n8n workflows for an online store, published as importable templates. Each one solves a problem a small business actually has: customer emails, shop chat, social comments, late parcels, stock-outs, a cluttered task board.
Each one is derived from a workflow that runs unattended on a real store, generalized so you can adapt it: prompts and business rules are templates to fill in. The design choices in each README come from what broke along the way.
| Workflow | What it does | Main services |
|---|---|---|
| ai-email-support-agent | Reads the support inbox every 3 minutes and drafts answers grounded in your knowledge base and order data (auto-send for safe cases is off by default) | Gmail, Shopify, Postgres + pgvector, any OpenAI-compatible model |
| ecommerce-chat-agent | Website chat agent that answers product questions and looks up an order only when the customer gives both its number and its email | Shopify, Postgres, OpenAI-compatible model |
| social-comments-agent | Answers questions under Facebook and Instagram posts and ads, hides only abuse, spam, scams, personal data and the authors you block yourself, and sends criticism and safety concerns to a human instead of hiding them | Meta Graph API, Telegram, OpenAI-compatible model |
| shipping-exception-alerts | Catches delivery problems from carrier webhooks before the customer writes in | 17TRACK, Shopify |
| low-stock-alerts | Warns before a product runs out, with thresholds computed from real sales | Shopify, Telegram |
| notion-nightly-archive | Moves finished tasks out of a Notion TODO database every night | Notion, Telegram |
- In n8n, Workflows → Import from file, pick the
workflow*.jsonof the folder. - Create the credentials listed in that folder's README (every credential in the
file is set to
REPLACE_ME). - Replace each
YOUR_…placeholder. The README lists all of them. - Run it by hand first. The email agent ships in draft mode; the chat and comment agents answer live, so test them on a staging site or a private Page before you point them at customers.
- Deterministic first, AI second. Sorting, deduplication and the decision to send are rules you can read. The model drafts; it does not decide alone.
- Ground every answer. Agents look up the knowledge base and the order before they write, and say so when they do not know.
- A human stays on what engages the brand. Doubtful cases, criticism and anything about safety become drafts or alerts for a person, never a silent hide.
- Say it is an AI. Messages written for customers carry a short, configurable AI disclosure, on by default.
- Webhooks that trigger an action check who is calling. They verify a signature or a header token before doing anything. The public chat widget cannot keep a secret: it filters the origin and the message length, and must sit behind a proxy that limits the request rate.
- Collect less, keep it shorter. Models only see what they need, and memory tables have a retention period you set.
- Leave a trail. Every run is logged, and questions the agent could not answer land in a review queue, so a human can grow the knowledge base.
I build and run automations and AI agents like these. LinkedIn · Malt
MIT licensed.