What does automation cost? Count tool calls, not workflows

What does automation cost with an AI agent?

What does automation cost with AI agents? It hinges on a number no demo shows you: tool calls. A tool call happens each time the model looks something up or takes an action, such as fetching an order or writing a note. Ask what automation costs and you usually get a price per workflow or per month. For a fixed workflow that still holds, because you know in advance how many steps it takes. For an agent it does not. An agent decides while working how many tool calls it needs, and in doing so it partly decides your bill.

Platforms bill for that in very different ways. On Zapier, an AI step costs the model rate once, and then again for every tool call. On n8n, one turn with an agent is one execution, however many tools it calls. The same agent can cost a fixed amount on one platform and grow with every extra step the model takes on another.

So do the maths before the demo. Count how many tasks come in each month, how many tool calls the agent needs per task on average and how far that number can spike. With those numbers you can compare platforms. Below, n8n, Zapier, Make and Notion sit side by side, with their own pricing and help pages as the source. After that comes a worked example you can rerun yourself.

Which billing unit does each platform use?

Every platform picks its own unit to charge you by. That unit decides what happens when an agent takes more steps, and that is where the platforms differ. The model you plug in is easier to swap. The billing unit only changes when you move to another platform.

Sources: n8n pricing, Zapier task usage rates, Make operations, Notion Workers pricing, Microsoft. Accessed 3 October 2026.

n8n puts it most plainly: an execution is one run of your entire workflow, "regardless of complexity". Zapier sits at the other end. Its rates page says the agent decides which tools to use at runtime, and each tool call is billed separately. Make sits in between. A scenario with fixed modules is easy to predict, until more bundles pass through a module than you expected.

Microsoft belongs on this list for a different reason. If you have a Copilot licence, you might assume AI is paid for. For everyday work, it is. But according to Microsoft, Cowork, Code and Autopilot, the new agentic parts, are billed on usage. No prices are listed. So with Microsoft too, you pay differently for an agent than for an assistant.

How does Zapier bill an AI step with tool calls?

Zapier publishes the formula itself: tasks per run = (1 × model rate) + (number of tool calls × model rate). The model rate depends on the tier you pick. Standard costs 1 task per step but cannot use tools. Advanced costs 3, Premium 5. Zapier's own example: Premium with two tool calls comes to 5 plus 10, so 15 tasks for a single AI step.

New AI steps on a paid plan default to Premium. If you change nothing, you pay the highest rate. And since Standard cannot use tools, any agent that needs to look something up needs at least Advanced.

Tasks per run for one AI by Zapier step. Our own calculation using Zapier's formula and rates. With your own AI account (your own key with a model provider), you also pay the model costs directly to that provider.

At fourteen tool calls on Premium, a single run reaches 75 tasks. That is exactly where Zapier pauses the step by default and asks for approval. You can raise that limit to 500 tasks. It is a useful brake, and also a signal: Zapier itself expects runs to climb that high.

If you use Zapier through MCP, from an AI assistant such as Claude or Cursor, a different rate applies: 2 tasks per tool call. Regular steps outside the AI cost 1 task each. A number of built-in steps cost nothing, and that is exactly where you can save.

What do you pay for an agent on n8n?

On 25 September, n8n launched Agents as a product type of its own, alongside workflows. You describe the task in plain language, choose a model and give the agent the workflows and tools it may use. On cost, the announcement is brief: one turn with an agent is one execution, and tool calls to your workflows and sub-agents do not count separately. Agents and workflows draw on the same quota.

The subscription itself is measured in executions. Billed annually, it comes to €20 per month for Starter, €50 for Pro and €667 for Business. The table below shows how many executions per month you get for that. n8n hosts Starter and Pro itself, on servers in Frankfurt. According to the same page, Business is for now only available if you run n8n on your own server.

Source: n8n pricing, accessed 3 October 2026. Monthly billing costs more.

