n8n Automation Agency: Workflows and AI Agents Built, Hosted and Kept Running

Updated September 2026 · Written and maintained by the Progression Agency strategy team

An n8n automation agency designs, builds and runs business workflows on n8n, the fair-code platform that connects apps, APIs, databases and AI models through visual nodes, code and webhooks. Progression Agency maps each process with your operations team, builds the workflows with error handling, logging and access controls from the first version, and runs them on n8n Cloud or on infrastructure you own. Progression Agency is based in New York City and works with clients across the United States and worldwide.

What an n8n automation agency builds

Mostly the unglamorous work that holds a company together: moving data between systems, reacting to events, producing documents and reports, and routing decisions to the right person. Increasingly it also includes AI steps that read, classify or draft, wrapped in rules that decide what happens next.

n8n suits this work because one workflow can mix no-code nodes, HTTP requests to any API and code, and because the platform can run inside your own environment. Its integrations directory listed 2,279 integrations when we checked on September 30, 2026 (n8n integrations), and a system without a node is usually reachable through the HTTP Request node or a custom node.

• New-customer onboarding across the CRM, billing, email and project tools.
• Invoice and payment reconciliation between billing and accounting systems.
• Support ticket triage: AI classification, routing and a drafted first reply for an agent to check.
• Order and inventory sync between an online store and an ERP.
• Weekly KPI reports assembled from several systems into one table.
• Document intake: parse, validate, file and notify the owner.
• Employee onboarding and offboarding across software accounts.
• Lead enrichment, deduplication and routing into the CRM.

Operations workflows

Order-to-invoice handoffs, onboarding checklists that create accounts across several tools, supplier updates parsed from email, and approvals that pause for a manager and resume. The value is in removing retyping and the errors it causes.

Integrations that have no ready-made node

Internal APIs, legacy systems and niche software are reached through the HTTP Request node first and a custom node when the same connection is reused widely, so workflows do not have to change as the connection matures.

AI steps inside ordinary workflows

Classifying inbound requests, extracting fields from documents, summarizing threads and drafting replies are the usual AI tasks. Each one sits between deterministic steps that validate the output before anything is written to a system of record.

Workflows that act as internal endpoints

A Webhook node paired with a Respond to Webhook node turns a workflow into a small internal endpoint that forms, apps or other workflows can call. It is often the fastest way to give a team a working tool before anyone commits to custom software.

Should you run n8n Cloud or self-host it?

Use n8n Cloud to start quickly without running servers; self-host when data must stay in your environment, when you need more control over scaling and networking, or when you need features sold with the self-hosted editions. n8n frames it the same way: Cloud is fully managed and needs no technical expertise, while self-hosting gives full control but needs technical skill to install and configure (n8n docs: choose how to use n8n).

When is n8n Cloud the right call?

Small teams, moderate volumes, no hard data-residency rule and nobody who wants to be paged about a server. The Starter and Pro plans include 5 and 20 concurrent executions respectively (n8n pricing), which covers most back-office work.

When does self-hosting pay off?

High execution volumes, workflows that touch data you cannot send to a third party, private networks, custom nodes, or a need to scale workers independently. The trade is operational: backups, upgrades, security patches and monitoring become your job or ours.

What does the free Community edition leave out?

The self-hosted Community edition has no environments or Git-based version control, no external secrets, no log streaming, no SSO and no projects with role-based access, and workflows cannot be shared: only the owner and the creator can access them (n8n edition comparison). Registering the Community edition for free adds folders, debugging in the editor and custom execution data. These gaps matter more as the number of builders grows.

What the Sustainable Use License allows an agency to do

It allows an agency to design, set up and maintain n8n workflows for your business and to charge for that work; it does not allow anyone to resell n8n itself as a hosted service that clients log into. n8n’s license FAQ says consulting and automation services are explicitly permitted (n8n license FAQ).

n8n moved to the Sustainable Use License on March 17, 2022 (n8n docs). The license lets you use or modify the software only for your own internal business purposes or for non-commercial or personal use, and distribute it only free of charge for non-commercial purposes (license text). n8n describes itself as fair-code, distributed under the Sustainable Use License and the n8n Enterprise License (n8n on GitHub); fair-code means source anyone can read and extend, with commercial rights kept by the authors (faircode.io).

