n8n Data Pipelines Explained: Split Out, Aggregate, Switch, Loops and AI Agents on Real Workflows

Short answer

Real workflow: a sales data pipeline with three outputs

What it does. This workflow was built on the training dataset of n8n's official course, so it uses the same data shapes as real e-commerce exports.^[9] It downloads a sales file containing a list of orders. It splits the list into one item per order and keeps the fields that matter. Then it produces three results: all orders sent to a data warehouse, a summary of delivered orders by region (total, count, average), and the same summary as a CSV file with a timestamp.

Notice that one node's output can feed several branches at once. The order list feeds both the raw export and the regional analysis, so one download serves three audiences. In a real business, you would replace the manual start with a schedule or webhook trigger.

Split Out, Aggregate, Summarize: the shape-shifters

If you've read how JSON works, you know a file can contain a list inside a field, such as one sales file with fifty orders, or one invoice with eight product lines. These nodes change how many items you have, without changing the data itself.

The rule of thumb: split when you need to look at each line (check, enrich, route), and aggregate or summarize when the next system wants one message (one API call, one email, one row). The maintenance-company example in part 2 uses exactly this pair: Split Out on the parts and hours of a job report, then Aggregate into one invoice.

If, Switch, Filter, Stop And Error: routing decisions

• Filter keeps matching items and drops the rest. Use it when the others don't matter, such as orders that aren't paid yet.^[13]
• If splits items into a true and a false branch. Use it when both sides need handling, for example delivered orders to the analysis and the others to a follow-up.^[14]
• Switch sends items to one of many routes based on rules, with an optional fallback.^[15] Typical uses: support tickets by language, orders by country to different couriers, files by type.
• Stop And Error deliberately fails the run with your own message, which also fires the error workflow.^[16][17]

Real workflow: a healthcare CSV pipeline that adapts itself

What it does. This is our most complete pipeline, built on public hospital datasets.^[9] When a CSV lands in a Google Drive folder, the workflow checks the file's size. Small files are downloaded, cut into chunks and inserted into a MySQL database. Large files go to the self-calling webhook from part 2, which streams them in 2 MB slices. A Switch node then recognises the dataset type from its columns: cardiology admissions, hospital-level data or weekly respiratory data. Each type gets its own cleaning and KPI queries. Unknown formats go to a fallback route that calls an auto-adapt sub-workflow. It infers the column types, creates the tables, cleans the data, and emails a warning if some columns are ambiguous.

Sub-workflows: build a step once, use it everywhere

The auto-adapt step is a separate workflow that starts with an Execute Sub-workflow Trigger. Both the normal pipeline and the streaming webhook call it with the Execute Sub-workflow node.^[18][19] n8n's documentation describes sub-workflows as a way to build modular, “microservice-like” workflows, which also helps when a workflow grows too large for memory.^[7] For a business, that means one “clean customer file” or “create invoice” step shared by every workflow that needs it, so a fix happens in one place.

Where else this pattern fits

• Retail: each supplier sends price lists in a different format; a Switch on the column names routes each one to the right cleaning step before prices update.
• Accounting: bank exports land in a folder; transactions are stored, categorised and summed per month without opening a spreadsheet.
• HR: monthly timesheet exports are checked, and any row with impossible hours triggers Stop And Error and an alert to the manager.

Looping in n8n: three ways, and when to use each

People coming from spreadsheets or code often look for a “for each” block. In n8n you rarely need one, because nodes already run once for every incoming item.^[2]

The documentation lists a few exceptions where n8n does not loop for you. For example, the HTTP Request node doesn't page through results by itself, and the Code node in “run once for all items” mode receives everything at once.^[2]

Real workflows: AI agents that answer from your data

What it does. Once the healthcare data is clean, anyone on the team can ask a question in Slack, for example “average length of stay for emergency admissions in March?”, and get the answer in the same thread.^[9] The agent has two tools, both small sub-workflows: one lists the available tables and their columns, and one runs a database query. Its memory is keyed to the Slack thread, so follow-up questions keep their context.

• AI Agent: the model reads the question, decides which tool to call, reads the result and writes the answer.^[8]
• Call n8n Workflow Tool: turns any sub-workflow into a tool the agent can use, with a description that tells the model when to use it.^[21]
• Simple Memory: keeps the last messages of a conversation. Using the Slack thread as the key gives each discussion its own memory.^[22][23]

A second, smaller agent from the same n8n course dataset handles customer service in a chat window. It uses a Data Table tool to look up customers and an HTTP Request Tool to fetch order details.^[24][25][26] The pattern is the same: a trigger where people already talk, an agent, and tools limited to the data it needs.

Why it matters: in a 2025 study of 5,172 customer-support agents, an AI assistant increased issues resolved per hour by 15% on average, with the biggest improvements for less experienced workers.^[29] An internal data assistant applies the same idea to every colleague who used to wait for “the person who knows SQL”.

Apply it: five everyday business pipelines

Each of these replaces a weekly or daily routine with one that runs on its own and only asks for attention when something is wrong. Add the error-alert workflow from part 2, and for outreach and reporting ideas see part 3 on lead generation and marketing reports.