
Six of our n8n builds in production cover four departments. Marketing runs two content pipelines that research, write and check articles. Customer support runs an after-hours chat agent that answers in the chat and files a ticket when a person must follow up. Finance runs invoice intake that registers supplier invoices, and a daily alert that reads the management information (MI) pack. Operations runs appointment scheduling on top of an ERP. Five of the six builds have a case study on our site and the sixth is written up on our AI proof of concept page, and every figure about a build below comes from those pages.
In n8n, each step of a workflow is a node. A trigger starts a run, and the nodes after it call APIs, move records between systems, call a language model or run code. The Code node runs JavaScript or Python where no built-in node fits, as n8n's documentation on the Code node describes. Sales and HR have no section here, because none of our published builds sits in either.
A product customisation platform for e-commerce stores paid about $700 an article and published four or five articles a month. We built a pipeline of AI agents on n8n that starts from one target keyword. A research agent clusters keywords and analyses the search results and competing pages, a brief agent sets the heading structure and word count, and content agents write, fact-check and polish the draft. No step waits for a person until the draft is ready to review. An article now costs $12.11 and takes 29 minutes, and capacity has gone from four or five articles a month to 43, as our product customisation case study reports.
A travel insurance group sells cover through four brands in the UK, Ireland and Australia, where one wrong word or one wrong statistic in an article is a legal problem. Its content pipeline runs 11 stages in n8n from a single trigger. It researches, plans topics per brand and region, builds each brief from the brand's context in Notion, drafts in the brand's own voice, and checks every draft against the brand's compliance rules before a person reads it. The group phased out external agencies and freelancers in 2026, in its own words in our compliance-first content case study.
A Dutch and Belgian supplier of branded merchandise staffed its chat desk for 45 hours of a 168-hour week, so evening and weekend questions waited for 08:30 on the next business day. On n8n, one agent answers in Dutch or English from the supplier's own resolved support conversations and checks live order status in HubSpot and AfterShip. A second agent, the Gatekeeper, checks every draft before the customer sees a word, and a triage agent reads the finished chat and files a sorted HubSpot ticket when a person must follow up. Chat now runs all 168 hours with a median first response of 7.7 seconds, and the agent went live seven weeks after kickoff, as our after-hours chat agent case study reports.
The chat agent has no permission to create tickets. Where it promised a follow-up, the triage agent creates the ticket the chat agent only claimed to make, so a promise made in the chat always reaches a person. The case study counts 704 conversations between 29 December 2025 and 26 August 2026, and 196 of them closed in the chat with no ticket.
An AI company in Pisa keyed in more than 180 supplier invoices a month by hand, and Italian rules require each invoice from outside the EU to be rebuilt as an autofattura, a self-invoice the buyer issues. Its invoice intake runs on the company's own n8n instance. The pipeline reads the purchasing inbox every 5 minutes, a browser agent collects invoices from 3 vendor portals each month, and an OpenAI step reads each invoice. n8n then sends it down the EU or non-EU path, converts non-EU amounts at the Banca d'Italia rate and registers it in Fatture in Cloud. It asks a person on Slack only when VAT or bank details are missing, and the accountant reviews before anything goes to SDI, the Italian tax exchange. The company reports that nine in ten invoices now reach Fatture in Cloud untouched, as our invoice automation case study sets out.
The same travel insurance group sent its 25-page MI pack to around 50 people every morning, and a sharp drop in the quote funnel took 20 days to notice. An n8n watcher picks up each pack from OneDrive, works out which brand it belongs to and copies its 17 worksheets into a Supabase database, where database views calculate every figure. A language model only writes the sentence around the finished numbers, and the alerts reach six Microsoft Teams audiences by 12:00. Finance sets each metric's band itself, and made 34 threshold changes in six weeks with no developer ticket. Detection now takes 30 minutes, and the system caught 4 of 4 replayed incidents, as our daily MI alerting case study reports.
