Costs, Monitoring and Launch Checklist
Your project now reads messages, looks up data, handles errors and waits for approval. The last job is to run it safely over time. This lesson shows you how to work out the cost, what to monitor, and how to launch in stages.
What You'll Learn
- How n8n, Make and Zapier count usage
- How to estimate the cost of one AI run
- What to watch each week
- How to launch without risking every customer at once
You Pay Twice
An AI workflow has two separate bills: the automation platform and the AI model.
Plans and prices change often. Check each tool's current pricing page before you decide.
| Criteria | What it counts | What this means for you |
|---|---|---|
| n8n Cloud | Workflow executions. One run is one execution, however many steps it has | Long workflows cost the same as short ones. Frequent triggers add up |
| n8n self-hosted | Your server cost | Cheap at small scale, but you handle updates and backups |
| Make | Credits, usually per module run | Long scenarios and frequent checks cost more |
| Zapier | Tasks, counted per successful action | Each action step on each item adds to the count |
| AI model | Tokens in and out, billed by the provider | Long prompts, memory and agents cost more |
What it counts
- n8n Cloud
- Workflow executions. One run is one execution, however many steps it has
- n8n self-hosted
- Your server cost
- Make
- Credits, usually per module run
- Zapier
- Tasks, counted per successful action
- AI model
- Tokens in and out, billed by the provider
What this means for you
- n8n Cloud
- Long workflows cost the same as short ones. Frequent triggers add up
- n8n self-hosted
- Cheap at small scale, but you handle updates and backups
- Make
- Long scenarios and frequent checks cost more
- Zapier
- Each action step on each item adds to the count
- AI model
- Long prompts, memory and agents cost more
If you use the AI built into Make, token usage can also come out of your Make credits. Make AI agents use credits for each operation too. Check the credit usage page for your setup.
Estimate One Run
Tokens are small pieces of text. As a rough guide, 1,000 tokens is about 750 English words. Use this method:
- Open a real run and find the token count on the AI node, or in your provider's dashboard.
- Multiply input tokens by the input price. Multiply output tokens by the output price.
- Add up every AI step in the run.
- Multiply by your expected runs per month.
An agent makes several model calls, and each call resends the conversation so far. One agent run can cost many times more than one classify step. Measure it. Do not guess.
Use AI to check your math, then confirm the prices on the provider's own pricing page.
Four Ways to Cut Cost
- Filter before AI. Use a free IF node to drop spam and auto-replies before any model call.
- Use a small model for easy steps. Classification often works on a cheaper model. Keep a stronger model for drafting if tests show that it helps.
- Send less text. Cut signatures and old email threads from the input.
- Set a spending limit in your AI provider's account, so a loop cannot create a large bill.
What to Monitor
A workflow can fail without an error. It can run every day and give worse answers. So track two kinds of signals.
Check these once a week. It takes about ten minutes.
| Criteria | Signal | Where to find it |
|---|---|---|
| Did it run? | Failed runs, and runs that did not happen | n8n Executions list and your error workflow |
| Is it good? | Percent of drafts approved | The decisions you log in your table |
| Is it affordable? | Tokens and runs per week | AI provider dashboard and platform usage page |
| Is it stuck? | Items waiting for approval too long | The status column in your table |
Signal
- Did it run?
- Failed runs, and runs that did not happen
- Is it good?
- Percent of drafts approved
- Is it affordable?
- Tokens and runs per week
- Is it stuck?
- Items waiting for approval too long
Where to find it
- Did it run?
- n8n Executions list and your error workflow
- Is it good?
- The decisions you log in your table
- Is it affordable?
- AI provider dashboard and platform usage page
- Is it stuck?
- The status column in your table
n8n also has an Insights area that shows production runs, failures and failure rate over time. What you can see depends on your plan.
One failure that the error workflow cannot see is a workflow that never starts. A login may expire on the trigger, or a webhook may change. Add a small scheduled check: if no messages have been handled in 24 hours on a working day, send an alert.
AI Automation Fundamentals explains this heartbeat habit in more detail.
Launch in Stages
- 1. ShadowRuns on real messages, sends nothing
- 2. Approve allA person checks every draft
- 3. One categoryRemove approval for one safe type
- 4. Review weeklySample results, widen slowly
Move to the next stage only when the numbers support it. For example, wait until 95 percent of drafts in one category are approved for a few weeks. Choose your own number before you start, so you are not tempted to change it later.
Launch Checklist
Before you turn on the workflow, confirm each item:
- All bad inputs from your test set follow the right path.
- AI steps return fields, and unexpected values go to a person.
- Read steps retry. Write steps do not repeat.
- An error workflow is set, and you have tested the alert.
- Risky actions need approval, and each approval has a time limit.
- A unique ID stops the workflow from handling an item twice.
- The AI gets only the customer data it needs.
- Agent steps have Max Iterations and few tools.
- A spending limit is set with your AI provider.
- One person is named as the owner of this workflow.
Save a copy of each working version, so you can go back if a change causes harm.
Key Takeaways
- AI workflows have two bills: the platform and the model.
- Measure token use on real runs, especially for agents.
- Filter before AI, use small models where they pass tests, and set a spending limit.
- Monitor whether it ran, whether it is good, and what it costs.
- Launch in stages, starting with no customer impact, and name an owner.