AI Lead Inbox → CRM Autopilot

A lead submitted on a public form is scored, enriched, written to a CRM-shaped store, given a draft reply, and announced with a Slack-compatible alert. The same rules run in this Next.js app and inside an n8n workflow you can import.

Northline, the studio named in the draft emails, is fictional. The pipeline is not. Mock mode needs no API key.

Architecture

flowchart TB
  form[Public lead form]
  form --> app[TypeScript pipeline]
  form -. optional relay .-> hook[n8n webhook /webhook/lead]
  subgraph rules [Shared lead logic]
    score[Normalize, score, route]
  end
  app --> score
  hook --> score
  score --> mode{LLM_MODE}
  mode -->|mock default| mock[Deterministic enrichment and draft]
  mode -->|openai| live["POST /v1/chat/completions"]
  mock --> crm[(SQLite leads)]
  live --> crm
  crm --- csv[leads.csv]
  crm --> alert{SLACK_WEBHOOK_URL}
  alert -->|set| slack[Slack-compatible webhook]
  alert -->|empty| inbox[Alert stored in the inbox UI]

Loading

The n8n graph scores in a Code node, then calls this app for the AI step, the CRM write, and the alert. Point the AI node at OpenAI directly when you want n8n to hold the model credential. See n8n.

What a run produces

Lead	Score	Tier	Owner	Channel	SLA
Priya Shah, Helio Clinic	100	hot	Avery Chen	#hot-leads	15 min
Marcus Adeyemi, Bright Parcel	64	warm	Jordan Hale	#inbound	4 hours
Sam Rivera, Rivera Studio	22	cold	Nurture sequence	#nurture-digest	1 day
Alex Guest, Rank Farm	0	review	Review queue	#review-queue	1 day

Captured with npm run autopilot -- --write-sample in mock mode. Full records are in samples/e2e-run.json, the sheet is samples/leads.csv, and the hot alert is samples/slack-alert.json.

The same path in the browser, still in mock mode:

The form also stacks on a phone-width viewport: docs/screenshots/intake-mobile.png.

Hot alert text:

Hot lead: Priya Shah at Helio Clinic — score 100 — Avery Chen — 15m SLA

Draft subject: Helio Clinic's lead inbox, scored before anyone is pinged.

The Rank Farm row is a guest-post pitch from a disposable inbox. Sales is not paged. The draft is an internal note that starts with "Do not send".

Scoring and routing

Points, capped at 100:

Signal	Points
Budget $15,000+ / $5,000–$14,999 / $1,500–$4,999 / under $1,500 / not specified	30 / 22 / 14 / 6 / 4
Timeline this week / this month / this quarter / this year / exploring	25 / 22 / 16 / 8 / 4
Work email / free inbox (Gmail, Outlook, iCloud, …) / disposable	15 / 4 / 0 and a review flag
Founder, CEO, owner, partner / VP, director, head, chief / manager, lead / everyone else	15 / 12 / 8 / 4
Note mentions automation (n8n, Zapier, Make, workflow)	+4
Note mentions a CRM, HubSpot, Salesforce, Airtable, or a spreadsheet	+3
Note mentions leads, inbound, or pipeline	+3, intent capped at 10
Website present	5
Disposable inbox	−40 and tier disqualified
Guest-post, backlink, crypto-airdrop, or follower-buy language	−50 and tier disqualified

Routing uses the score after penalties:

Score	Tier	Owner	Slack channel	SLA
70–100	hot	Avery Chen, Senior AE	#hot-leads	15 min
40–69	warm	Jordan Hale, Account executive	#inbound	4 hours
0–39	cold	Nurture sequence	#nurture-digest	1 day
spam or disposable flag	disqualified	Review queue	#review-queue	1 day, no sales email

Boundaries are tested: 70 is hot, 69 is warm, 40 is warm, 39 is cold. A founder with a huge budget still goes to review if the message is a backlink pitch.

Jev decision engine (optional)

When TYPESAFE_API_KEY is set, the TypeScript pipeline asks Jev (jev-1.13.0) to score and route the lead before the local rubric is used. One POST https://api.typesafe.ai/v1/systemone per lead sends the form as state and these questions together:

Question	Type	What the app does with the answer
score	score, 10 ordered levels	Maps the 0–9 position onto the 0–100 score stored in the CRM
route	choice: avery, jordan, nurture, review	Picks the owner, channel, and SLA
intent	choice: automation, crm, inbound, partnership, other	Shown on the receipt and the lead row
spam	noul	At 0.50 or higher, hold for human review
urgent	noul	Shown on the receipt. At 0.50 or higher, the row is flagged urgent
needs_human	noul	At 0.50 or higher, hold for human review

Choice and score answers also return confidence. If score, route, or intent confidence is below JEV_CONFIDENCE_THRESHOLD (default 0.6), the lead goes to the review queue instead of the chosen owner. The receipt, the CRM row, and the Slack text show those confidences and the three gate probabilities. The local rubric bars stay on the lead page so you can compare them with Jev's score.