Where does the Enterprise license come in?

Files with .ee in their names fall under the n8n Enterprise License and need a valid license to use (n8n license). The paid capabilities, such as SSO, environments, external secrets and log streaming, come with the Business and Enterprise editions, so if a requirement depends on one of them, the plan decision comes before the build.

Who owns the instance and the workflows?

You do. We set n8n up in your account, under your license if one is needed, with your people as owners, and every workflow can be exported. Nothing we build depends on our continued involvement.

How is n8n priced?

By executions: one execution is one full run of a workflow however many steps it has, and n8n’s plans are sized by executions per month. Self-hosting the free Community edition has no license fee; you pay for servers and for the people who run them.

Per-step pricing works differently. Zapier counts a task for each successful action step, while triggers, filters and paths use none (Zapier help), and Make charges a credit for each module action in a scenario (Make pricing). A workflow with one trigger and nine actions therefore uses nine Zapier tasks per successful run and one n8n execution. That gap is why teams with busy, multi-step automations look at n8n, and why an honest comparison starts from your real run counts.

Costs that remain when you self-host

Servers, a PostgreSQL database, backups, monitoring and upgrade testing. For light workloads that is a small monthly bill; queue mode with several workers is a platform to operate, which our managed DevOps service covers. Our guide to what AI automation costs separates build cost from running cost.

Nodes, the Code node and custom nodes

Reach for the built-in node first, the HTTP Request node second, the Code node third, and a custom node when the same integration will be reused across many workflows. Each step down the list adds power and adds maintenance.

The Code node

n8n’s Code node runs JavaScript on Node.js and Python; native Python through task runners arrived in version 1.111.0 and is stable in n8n 2, which removed the older Pyodide approach (n8n docs: Code node). By default it runs once for all items, or once per item when chosen. We keep Code nodes short and tested, and move anything larger into a custom node or a service.

Building a custom node

n8n supports two building styles: declarative, the default and recommended starting point, and programmatic, which n8n treats as the exception for cases such as trigger nodes, non-REST APIs and heavy data transformation (n8n node building styles). Node development needs JavaScript, TypeScript and npm packaging, and a finished node can stay private, be shared with the community or be submitted for verification (n8n docs: create nodes).

Community nodes: useful, but vetted

Community nodes have full access to the machine n8n runs on and to the data in your workflows, and an update can break existing workflows. n8n verifies some community nodes against security requirements, and administrators can switch off installation entirely (n8n docs: community node risks). Our default is verified nodes only, pinned versions and a review before any upgrade.

AI agent workflows with approvals and tests

n8n’s AI Agent node lets a model choose among the tools you connect, and it needs at least one tool to work (n8n docs: AI Agent). We build agents inside workflows so every tool call is logged, risky actions wait for a person, and quality is measured against test cases before and after launch.

Human approval before risky actions

n8n can pause a workflow when an agent wants to use a tool that needs review, and wait for a person to approve or deny it through Slack, Microsoft Teams, Gmail, Outlook, Telegram, WhatsApp Business Cloud, Discord, Google Chat or n8n’s own chat (n8n docs: human-in-the-loop). We switch it on for anything that sends messages, changes records or deletes data, and relax it only when the logs justify it.

Evaluations, because models are not deterministic

n8n’s evaluations run a dataset of test cases through a workflow: a handful of examples while building, then a larger dataset drawn from production and scored with metrics once the workflow is live (n8n docs: evaluations). Every agent we ship has a test set, and prompt or model changes are checked against it before they reach production.

MCP in both directions

The MCP Server Trigger node lets n8n act as a Model Context Protocol server that exposes workflows as tools to MCP clients, with bearer or header authentication. With several webhook replicas, n8n says all /mcp traffic must go to one dedicated replica (n8n docs: MCP Server Trigger). n8n’s MCP Client Tool node goes the other way, giving an n8n agent tools hosted on external MCP servers.

The failure modes are well documented. OWASP’s 2025 Top 10 for LLM applications lists prompt injection first and excessive agency sixth (OWASP GenAI Security Project); narrow tool permissions, approval steps and output validation are how a workflow limits both. Agents that outgrow a workflow move to our custom AI agent builds, with the same testing discipline.