Two caveats. First, the execution is the platform bill, not the model bill. You pay for the language model separately, with your own key from an AI provider or through n8n's Gateway credits. With every tool call, the model gets a result back and has to reason again. That costs tokens. On n8n, then, the platform bill does not grow with tool calls, but the model bill does.

Second, Agents is still in preview. n8n itself advises testing before you publish and keeping approvals on for anything sensitive. A tool you mark as sensitive waits for a human to Approve or Reject. That protects your systems from a wrong write action. It does not protect your model bill, because nothing caps the number of lookups.

What does the same customer service agent cost on two platforms?

This is a worked example, not a client case. Take an online shop that lets an agent handle customer emails. For each email, the agent fetches the order, checks the shipping status, creates a return label if needed and drafts a reply in the helpdesk. That averages four tool calls per email. A simple question takes two, a tricky return with follow-up questions eight. A thousand emails come in each month.

Our own worked example, AI step only. Formula and rates from Zapier, executions according to n8n. On Zapier the model costs are included in the tasks, except with your own AI account.

Why a range for n8n? Email is not among the channels n8n lists for Agents. So an email arrives through a workflow that calls the agent. The announcement does not say whether that counts as one or two executions per email. To be safe, we assume two.

Now look at the second column. No extra emails came in. The model took two more steps per email, for instance because the questions were harder or a new model version behaves differently. On Zapier Premium, that pushes usage up by ten thousand tasks a month. On n8n, platform usage stays flat and only the model bill rises.

We deliberately left euros out of the Zapier rows. What a task costs depends on your plan and your volume, and both change. Put the figure from your own plan next to it. If you are close to your plan's ceiling, you will see straight away what two extra tool calls per email cost you.

Zapier often wins too. A Zap with three fixed steps and one summary on Advanced is perfectly affordable, and Zapier has ready-made connections to thousands of apps. The difference only becomes large when the model chooses the number of steps itself. Then Zapier bills you per step, and n8n per turn.

Where does the multiplication happen on Make and Notion?

Make and Notion multiply as well, just in a different place. On Make it happens in the bundles. An operation is one run of a module, and each bundle that passes through a module counts as a run of its own. Make walks through a scenario with a form trigger, ten bundles and three modules. That adds up to 31 operations per run: one for the trigger and ten for each other module.

A second example on the same page shows what bundling does. A scenario sends the sales manager a separate email for every new contact. With a Text aggregator, all new contacts go into one email, and usage drops from 11 to 3 operations. The same information arrives, at roughly a quarter of the usage.

On Notion, the multiplication depends on who calls the Worker. According to Notion, a Worker typically costs $0.0023 per run. A sync that runs every fifteen minutes comes to $6.62 a month in Notion's own example. When an agent uses the Worker as a tool, every tool call counts as a run. Notion's example of a support agent with five Worker runs per agent run and nine thousand agent runs a month comes to $103.50.

Source: Notion Help Center. Worker runs only: the Custom Agent's own actions consume credits on top. Workers are free on Business and Enterprise during the beta.

These are two different jobs, so comparing them like for like makes little sense. With the sync, you decide how often it runs. With the agent, the model decides how many runs each conversation adds, and the agent's own credits come on top. Each run is cheap, but it is still a multiplier you do not control.

How do you count tool calls before choosing a platform?

You do not need a platform to count tool calls. An hour with a spreadsheet and a few real examples is enough. Here is how:

1. Pick one flow. Customer emails or quote requests, something that recurs every week.
2. Take ten real cases. Not the easiest ones. Include the customer who emails back three times and the order that shipped in two parts.
3. For each case, write down what a person looks up and does. Every system they open and every action they take becomes a tool call later. Looking up an order is one tool call, creating a return label another.
4. Note the average and the outlier. Four on average and twelve in the worst case is a different story from four on average and five in the worst case.
5. Multiply by your volume. Tasks per month times tool calls per task. Run that through the billing rule of every platform you are considering.
6. Run a scenario with two extra calls. An agent sometimes double-checks something or retries after an error. You want to know that up front.

