In plain English
- What happened
- AI agent tools sold by vendors such as OpenAI, Anthropic, and Zapier are billed using usage-based pricing shown on their own pricing pages, not a single flat subscription, which is what makes a round savings figure hard to verify. Zapier's agent pricing is based on the number of tasks completed each month rather than a flat fee, and Anthropic's model pricing works the same way.
- Why it matters to your business
- A savings figure on its own does not tell a reader whether their situation is anything like the one being described, since hourly rates, task volume, and what was already being spent all change the outcome. The vendors' own usage-based pricing pages are the only reliable starting point for estimating what a similar setup would cost to run.
- Do this
- Before comparing any savings claim to a real decision, write down the current hours and cost of the task being considered, then check the relevant vendor's usage-based pricing page to project what running an agent would actually cost at that volume.
Maybe, but the number on its own does not tell a reader much. What matters is whether the task being replaced was expensive enough to produce six figures of savings in the first place, and whether what a business ends up paying to run an agent is smaller than what it was paying before.
How agent costs actually work
Most AI agents sold are not a single flat-fee product. They run on top of a model that is billed by usage — tokens processed, tasks completed, or API calls made — as shown on OpenAI's current API pricing page and Anthropic's current pricing page. A subscription like ChatGPT Plus covers chat use for $20 a month, but multi-step agent work that takes actions on a business's behalf is usually billed separately, through the API, by how much work it actually does.
Automation platforms built on top of those models, such as Zapier's agents product, price their plans by the number of tasks an agent completes each month rather than a flat number. That means the total bill rises as the agent does more work, which is the opposite of how a one-time software purchase behaves.
What a six-figure claim would actually require
A savings claim in six figures implies one of two things: either the task being automated was already costing tens of thousands of dollars a month in labor or software, or the figure spans a long enough period that smaller weekly savings added up. Neither of those shows up in a short claim by itself, and the vendors' pricing and usage policy pages describe rates and rules, not an individual customer's hours saved or dollar figure.
That is not unusual — software vendors rarely publish case studies with the underlying math. It does mean a reader cannot take a round number and apply it to their own business without first working out their own baseline cost for the task in question.
Reality check
A single success story, however real for the business involved, is not a benchmark. Hourly rates, task volume, and how much the owner was already paying for the tools being replaced all change the math, and none of that is visible in a number by itself.
- 1
Write down what the task costs today
Add up the hours spent on it each week, the hourly rate of whoever does it, and any software already being paid for to handle it. This is the baseline an agent would need to beat.
- 2
Read the vendor's usage-based pricing page before signing up
Check pages like OpenAI's API pricing or Zapier's agent pricing to see whether the agent is billed per token, per task, or per month, since the bill scales with how much work it does.
- 3
Run the agent on one task for a set period
Thirty days is enough to compare the actual bill against the baseline cost from step one, before deciding whether to expand it to other tasks.
- 4
Count the time still spent checking its work
Most agents still need a person to review what they produced before it reaches a customer or a set of books. That review time is part of the real cost, not separate from it.
- 5
Treat a savings number without a breakdown as an anecdote
Ask what hours were saved, at what rate, and what the vendor bill came to before deciding a result from one business applies to another.
I currently spend about [X hours per week] on [task], which costs roughly [$Y per month] in labor or software. I'm considering an AI agent billed at [vendor's rate, e.g. $ per task or per token]. Help me calculate how many hours or tasks per month the agent would need to handle before it costs less than what I'm already spending, and list anything that would still need my own time to check.
OpenAI's and Zapier's current pricing pages publish rates and rules, not a dollar figure tied to any one customer's result.
The honest answer is that an agent can save real money, sometimes a lot of it. The only way to know if it would for a particular business is to do the arithmetic with that business's own numbers rather than someone else's headline.
Sources