A Dutch implementation partner spent a year trying to build appointment scheduling for a dental care group inside the group's ERP, and across 200 practitioners the rules outgrew the ERP's data model. We built the scheduling in n8n on top of the ERP, with a planning copy of the data in Supabase. Five sync workflows run every night, an AI agent picks the best free slot, and each appointment it cannot place comes with a report saying why, so a planner picks it up with the reason in hand. The agent schedules about 1,500 appointments a week; the build is written up as a case study on our AI proof of concept page, and our n8n page lists it among our builds in production.
Our n8n page names three things every production workflow should carry: retries, an error workflow and an alert that reaches a person. In n8n, a node set to Retry On Fail reruns when it fails, up to five tries, as the retry code in n8n's repository sets it. If the run still fails, the error workflow named in the workflow's settings starts from an Error Trigger node and can send an email or Slack alert to a person, and the same error workflow can serve several workflows, according to n8n's documentation on error handling.
Execution history needs a limit as well, because each execution log holds the data that run touched. n8n prunes finished executions by default once they are 14 days old or once 10,000 are stored, and a self-hosted instance can stop saving successful runs altogether, according to n8n's documentation on execution data. Execution data pruned on a schedule is one of the checks we run on every n8n instance, so personal data does not pile up in logs.
Where the data is sensitive, self-hosting keeps the n8n editor, your credentials and every execution log on infrastructure you control, and it also makes the setup your responsibility. The invoice intake above runs on the company's own n8n instance, and every workflow, credential and log stays inside its account. A company can self-host the free Community edition for its own internal business purposes under n8n's Sustainable Use License, which n8n calls fair-code rather than open source, as n8n's license page explains.
Five of the six builds started from a figure the client already had: $700 an article, 45 staffed chat hours in a 168-hour week, more than 180 invoices a month keyed by hand, 20 days to notice a drop in the quote funnel, and a year spent on scheduling inside the ERP. Pick the workflow in your department whose cost you can state the same way, and map its trigger, its steps, the systems it touches and the point where a person checks the result before anything is built.
We design, build and run workflows and AI agents on self-hosted n8n as an official n8n Select Partner, and each workflow has a named owner. A senior partner leads from the first call, and the workflow is scoped in writing before you commit.
n8n automates workflows that a trigger starts and a fixed set of steps completes, and it can run AI agents inside them. In our builds that covers content pipelines in marketing, an after-hours chat agent in customer support, invoice intake and a daily alert on the MI pack in finance, and appointment scheduling in operations. Each step is a node that calls an API, moves a record, calls a language model or runs code.
Yes. The content pipelines run research, brief and content agents, the after-hours chat agent answers customers while a second agent, the Gatekeeper, checks every draft before the customer sees it, and the scheduling build uses an AI agent to pick each free slot. In each of these builds a person still checks the output or picks up what the agent could not settle, from the draft review in marketing to the planner who takes an appointment the agent could not place.
A node set to Retry On Fail reruns when it fails, up to five tries. If the run still fails, the error workflow set in the workflow's settings starts from an Error Trigger node and can send an email or Slack alert to a person, according to n8n's documentation on error handling. Our n8n page names retries, an error workflow and that alert as the three things every production workflow should carry.
Self-host when your data must stay in your own cloud or you run high volumes. Self-hosting keeps the n8n editor, your credentials and every execution log on infrastructure you control, and the invoice intake in this post runs on the company's own instance. It also makes the setup your responsibility, so plan the error workflow and the pruning schedule when you design the workflow.
Yes. The Code node runs JavaScript or Python as a step in a workflow, for logic no built-in node covers, according to n8n's documentation on the Code node. On a self-hosted instance you can import npm modules into the Code node once you enable them, while n8n Cloud offers two built-in JavaScript modules and no imported libraries for Python.
By default n8n prunes finished executions once they are 14 days old or once 10,000 are stored, according to n8n's documentation on execution data. A self-hosted instance can change both limits and can stop saving successful runs altogether. Each execution log holds the data that run touched, so a pruning schedule is one of the checks we run on every n8n instance.
Choosing your first n8n workflow?
Book a discovery call. A senior partner runs the call, and you bring the process.