Error handling: what happens when a workflow fails

A failed run should raise an alert, keep its data and be safe to retry. In n8n that means an error workflow behind every production workflow, retries on flaky steps, and writes designed so a repeated run cannot post twice.

An error workflow starts with the Error Trigger node and runs automatically when an execution fails; one error workflow can serve many workflows, and the Stop And Error node lets a workflow fail on purpose when a business rule is broken (n8n docs: error handling). We send the failure, the execution link and the input to the channel your team watches, and log it where it can be counted. Every failure alert carries:

• The workflow name, the failing node and the error message.
• A link to the failed execution, where the execution was saved.
• The input item that failed, or a pointer to it in the holding table.
• Whether a retry is safe, and who owns the fix.
• A count of similar failures in the last day, so one-offs and outages look different.

Retries that cannot double-post

Before a workflow creates an invoice, sends an email or records a payment, it checks whether that action already happened, using an idempotency key or a lookup. A retry after a timeout then produces one invoice, not two.

Holding tables instead of silent drops

Items that fail validation go to a holding table with the reason attached, and someone owns reviewing it. Nothing disappears, and the table doubles as a list of process problems worth fixing upstream.

Queue mode: scaling n8n beyond one server

Queue mode splits n8n into a main instance that receives triggers and webhooks, a Redis queue and worker processes that run the executions, so capacity grows by adding workers. It needs PostgreSQL: n8n says queue mode on SQLite is not recommended and a distributed setup on SQLite is not supported (n8n docs: queue mode).

The main instance handles timers and webhook calls and creates executions without running them; workers take execution IDs from Redis, run the workflow and write results to the database. Every worker needs the main instance’s encryption key to read credentials. Worker concurrency defaults to 10, n8n recommends 5 or higher, and optional webhook processors take incoming requests off the main instance.

Database and execution data

n8n uses SQLite by default and supports PostgreSQL, currently versions 17 and 18, plus 16 for compatibility (n8n docs: database). Execution history grows quickly, so pruning is on by default with a 336-hour (14-day) age limit and a 10,000-execution count limit (n8n docs: execution data). We tune what gets saved, keeping failures and trimming routine successes.

Concurrency limits on a single server

Without queue mode, production concurrency control is disabled by default; setting N8N_CONCURRENCY_PRODUCTION_LIMIT queues the excess and processes it first in, first out (n8n docs: concurrency). It is a simple guard for one server receiving bursts of webhooks.

Securing a self-hosted n8n instance

Treat n8n as what it is: a system holding credentials to most of your other systems. Harden it before the first production workflow and audit it on a schedule.

n8n’s security audit produces five reports (credentials, database, file system, nodes and instance) covering unused credentials, SQL expressions, risky nodes, unprotected webhooks and outdated versions, and runs from the CLI with n8n audit, through the API or as a node (n8n docs: security audit). Once enabled, SSRF protection blocks private, loopback and link-local ranges (n8n docs: SSRF protection); the rest of the checklist is in n8n’s security documentation.

n8n marketing: what an n8n marketing agency automates

An n8n marketing agency uses n8n to move leads, content and reporting data between marketing tools without per-step fees, and puts a review step wherever AI drafts copy. Strategy (who gets which message, and when) comes first; the automation carries it out.

“n8n marketing agency” draws about 1,300 US searches a month with no advertiser bids recorded, while “n8n marketing” draws about 30 at a $4.08 cost per click (Ubersuggest, September 2026). n8n’s public template library has a marketing category with thousands of community workflows (n8n marketing templates): a good place to see what is possible, and a poor place to copy production logic from.

Lead capture to CRM in seconds

Form fills, ad leads and chat transcripts are normalized, enriched, deduplicated and pushed to the CRM with source fields intact, then routed to the right owner. Completeness matters more than speed: a lead with no source is a lead nobody can credit. Our CRM consulting team sets the rules the workflow enforces.

Reporting pipelines

Ad spend, CRM stages and web analytics land in one table on a schedule, so the weekly report builds itself and the numbers match across teams.