After one flow, you can see whether the platform you were looking at fits. A flow with many fixed steps and few choices suits a unit per step. If the model has to look up a lot and make its own choices, a unit per run suits it better.

Then take three questions to every demo, ours included. What is the billing unit, and what happens to it when the agent takes one more step? How many tool calls did the agent make in this demo? And where is the brake: a limit per run, approval on sensitive tools, an alert when usage drifts? The answer to the second question is in the log. n8n, for instance, says you can see inputs, tool calls, outputs and errors per session. Ask about it.

Which choices lower the bill without making the agent dumber?

Every decision that is fixed belongs in the workflow. There, a step costs little or nothing. Almost every saving below comes down to that.

Put fixed decisions in the workflow. On Zapier, Formatter, Paths, Filters, Delay, Looping, Sub-Zaps, Digests, Storage, Tables and Forms cost zero tasks. Should an email about returns always go to another team? Let a filter make that call, not the model. The forwarding itself is still a regular step of 1 task. On Premium that saves 4 tasks per run, on Advanced 2.

Choose the lightest model that can handle each step. Classifying and short summaries often work on a lighter model. On Zapier, Standard does that for 1 task, as long as no tool is needed. If a tool is involved, Advanced instead of Premium saves two fifths per run. With your own AI account you pay 1 task per step and 1 per tool call, and the model costs go directly to your provider.

Bundle where you can. Make's example of going from 11 to 3 operations shows what a single aggregator saves. The same email arrives, at roughly a quarter of the usage.

Set a limit and check who enforces it. Zapier pauses an AI step at 75 tasks by default. The n8n Agents announcement mentions no such limit per run, only approval on sensitive tools. That stops write actions but leaves the number of lookups uncapped. A limit lets you spot an outlier before it shows up on the invoice.

Measure again after a week. Your count on paper is an estimate. The log from the first week shows how many tool calls the model really makes. If it differs, redo the maths before you scale up.

What is hard to change once you want to switch?

The usual fear is getting stuck with a model. That is the lowest hurdle. Zapier lets you pick models from a list, n8n lets you use any model you have a key for. Swapping a model is quick. What it does to your bill only shows in the log: another model may need more tool calls, and on Zapier the rate changes too.

Other things are not so quick to change. First, the place where your connections and credentials live. Then the history of your runs: which customer got which answer, which order was changed. And the billing unit itself, because it is woven into how your workflows are built. A Zapier workflow designed to use few tasks looks different from an n8n workflow optimised for executions.

So ask the cost question before you build. If you start on Zapier with an agent on Premium and find after three months that usage is climbing fast, you cannot just move. Every workflow, connection and exception would have to be rebuilt.

You can switch later. But make the first choice with the right number in hand. For an online shop with a few fixed connections, the number of workflows may genuinely be the best measure. If you let an agent search, decide and write on its own, the number of tool calls is the measure. That number belongs in our proposals. Ask other builders for it too.

How do we cost out an automation at Oase Creative?

We build automations and agents for clients under technology and AI, alongside the websites and Shopify stores that automation is often tied to. We also run an agent system of our own that helps run our agency. What that costs in upkeep and attention is in a separate article, measured over 134 days. This article covers the other half: the bill.

For a new automation, we want to see the count from the steps above before we propose a platform. How many tasks per month and how many tool calls per task, with the worst case included. Only once those numbers are on the table do we decide where the logic sits. In the workflow it is cheap, in the model it is flexible. Usually it ends up a mix.

A proposal like that is less spectacular than a demo. It contains a table of numbers, not an agent helping a customer live. If that table shows that a simple Zap with two filters is enough, we say so. That is cheaper for you and less maintenance for everyone.

Not sure whether to build this yourself or have it built? How to choose an AI agency lists the questions to ask a builder. The last 20 percent of AI in an online shop covers the work that starts after the demo. If you connect several systems to an online shop, the build counts too; what a Shopify store costs is in a separate article. In the middle of choosing? Tell us where you are stuck. Bring your count of ten cases, and we will have something concrete to look at straight away.