Bring your own key. Copy .env.example to .env and set TYPESAFE_API_KEY. Do not commit .env. The key is read only from that environment variable and is sent as Authorization: Bearer …. It is not written into the repo, the workflow JSON, or the CRM row.

If the variable is empty, or the call times out, returns a non-200, or comes back without the six answers, the pipeline keeps the existing rubric score and route. The run still files the lead. A failed call leaves a Jev step that says it fell back to the local rubric.

The n8n Code node still embeds the local rubric. Jev runs when the form (checkbox off), POST /api/leads, or npm run autopilot uses this app. Docker Compose forwards TYPESAFE_API_KEY into the web service when you export it in the shell.

Live check, after you put your own key in the environment:

TYPESAFE_API_KEY=your-key npm run jev:smoke

That posts the hot clinic fixture once and prints the score, route, intent, gate probabilities, and whether the confidence threshold held it. npm test never calls TypeSafe. It mocks fetch.

CRM shape

SQLite lives at $DATA_DIR/crm.sqlite (default ./data, gitignored). Every write rewrites leads.csv in the same folder, UTF-8 with a BOM, RFC 4180 quoting. Headers:

Lead ID, Created At, Full Name, Email, Company, Role, Website, Need, Monthly Budget USD, Budget Label, Timeline, Source, Score, Tier, Owner, Owner Role, Slack Channel, SLA Minutes, Route Reason, Industry, Company Size, Intent, Pain Points, Personalization Hooks, Next Action, Draft Subject, Draft Body, LLM Provider, LLM Model, Alert Status, Origin

Airtable field types for the same columns: single line text for names and ids, email for Email, url for Website, long text for Need, Intent, Pain Points, Personalization Hooks, Next Action, and Draft Body, number for Monthly Budget USD, Score, and SLA Minutes, single select for Timeline, Tier, Budget Label, and Origin, date time for Created At.

Google Sheets: File → Import → Upload leads.csv, or append the same columns with a Sheets node in n8n.

LLM providers

src/lib/llm.js picks the provider from the environment.

Variable	Default	Purpose
LLM_MODE	mock	mock or openai
LLM_BASE_URL	https://api.openai.com/v1	Any OpenAI-compatible /chat/completions base
LLM_API_KEY	empty	Required only when LLM_MODE=openai
LLM_MODEL	gpt-4o-mini	Model name sent to that API
LLM_TIMEOUT_MS	20000	Request timeout
SLACK_WEBHOOK_URL	empty	Incoming webhook. Empty keeps the alert in the app
N8N_WEBHOOK_URL	http://127.0.0.1:5678/webhook/lead	Used by the form's n8n checkbox
DATA_DIR	./data	SQLite and CSV directory
TYPESAFE_API_KEY	empty	Optional Jev key. Empty keeps the local rubric
JEV_CONFIDENCE_THRESHOLD	0.6	Score, route, or intent confidence below this goes to review
JEV_TIMEOUT_MS	20000	Jev request timeout

Mock mode is a pure function of the lead, the score, and the route. The same input always returns the same industry, pains, and draft. It does not call the network. OpenAI mode sends one JSON-object chat completion and refuses the response if the draft or the other required keys are missing. If LLM_MODE=openai and the key is empty, the run fails before a CRM row is written.

Copy .env.example to .env only when you want a live model or a real Slack webhook. .env is gitignored. This repo does not contain keys.

Run it locally

npm install
npm test
npm run dev

Open http://127.0.0.1:4317. Submit the form, or use the sample chips, then open Inbox. "File the four sample leads" runs the same fixtures through the API.

Replay the captured path without the server:

npm run autopilot -- --write-sample

npm test covers the rubric, tier boundaries, mock determinism, CSV quoting, the OpenAI-without-a-key failure, and the fact that the n8n Code nodes contain the same scoreLead / routeLead source as the app.

Docker Compose

docker compose up --build

• App: http://localhost:4317
• n8n: http://localhost:5678

On first boot the n8n container imports n8n/ai-lead-inbox.workflow.json. Create the local owner account n8n asks for, open the workflow, and publish it. Production webhook: POST http://localhost:5678/webhook/lead.

The Code node pipelineBaseUrl is http://web:4317, which is the app on the compose network. If you run n8n on the host instead, change that constant to http://127.0.0.1:4317.

The form checkbox "Send this lead through the n8n webhook" posts to the app, which relays to N8N_WEBHOOK_URL. Inside compose that URL is http://n8n:5678/webhook/lead. Leave the checkbox off to run the TypeScript pipeline with n8n stopped.