Content and social operations with a review step

AI drafts, a person approves, n8n publishes and logs. Approval is not optional for anything public, which is exactly what n8n’s human-in-the-loop feature is for.

Nurture programs and lifecycle email are designed by our lifecycle marketing automation team; n8n often carries the data they depend on.

Automation software like n8n: how the alternatives compare

The main alternatives are Zapier and Make for hosted no-code automation, Microsoft Power Automate for Microsoft 365 organizations, Activepieces and Node-RED as open-source options, and custom code. The right choice depends on run volume, hosting requirements and who will maintain the workflows.

Sources: Zapier task usage, Make pricing, Microsoft Learn: Power Automate, Activepieces on GitHub and Node-RED.

When is Zapier or Make the better answer?

When a handful of simple automations run a few hundred times a month and the people maintaining them are not technical, the easiest tool wins. Moving them to n8n would add hosting and skills without saving much.

When does custom code beat any platform?

Logic that is core to your product, needs full test coverage or must scale predictably belongs in code, often behind an endpoint that n8n calls. Our custom software team builds that layer.

How ops leaders ask AI assistants to find an n8n expert

Buyers describe the job, the volume and the constraint, then ask ChatGPT, Claude, Perplexity, Gemini or Copilot who can do it. The assistants answer from pages written at that level of detail, from n8n’s own documentation, and from community threads and list articles.

Typical prompts

The wording is operational, and it usually names the constraint that rules providers in or out:

• “Find an n8n expert to self-host n8n in queue mode on AWS with PostgreSQL.”
• “Agency to move our Zapier zaps to n8n and cut our task bill.”
• “Who builds n8n AI agents that ask a manager before emailing customers?”
• “n8n developer to write a custom node for our internal inventory API.”
• “Is n8n or Make cheaper for multi-step workflows, and who can set it up?”
• “n8n marketing agency to sync HubSpot, Google Ads and our data warehouse.”

What gets cited

When we searched “n8n automation agency” on September 30, 2026, the results mixed specialist agency pages, n8n’s own AI agent pages and a list article ranking n8n agencies. Assistants lean on the same material: n8n’s documentation and pricing for facts, community discussions and list articles for names, and agency pages with node-level detail for the recommendation itself.

What to publish so assistants name you

Write workflows up at the level a buyer asks about: the trigger, the nodes, the error path, the hosting model and the run volume. Explain how your hosting model fits the Sustainable Use License, publish price ranges, and keep crawlers welcome. Perplexity says PerplexityBot surfaces and links websites and is not used to train foundation models (Perplexity); Anthropic runs a separate Claude-SearchBot for search quality (Anthropic); and Microsoft Copilot draws web results from the Bing search service (Microsoft Learn), which takes IndexNow notices for new pages. Our generative engine optimization team applies this to service companies.

How to choose an n8n automation partner

Choose on evidence of production discipline, not on the number of templates someone has published. Ask how they handle failure, security, licensing and handover, and check the answers.

A partner who cannot say when n8n is the wrong tool will fit every problem to it.

What n8n development costs

As published planning ranges, a set of production workflows costs $15,000–$50,000 over four to ten weeks, a complex integration or custom node $15,000–$40,000, and an AI agent proof of concept $15,000–$40,000. n8n subscription or server costs are separate.

These are planning ranges, not quotes; the quote follows a written scope that lists every workflow, system and volume.

Our n8n build process and timeline

A first set of workflows goes from process mapping to monitored production in four to ten weeks. Each stage ends with something you can check.

Services that connect to your n8n work

n8n rarely works alone. These services cover the systems and decisions around it.

• AI consulting and automation for small businesses: the wider program when n8n is one tool among several.
• Custom AI agents: agents that outgrow a single workflow.
• APIs and integrations: the endpoints your workflows call.
• Lifecycle marketing automation: the strategy behind marketing workflows.
• CRM consulting: the routing and data rules lead workflows enforce.
• Managed DevOps: hosting, monitoring and upgrades for self-hosted n8n.
• AI integration and AI chatbots that hand work to workflows.
• Dashboards fed by reporting workflows.
• Python development for services your Code nodes outgrow.
• Salesforce development when workflows meet Salesforce limits.
• What AI automation costs: build versus running cost, explained.

Frequently asked questions