How Much Does AI Automation Cost to Build and Run?
What AI automation costs to build and run, what n8n and AI model calls cost per workflow run, and how to pick the workflow that pays back fastest.
By Sameer Rahman- 8 min read
On this page
Most people ask what an automation costs and expect one number. In practice you pay for three different things, and they behave very differently. The build is a one-time cost. Running costs come every month, and they're usually tiny. Upkeep is what keeps the whole thing working a year from now.
This guide breaks down all three with real platform prices. You'll see what n8n and AI model calls cost per run, when self-hosting saves money, what makes a build bigger, and how to pick the workflow that pays for itself fastest.
Key takeaways
- AI automation has three costs. The build, the running costs, and upkeep. The build is usually the biggest by far.
- Running costs are small. n8n Cloud starts at about $22 a month for 2,500 runs, and an AI step costs about $0.0002 a run on Claude Haiku 5.5.
- n8n charges per workflow run, not per step, so a careful 20-step workflow costs the same to run as a 2-step one.
- Self-hosting n8n is free software on a $12 to $24 server, but someone has to own updates, backups, and security.
- Pick by payback, not build price. In our example, a $6,000 onboarding workflow pays back faster than a $3,000 reporting one.
How much does AI automation cost?
AI automation costs a one-time build fee, then running costs that are often under $100 a month, plus monthly upkeep. The build is priced by how many workflows you need, how many tools each one connects, and how many exceptions it has to handle.
AI automation is software that does repetitive work across the tools you already use, with AI handling the steps that need reading or judgment. For example, a workflow might read an incoming invoice, pull out the amounts, check them against the purchase order, and enter them in your accounting tool.
| Cost | When you pay it | What decides it |
|---|---|---|
| Build | Once | Number of workflows, tools, exceptions, and AI steps |
| Platform | Monthly | n8n plan or server, based on how often workflows run |
| AI model calls | Monthly | How much text each AI step reads and writes, and which model |
| Upkeep | Monthly | Monitoring, fixes when tools change, and small changes |
What it costs to run
Running costs are the platform plus the AI model calls. Both are cheaper than most people expect.
The platform
We build on n8n. It bills per workflow run (opens in a new tab), and a run counts once however many steps it has. Here's what that looks like right now. n8n prices its plans in euros, so these are rounded to dollars.
| Option | Monthly cost | Workflow runs | Who looks after it |
|---|---|---|---|
| n8n Cloud Starter | About $22, billed annually | 2,500 | n8n |
| n8n Cloud Pro | About $56, billed annually | 10,000 | n8n |
| Self-hosted Community edition | Free software, plus a server at $12 to $24 (opens in a new tab) | No limit | You |
On Pro, each run costs about half a cent. Because steps don't count, you can build the careful version of a workflow, with checks, retries, and alerts, without paying more each time it runs. That's a big difference from tools that bill every action. For more on that, see n8n vs Make vs Zapier.
When self-hosting saves money
The self-hosted Community edition is free, and a server with 2 to 4 GB of memory runs $12 to $24 a month on DigitalOcean. So on paper, it's the cheapest option at any volume.
The catch is the work. n8n's own docs list maintenance of a self-hosted setup as "your responsibility" (opens in a new tab), and recommend n8n Cloud if you don't have the technical expertise. Updates, backups, security, and fixing it when the server goes down all land on someone.
So self-host when you already have someone who looks after servers, or when your data has to stay on your own infrastructure. Otherwise, n8n Cloud costs a few dollars more and takes that work off your plate.
The AI steps
AI model calls are billed by the token, a small chunk of text, roughly part of a word. At Claude's list prices (opens in a new tab), Haiku 5.5 costs $0.10 per million tokens in and $0.50 out, and Sonnet 5.5 costs $2 in and $10 out.
| AI step | Haiku 5.5 per run | Sonnet 5.5 per run | 10,000 runs on Haiku | 10,000 runs on Sonnet |
|---|---|---|---|---|
| Sort an email (1,000 tokens in, 200 out) | $0.0002 | $0.004 | $2 | $40 |
| Read a long document (10,000 in, 300 out) | $0.0012 | $0.023 | $12 | $230 |
Sonnet costs about 20 times more per step. So use the smaller model for routine steps like sorting, tagging, and pulling out fields, and save the bigger one for steps where a wrong call is expensive. Most workflows only need the bigger model on one step, if any.
What makes the build bigger
The build is where most of the cost is, and it follows the work. Here's what adds to it.
More tools. Every app the workflow connects to needs its own connection, field mapping, and testing. A workflow between two tools is a much smaller build than one that touches five.
Writing, not just reading. Pulling data out of your CRM is simple. Creating and updating records means handling duplicates, missing fields, and what to do when the update fails.
Exceptions. The normal path is quick. The invoice in a different format, the client with no email address, the form someone filled in twice. Each exception needs a rule, or a handoff to a person.
