What Is n8n, and Should a Small Business Use It?

n8n is a workflow automation tool, like Zapier or Make, that connects your apps and runs multi-step jobs, including AI steps. n8n Cloud starts at about $23-$27 a month, or you can run the free Community Edition on your own server. It suits a small business with someone technical to hand; without that, Zapier or Make is usually simpler.

What sets n8n apart is how it charges: per workflow run, not per step. A 12-step workflow that runs 1,500 times a month counts as 1,500 executions on n8n but 16,500 tasks on Zapier. The trade-off is that n8n expects more of you: reading error messages, reshaping data between steps and, if you self-host, keeping a server secure and up to date.

Follow me on Instagram@sagnikteaches

n8n in plain English

Picture a canvas where each box is one action: "when an email arrives", "pull out the order lines", "look up the customer", "create the sales order", "message the team". You join the boxes with lines, and n8n runs the chain every time the first box fires. That chain is a workflow, and the boxes are called nodes.

Connect on LinkedInSagnik Bhattacharya

n8n calls itself "fair-code": the source code is public and you can run it yourself for free, but the licence limits reselling it as a product. It has 400-plus built-in integrations by its own count, a Code node for JavaScript or Python when nothing built-in fits, and an HTTP Request node that can talk to almost any system with an API. That flexibility is why developers like it, and why it can feel less guided than Zapier to someone who has never seen a JSON file.

Subscribe on YouTube@codingliquids

The four building blocks: triggers, nodes, credentials, executions

• Trigger: what starts the workflow. A new email, a form submission, a row added to a sheet, a schedule ("every weekday at 7am") or a webhook, which is a web address another app calls the instant something happens. Our explainer on what a webhook is covers why some automations run instantly and others wait.
• Nodes: each action or decision. Some connect to apps, some transform data (split, merge, filter, rename fields), some make decisions (IF, Switch), and some call AI models.
• Credentials: the stored logins and API keys that let nodes act on your accounts. n8n encrypts them; whoever can edit workflows can use them, so access matters.
• Executions: each complete run of a workflow. n8n's pricing is built on these, and it counts one execution however many nodes the run passes through.

Between those blocks sits the part that makes n8n feel technical: data passes from node to node as fields, and fields rarely arrive in the shape the next app wants. A customer's email says the delivery date is "6/10/26"; the order system's API wants "2026-10-06". A spreadsheet stores quantities as text, so "200 " with a trailing space fails a number check. In Zapier, a Formatter step handles many of these conversions with menus. In n8n you often write a short expression in the field instead, or add a small node to tidy the data. Neither is hard once you've done a few, but it's the skill that separates a workflow that runs from one that runs every time.

Why per-execution pricing changes the sums

Zapier bills per task, meaning each successful action step (triggers and filters don't count). Make bills per credit, and each module action generally uses one. n8n bills per execution. For short workflows the difference is small; for long ones it's large. Take one workflow with a trigger and 11 action steps running 1,500 times a month:

Flip the shape and the answer flips. A two-step workflow (trigger plus one action) running 1,500 times is 1,500 Zapier tasks, and Zapier's 2,000-task Professional tier at $49 a month billed annually is competitive once you count the time you'd spend learning n8n. The rule: n8n's pricing wins for long workflows with many steps; for short ones, choose on ease of use.

n8n Cloud prices are set in euros, so the dollar cost moves with exchange rates. The plans, per n8n's pricing page: Starter with 2,500 executions a month; Pro at 50 euros a month billed annually (60 monthly) with 10,000; a self-hosted Business plan with 40,000 at a much higher price; and Enterprise on quote. Every plan includes unlimited users, workflows and steps.

The AI side: agent nodes, models and whose bill it is

n8n has two ways to use AI. The simple one is a node that sends text to a model with an instruction ("extract the order lines as JSON") and passes the answer on. The more ambitious one is the AI Agent node, which connects a chat model to "tools" (other nodes it can call, such as a stock lookup) and memory, and lets the model decide which tool to use for each request.

For most small-business jobs the simple node is enough, and more predictable. Either way you connect your own account with a model provider such as OpenAI or Anthropic, and the token costs are billed by them, not by n8n. The costs are usually small. Extracting order details from 1,500 emails a month, at roughly 2,000 tokens in and 400 out each, on OpenAI's gpt-5.6-luna ($0.20 per million input tokens, $1.20 per million output) comes to about $0.60 plus $0.72, so under $1.50 a month. On Claude Haiku 4.5 ($1 / $5) it's about $6.

