcostper_run (tokens × price/1k + vector/db/query costs)
• maintenancehoursper_month
• targetgrossmargin (e.g., 0.6)

Formulas:

• setupfee = setuphours × hourly_rate + one‑time tooling (if any)
• monthlydirectcost = platformmonthly + (avgrunspermonth × llmcostperrun) + (maintenancehourspermonth × hourly_rate)
• recommendedmonthlyprice = ceil(monthlydirectcost / (1 - targetgrossmargin))
• anchor_ranges (market‑checked): entry $99–$299/mo per unit; mid €500–€3,000/mo; high‑touch $2,500–$5,000+/mo. Use your cost model to pick a lane.

Packaging notes:

• Productize by agent: one setup + one monthly per agent. Offer a bundle discount for 2–3 agents.
• SLOs in proposal: parse rate ≥ 98%, FP rate ≤ 1%, reporting on Mondays by 09:15, incident response < 2 hours during support window.
• Change management: 90‑day review to tune thresholds and cut human review from 20% → 10% once stable.

Run this before you turn on fully unattended mode.

Go‑live checklist:

1. One job per agent; JSON Schemas validated and versioned.
2. Golden set (≥50) loaded; nightly eval passing; deploy gate active.
3. Retries/backoff implemented globally; rate‑limit quotas set.
4. Auth rotation tested; failure alerts include re‑auth instructions.
5. PII masking verified end‑to‑end (logs, alerts, data store).
6. Human‑in‑the‑loop paths wired; sensitive cases always human‑owned.
7. HTML validation (reporting) green for two consecutive Mondays.
8. KPIs visible to client (shared dashboard or weekly footer block).
9. Incident SOP rehearsed once with a forced failure.
10. Lisbon Test: pull laptop power and Wi‑Fi for 10 minutes mid‑run — does it recover without you?

Minimal SLO card to paste in your client doc:

• Lead‑qual: ≥98% parse rate; FP ≤1%; median FRT ≤5 min.
• Reporting: On‑time Mondays by 09:15; validation pass ≥98%.
• Inbox: Sensitive‑case catch rate ≥95%; deflection rate target agreed per client.

What to iterate next:

• Reduce manual review from 20% → 10% once KPIs hold for 4 weeks.
• Add canary checks (1 synthetic message/report per day) to detect silent failures.
• Graduate autonomy only after three green eval cycles post‑change (model or prompt).