AI decisions. An AI step that sorts or reads needs a set of real examples with known right answers, and testing against them.
Error handling. Retries, alerts, and a clear record of what failed and why. This is what separates an automation you can forget about from one you have to babysit.
What the manual work costs you
The cost that matters is the one you're paying now. Doing the work by hand costs staff time on every run, every month.
For staff cost, use what the people doing the work cost you, benefits included. The private-industry average was $46.89 an hour (opens in a new tab) in June 2026, or about $0.78 a minute.
| Workflow | Runs per month | Minutes per run | Hours per month | Manual cost per month |
|---|---|---|---|---|
| Client onboarding | 20 | 90 | 30 | About $1,410 |
| Invoice entry | 400 | 4 | About 27 | About $1,250 |
| Weekly reporting | 4 | 120 | 8 | About $375 |
Two of those cost over $15,000 a year in staff time, and that's before the mistakes. A missed onboarding step or a mistyped invoice costs more time to find and fix than it took to make.
How to pick the workflow that pays back first
Once you know the manual cost, you can work out the payback for each candidate. It's one division.
Say the three workflows above are quoted like this.
| Workflow | Build | Running and upkeep per month | Monthly saving | Months to pay back |
|---|---|---|---|---|
| Client onboarding | $6,000 | $60 | About $1,350 | About 4.5 |
| Invoice entry | $8,000 | $80 | About $1,170 | About 7 |
| Weekly reporting | $3,000 | $30 | About $345 | About 9 |
The cheapest build pays back slowest. Weekly reporting looks like the easy first project, but at eight hours a month it takes twice as long to pay for itself as onboarding.
So rank your candidates by hours per month first, then compare payback. The best first workflow is usually a long task that happens often enough to matter, not the one that's quickest to build.
What upkeep covers
Automations break for boring reasons. A tool updates its API, someone renames a field in the CRM, a password changes, or a new kind of input shows up. None of that is the workflow's fault, but someone has to notice and fix it.
Good upkeep comes down to three things.
- Monitoring. An alert the moment a run fails, not when a client asks why they never got their welcome email.
- Fixes. Updating the workflow when a connected tool changes.
- Small changes. New fields, new steps, and tweaks as your process changes.
This is a monthly cost, and it's what keeps a $6,000 workflow saving $1,350 a month in year two.
Common mistakes when budgeting for AI automation
- Starting with the easiest workflow. Easy to build often means few hours saved. Run the payback numbers first.
- Using the biggest AI model for every step. Sorting and tagging work fine on a small model, at a twentieth of the price.
- Self-hosting to save $30 a month. If nobody owns the server, the first outage costs more than years of n8n Cloud.
- Skipping error handling. A workflow that fails silently is worse than none, because everyone assumes the work got done.
- No budget for upkeep. Every tool you connect will change at some point. Plan for it from day one.
Conclusion
AI automation costs a one-time build, small monthly running costs, and upkeep. On n8n, running costs start at about $22 a month, and the AI steps cost fractions of a cent per run. The build is the real cost, and it follows how many tools, exceptions, and decisions the workflow has to handle. Work out what the manual version costs you, then start with the workflow that pays back fastest.
We build AI automation on n8n in your own accounts, with monitoring and alerts from day one, and keep it running after launch. Every project is scoped and priced after a discovery call. For the full picture across all four services, see what it costs to hire an AI agency.
Frequently asked questions
You pay for the build once, then running costs and upkeep each month. Running costs are small, often under $100 a month, since n8n Cloud starts at about $22 a month and an AI step costs a fraction of a cent per run. The build is the main cost, and it follows how many workflows, tools, and exceptions are involved.
n8n Cloud is about $22 a month for 2,500 workflow runs on Starter and about $56 for 10,000 on Pro, billed annually. The self-hosted Community edition is free, and runs on a server from about $12 to $24 a month.
A step that reads about 1,000 tokens and writes 200 costs about $0.0002 on Claude Haiku 5.5 and $0.004 on Claude Sonnet 5.5. 10,000 runs comes to about $2 on Haiku or $40 on Sonnet.
On fees, yes. The software is free and a server costs about $12 to $24 a month. But you take on updates, backups, and security, so it only saves money when someone already looks after servers.
The one with the fastest payback, not the cheapest build. Multiply runs per month by minutes per run to get the hours it saves, then divide the build cost by the monthly saving after running costs.
Sources
- n8n, Pricing (opens in a new tab), read November 10, 2026.
- n8n Docs, Choose how to use n8n (opens in a new tab) and Hosting n8n (opens in a new tab), read November 10, 2026.
- DigitalOcean, Droplet pricing (opens in a new tab), read November 10, 2026.
- Anthropic, Claude pricing (opens in a new tab), read November 7, 2026.
- U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026 (opens in a new tab), released September 9, 2026, read November 10, 2026.