The detail that makes AI nodes reliable is asking for structured output: a fixed set of fields the next node can read. An illustrative result from an order email:

{
  "customer_account": "TR-0412",
  "po_number": "PO-88213",
  "lines": [
    {"supplier_code": "HX-40-SS", "description": "hose clamp 40mm stainless", "qty": 200},
    {"supplier_code": null, "description": "the blue caps like last time", "qty": 50}
  ],
  "delivery_date_requested": "2026-10-06",
  "needs_human": true,
  "reason": "line 2 has no product code"
}

The needs_human flag is the part to insist on. Without it, the model would guess which "blue caps" the customer meant, and a confident guess is worse than a question.

The agent node earns its place when a request could need different lookups each time. An illustrative testing laboratory gave an agent two tools, "look up sample status by client reference" and "look up the turnaround promised on the quote", and let staff ask it questions in a chat window: "Has the batch from the brewery client finished microbiology, and are we inside the promised time?" The agent picked the right lookups and answered in a sentence. The same lab kept the agent read-only: it can look things up but has no tool that changes records or emails clients, because an agent that can act can also act wrongly. Start agents with read-only tools, log every tool call, and add actions only after a month of reviewing those logs.

Cloud, self-hosted, and what the licence allows

n8n's own documentation is plain that self-hosting needs technical expertise to install and configure, and that maintenance is your responsibility. The Community Edition's Sustainable Use License lets you run n8n for your own business's internal workflows; what it restricts is selling a product whose value comes mainly from n8n itself. A wholesaler automating its own orders is fine. For the full cost comparison, including servers and maintenance time, see n8n self-hosted versus cloud costs.

A wholesaler's first workflow, node by node

An illustrative wholesaler of catering disposables takes about 1,100 emailed orders a month from trade customers, in every format: typed lists, forwarded spreadsheets, photos of handwritten notes. Sales admin spends about six minutes keying each one. The first n8n workflow looks like this:

1. Email trigger on the orders inbox.
2. AI node extracts customer, PO number and lines as structured JSON, with a needs_human flag.
3. Google Sheets lookup maps each customer's own product descriptions to the wholesaler's SKUs, from a table sales admin maintains.
4. IF node: every line matched, customer known, not flagged? If yes, carry on; if no, send to step 7.
5. HTTP Request node creates a draft sales order in the order system through its API.
6. Message to the sales channel: "Draft order 55120 for TR-0412, 6 lines, check and confirm."
7. Exceptions path: message with the original email attached and the reason, for a person to key.

That's 1,100 executions a month, comfortably inside Starter's 2,500. In an illustrative first month, 71% of orders went down the automatic path; checking a draft took about 90 seconds against six minutes to key it. Keying time for the automatic orders fell from roughly 78 hours to 20 hours of checking, and the exceptions stayed as before. Setup took a technically confident manager about 16 hours spread over three weeks, most of it on the SKU mapping table and the order system's API. The same job is described from the wholesaler's side, without n8n specifics, in reading emailed orders into a wholesaler's system with AI.

Who n8n suits, and who should pick something simpler

Two quick illustrations. A courier firm sending a tracking email for each of 4,000 parcels a month, in a workflow with a trigger and five actions, would use 20,000 Zapier tasks against 4,000 n8n executions: n8n is worth the learning curve there. A laboratory that wants one weekly summary of test turnaround times has maybe four runs a month, and should use whatever its team already knows. If you're unsure whether to learn a tool like this at all, hiring an automation consultant versus building it yourself weighs the options.