Image pin: n8nio/n8n:2.43.3. The app image is node:22-bookworm-slim so node:sqlite is available.

n8n

Import path: n8n menu → Import from file → n8n/ai-lead-inbox.workflow.json.

Nodes, in order:

1. Lead webhook — POST /lead, response comes from the last node.
2. Score and route — the shared JavaScript. Edit CONFIG.pipelineBaseUrl here.
3. Valid lead? — invalid payloads get a 400 from Reject invalid lead.
4. AI enrich — POST {pipelineBaseUrl}/api/llm/enrich. Honors the app's LLM_MODE.
5. Assemble record — merges enrichment into the CRM record and the Slack payload.
6. Write CRM row — POST /api/crm/commit.
7. Post Slack-style alert — POST /api/alerts. The app stores it and forwards to SLACK_WEBHOOK_URL when that variable is set.
8. Reply to caller — returns the stored lead and alert.

OpenAI direct is on the canvas and disabled. It is an HTTP Request to https://api.openai.com/v1/chat/completions with a Header Auth credential id of PLACEHOLDER_OPENAI_HEADER_AUTH. To use it: create a Header Auth credential whose name is Authorization and whose value is Bearer <your key>, attach it, disable AI enrich, and connect OpenAI direct between Valid lead? and Assemble record. For a non-OpenAI base URL, change the node URL and keep /chat/completions. Do not paste the key into the workflow JSON.

To swap the CRM write for Google Sheets or Airtable, replace Write CRM row with that node and map the fields listed above. To post from the native Slack node, replace Post Slack-style alert and map alert.text plus the blocks. Sticky notes on the canvas repeat this.

Regenerate the export after editing src/lib/lead-logic.js:

npm run workflow:build
npm run workflow:check

Make and Zapier

The scenario is the same seven steps. You do not need this repo running if you paste the scoring functions into a code module and call your own model and CRM.

Step	Make	Zapier
Trigger	Custom webhook, or a form's webhook	Webhooks by Zapier → Catch Hook, or the form's native trigger
Score and route	Tools → Run JavaScript. Paste scoreLead and routeLead from src/lib/lead-logic.js	Code by Zapier, JavaScript, same functions
Branch	Router on tier	Paths on tier
Enrich	HTTP POST to this app's /api/llm/enrich, or the OpenAI module with a JSON response and the prompt from buildEnrichmentPrompt	Webhooks POST to /api/llm/enrich, or the OpenAI action
CRM	Google Sheets → Add a row, or Airtable → Create a record	Google Sheets or Airtable, same columns
Draft	Already inside the enrichment JSON (draftSubject, draftBody). Optional: Gmail → Create draft	Same. Optional: Gmail Create Draft
Alert	HTTP POST to the Slack incoming webhook, body { "text", "blocks" }	Slack → Send Channel Message, or Webhooks POST

Hot is score ≥ 70, warm is ≥ 40, cold is the rest, and any of disposable_email, spam_pattern, or invalid_email is the review path regardless of score.

HTTP API

Method	Path	Role
POST /api/leads	Run the TypeScript pipeline	Form default
GET /api/leads	List CRM rows
GET /api/leads/:id	One row plus its alert
POST /api/llm/enrich	{ lead, scored, routing } → enrichment	n8n AI step
POST /api/crm/commit	Store a record the workflow already scored	n8n CRM step
POST /api/alerts	Store a Slack payload and forward if configured	n8n alert step
POST /api/samples	File the four fixtures	Inbox button
POST /api/relay/n8n	Forward a lead to the n8n webhook	Form checkbox
GET /api/health	Mode, model name, whether Slack is mocked	No secrets

Limitations

• Mock enrichment is a deterministic rubric, not a language model. It is there so the demo runs with no key. Turn on LLM_MODE=openai when you want a real completion.
• Jev is optional and only scores and routes. It does not write the draft. With TYPESAFE_API_KEY empty, no lead is sent to TypeSafe.
• The CSV and the column list are compatible with Google Sheets and Airtable. This demo does not call the Sheets or Airtable APIs. Swap the n8n HTTP node when you want a live base.
• Slack delivery is an incoming-webhook POST of Block Kit JSON. There is no bot token in the repo. With SLACK_WEBHOOK_URL empty, the alert is stored and shown in the inbox.
• The form and the API have no authentication. Do not expose the compose stack on the public internet as-is.
• node:sqlite is still marked experimental in Node 22. The CSV is the portable copy of the same rows.
• n8n must be published before its production webhook answers. The in-app pipeline does not need n8n.
• Docker was not executed in the environment that built this repo (no Docker daemon). The compose file, Dockerfiles, and the workflow JSON are included; import behavior follows n8n 2.43.3.
• Drafts are stored. Nothing is emailed.