The Hidden Cost of AI Agents Nobody Talks About

Your agent isn’t the only one doing the heavy lifting

Imagine hiring a digital teammate tomorrow. One that moves faster than you can type and makes light work of the jobs that slow you down.

I have been teasing AI agents for the past few weeks, and we are almost there! As you know, I like to be thorough, and there is one more topic we must cover together: the cost of running AI agents.

If you are a new reader, welcome! Scroll to the end of this article for a list of recommended reading so you can catch up.

Before we build, let’s make sure your toolkit is ready

I cannot possibly provide you with a comprehensive list of every possible way to create agents. However, I can tell you that there is one tool I use almost exclusively because it is so well-known in the Agentic AI world: n8n.

There are other great choices, such as Zapier and Make. For those of you on Microsoft 365, you have access to a similar technology called Power Automate. A new kid on the block is Postman’s AI Agent Builder.

Here is why n8n stands out for me:

• The company offers an open-source version, allowing you to host your flows locally on your computer or corporate IT infrastructure.

• They offer paid plans on their website for those who prefer not to use SaaS (Software as a Service) with predictable pricing.

• They have adopted the popular open-source approaches to AI workflows, which are very popular among software developers (such as Langchain).

Why is that last point so important? Imagine you’ve built an automation that organizes customer support requests. You realize you could package it up and sell it as a product. You can do that with n8n!

The new breed of AI agent builders

You may be aware that Anthropic’s Claude can create apps in a chat window using their Artifacts feature. Other products can make a full-blown app for you from a prompt, such as Loveable or Replit.

Even developers can create agents using code editors like GitHub Copilot, WindSurf, and Cursor.

All these products offer the promise of quickly transitioning from an idea to a product, with minimal knowledge of the underlying technology used to run them.

Unfortunately, the underpinnings of some products are so opaque that you don’t really know how they work.

When I review automation agents, I test them across different platforms before settling on the one that best fits the task. I prefer tools that let me visualize each step, since seeing the flow helps me understand exactly what is happening. That is why I often choose n8n, though I still use Zapier, Make, Power Automate, and other automation tools.

Clarifying prescriptive workflows vs AI automations

Imagine the following repetitive tasks where you and your team must send out a document to a large audience:

• Write a document.

• Send it out for review.

• Collect feedback.

• Revise and publish it.

A prescriptive workflow might:

• Send an email or Slack message to reviewers.

• Create and assign review tasks in a tool like Jira.

• Notify you when all tasks are complete so you know the document is ready.

A value-added AI automation could go further by:

• Help you draft the document.

• Use past feedback from your team and suggest edits.

• Assign reviewers and highlight topics that match their expertise.

• Rewrite the document based on input from your team and send you a final draft for approval.

• Publish the document, saving it to your repository, converting it to PDF, and even translating it.

The difference is clear. A prescriptive workflow is mostly an if-this, then-that process. An AI automation acts like a partner. It co-writes, suggests improvements, tailors tasks to the right people, and cuts down the time spent on manual edits.

Got it, but what about costs?

In general, you should expect to pay out-of-pocket expenses for the following:

1. Hosting your AI solution somewhere (whether you pay n8n or host it yourself). Call that $20/month.

2. AI API calls (read on for more on this important topic).

3. If your automation requires access to other tools, some are available for free, while others have a cost associated with them. If you want to perform basic automations with tools like Microsoft Office or Google Workspace, these will generally be free to use.

Aside from hosting costs, your most significant expense will likely be those AI API calls. Let’s understand what that means for you.

AI API tokens

AI tools use what we call tokens to define how much you pay. For example:

• 1 token ≈ 4 characters of English text

• A word averages about 1.3 tokens

Token pricing

Now that you understand the basics of what a token is, you should know there are typically two ways you will pay for them:

1. Input tokens represent your prompt (instructions) that define what you want from the AI.

2. Output tokens represent the AI’s response to your prompt.

Input tokens usually cost less. Output tokens typically cost more. That is because input tokens are your instructions, while output tokens represent the AI generating a detailed response. The added cost reflects the value of producing new and valuable content.

How to estimate token cost

---

First, a big pricing disclaimer

For this article, I outline how to estimate pricing and provide some estimates in $ USD. Your pricing can vary widely based on usage, contracts, and any number of other factors. Do not use this article as a way to estimate your AI automation costs.

Take the time to check current rates and compare options so you get the best results at a cost that meets your budget. To explore pricing models directly, visit:

• OpenAI’s cost comparison site

• Anthropic’s pricing page

• Google’s Gemini developer API pricing page

---