What goes wrong in the first three months

• A schedule that eats executions. "Check the sheet every minute" fires 60 × 24 × 30 = 43,200 times a month, more than 17 times Starter's allowance, even if there's rarely anything to do. Use a webhook or an email trigger, or check every 15 minutes during working hours.
• No error workflow. When a node fails, the execution stops and, unless you've set up an error workflow that messages someone, nobody knows. Orders silently go missing. Stopping automations breaking silently covers alerting in general.
• Short execution history. Starter keeps execution data for 7 days, Pro for 30. A customer complaining about an order from three weeks ago may be undebuggable on Starter. Log key results to a sheet as you go.
• AI output drifts. Without a strict output format, a model occasionally returns "Here are the order lines:" before the JSON, and the next node breaks. Use structured output and treat any malformed answer as an exception.
• Expired connections. Some app logins expire or get revoked when someone changes a password. Build a weekly test run that exercises every credential.
• One person knows it. The manager who built it goes on holiday and a workflow fails. Keep a one-page note per workflow: what it does, where the SKU table lives, how to switch it off.

The running cost nobody quotes: someone's time

The subscription is the small part. Plan for regular upkeep per active workflow: checking the exceptions queue, updating the mapping table when customers change their wording, fixing a node when an app changes its API, and reviewing AI outputs for drift. For the wholesaler above, that came to about an hour a week once things settled, mostly on new product descriptions ("the blue caps like last time") that had to be added to the SKU table. Self-hosting adds server updates on top: n8n releases new versions frequently, and security fixes are among them, so somebody has to apply updates promptly and test that workflows still run afterwards. If that time comes from an owner who is already stretched, include it in the comparison with Zapier, because an easier tool that costs more a month can still be the cheaper choice once the hours are counted.

How to trial n8n without betting the business on it

1. Choose one workflow you've already mapped, with a known volume and a clear exceptions path.
2. Build it on n8n Cloud's free trial rather than a server, so you're testing n8n, not your hosting skills.
3. Run it alongside the manual process for two weeks: people keep working as normal while you compare the drafts.
4. Count three things: how many runs completed correctly, how many went to the exceptions path, and how long the build and fixes took.
5. Decide on those numbers. If the build took three times longer than expected and one person barely understands it, that's a real cost, even if the subscription is cheap.

n8n is a capable tool whose pricing rewards ambitious workflows. For a small business the deciding question usually isn't the price; it's who will look after the workflows next year. If you have that person, or can hire one, n8n's per-execution model and flexibility are hard to beat. If you don't, a simpler platform with a higher bill will cost less overall. For a side-by-side view of all three platforms on AI work, read Zapier vs Make vs n8n for AI automation.

n8n questions owners ask before trying it

Do I need to know how to code to use n8n?

Not for simple workflows: you drag nodes onto a canvas and fill in settings, much like Zapier or Make. But n8n shows you raw data more often, and real workflows usually need small expressions to reformat dates, join fields or pick an item from a list. If reading a line like {{ $json.order_number }} makes you uneasy, budget for a contractor to build the first workflows and show you how to maintain them.

Can n8n connect to my accounting or stock software?

Check n8n's integrations list for a built-in node first. If there isn't one, the HTTP Request node can talk to any system with an API, using the vendor's API documentation and an API key. That route works well but takes technical effort. If your system has no API at all, n8n can still read and write files or emails, which is often enough for a CSV import.

Is n8n safe to use with customer data?

It can be, but safety depends on how you run it. On n8n Cloud, the vendor runs the servers and you manage user access and credentials. Self-hosted, security is your job: updates, backups, firewall settings and protecting the encryption key for stored credentials. Either way, check the data-processing terms, limit who can see execution data, and send only the fields each step needs to any AI model.

Further reads

• When Does Zapier Get Too Expensive? Finding the Tipping Point — Find the point where per-task pricing makes n8n worth the effort.
• How to Build Your First AI Automation in Make, Step by Step — If n8n feels too technical, build the same workflow in Make.
• How to Add Human Approval Steps to AI Automations — Add a person-approves step before an AI output goes anywhere.
• AI Agent vs Chatbot vs Automation: Which Does Your Business Need? — Decide whether you need an agent node at all.
• How to Hire Your First AI Automation Specialist — What to look for in someone who can run n8n for you.
• How to Map a Business Process Before You Automate It — Map the process first so the workflow handles its exceptions.
• iPaaS, Middleware or Custom Code: How Should Your Apps Connect? — iPaaS, middleware and custom code explained in plain English, with one real-shaped integration costed three ways over three years.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: n8n pricing page (plans, executions, retention, concurrency), n8n documentation (choosing a hosting option, Sustainable Use License, AI Agent node), n8n GitHub repository; Zapier pricing page (task tiers); facts sheet for API token prices.