What an AI automation actually costs per month

The AI model is the cheapest part of an AI automation. At small-business volume, a single classification or extraction call costs a fraction of a cent, and even a few thousand runs a month lands the model bill under five dollars. The number that decides your monthly cost is the automation platform's billing unit, how it counts your steps, and how often the workflow runs. The model you pick is a distant fourth lever.

Most cost guides answer "what does AI automation cost" with a flat range, usually 300 to 1,500 dollars a month. That range is not wrong, it is just unhelpful, because it bundles one-time build labor, a whole stack of automations, and high run volume into one number. If you want to know what a single automation costs to keep running, you have to price the three things that actually move the bill: the platform's unit of billing, the number of billable steps in the flow, and the run volume. The model tokens are a rounding error sitting underneath all three.

The model tokens are a rounding error

Take a common build: an automation that reads each inbound lead, classifies and scores it, and routes it. The AI step sends maybe 1,500 input tokens (the lead details plus your instructions) and gets back 200 output tokens (a score and a category).

On Claude Haiku 4.5, input runs 1 dollar per million tokens and output runs 5 dollars per million, as of June 2026. That is 0.0015 dollars of input and 0.001 dollars of output, so about 0.0025 dollars per run. On GPT-5 mini, at 0.25 dollars per million input and 2 dollars per million output, the same call is about 0.0008 dollars. Run that 1,000 times in a month and your model cost is 2.50 dollars on Haiku or 78 cents on GPT-5 mini.

That is the whole AI bill for a thousand handled leads. Prompt caching and batch processing cut it further. So when someone tells you AI automation is expensive because "the AI is expensive," they are pointing at the smallest line on the invoice.

Your platform charges by a unit, and the unit is everything

Here is where the real money is, and where the three big platforms diverge hard. They do not bill the same way, and the billing unit changes the answer by 10x or more for the exact same workflow.

What you pay for	Zapier	Make	n8n cloud	n8n self-hosted
Billing unit	per action step (task)	per module run (operation)	per full workflow run (execution)	nothing per run
Does the trigger count?	No	Yes, one operation	No, it is part of the execution	No
Do filters and routing count?	No, they are free	No, a filter blocks the next module	No	No
6-step lead-triage run	about 4 tasks	about 5 operations	1 execution	1 execution

The detail that catches people is what each platform considers billable. Zapier only counts action steps that actually send data to an app: triggers, filters, formatter, and paths are free, so a flow with an AI call, a CRM lookup, a CRM write, and a Slack alert is four tasks. Make counts every module that runs, including the trigger, so the same flow is five or six operations. n8n counts the entire run as one execution no matter how many nodes you wired up.

Now multiply by volume. At 1,000 runs a month, that is roughly 4,000 Zapier tasks, 5,000 Make operations, or 1,000 n8n executions for the identical automation. Make's Core plan at about 10.59 dollars a month (annual) includes 10,000 operations, so this fits. n8n's Starter at 24 euros a month includes 2,500 executions, and self-hosted n8n on a small 5-dollar VPS has no per-run charge at all. Zapier's 4,000 tasks push you well past the entry tier into a paid plan costed by task volume. Same logic, same model, same lead, and the platform line ranges from near-zero to the largest cost in the build.

Where the bill actually spikes

Two things blow up an AI automation bill, and neither is the model.

The first is step count multiplied by run volume on a per-step platform. Every step you add to a Zapier or Make flow is not a one-time cost, it is a tax on every single run forever. A flow that grows from four steps to eight has doubled its per-run bill. On n8n that same change costs nothing extra, because the execution count did not move. This is why "just add one more step" is a bigger decision on per-step platforms than it feels like in the editor.

The second is loops. When a step processes a list, each item is usually billed on its own. An iterator that handles 50 line items is 50 operations in Make, not one. A Zapier loop that fans out to 50 records runs its downstream actions 50 times. I have seen a flow that tested at three operations on a single sample record bill three hundred on a real batch. If your automation touches line items, order rows, or search results, price it against a realistic batch size, not the one clean record you used to build it.

Built-in AI features add their own meter. Make's AI modules consume credits scaled by token usage, and a Zapier MCP tool call counts as two tasks rather than one. None of this is hidden, but it means the platform's accounting, not the underlying model price, is still what you are optimizing.

So which model should you pick

Pick the cheapest model that is accurate enough for the job, then stop thinking about model price. For classification, extraction, routing, scoring, and short rewrites, a budget model like Haiku 4.5 or GPT-5 mini is almost always accurate enough, and the difference between it and a flagship model is a dollar or two a month at the volumes most operators run. Reserve the expensive models for the steps that genuinely need stronger reasoning, and let the cheap one handle the routing around them.

The mistake is treating model selection as the cost decision. It is the accuracy decision. The cost decision is the platform and the architecture. If your bill is too high, do not downgrade the model, because there is almost nothing to save there. Collapse steps, filter early so fewer modules run, batch where you can, and if you are running real volume, move the workflow to a per-execution model or self-host it. Those moves change the bill by multiples. A cheaper model changes it by pennies.

How to estimate your own automation in five minutes

Before you build, do this math instead of trusting a range. It takes five minutes and it is usually accurate within a few dollars.

First, count the billable steps in the flow for your chosen platform, using the rules above: action steps for Zapier, module runs for Make, one execution for n8n. Second, multiply by your realistic monthly run volume, and if any step loops over a list, multiply that step by the batch size. Third, map that total to the platform's plan tiers to get the platform cost. Fourth, add the model cost: input tokens plus output tokens per run, times volume, at the published rate for your model. The fourth number will almost always be the smallest, and that is the point.

If the platform line comes back larger than you want, that is a signal about architecture, not about AI. The fix lives in how the workflow is built and where it runs. If you want a second opinion on whether an automation is structured to stay cheap as it scales, that is the kind of tradeoff we work through on workflow automation systems, and the n8n vs Zapier vs Make decision covers the other half of the choice: where your data of record should live. When you are ready to price a specific build, tell us what it has to do and we will cost it by the unit, not by a range.