How much does it cost to run an AI automation? Estimate the running costs before you build

What you actually pay for

When a business owner asks what an AI automation costs, they usually mean the build. The build is only part of it. Once a workflow is live, you keep paying for it every month, and those costs are easy to miss in a demo.

It helps to split the running cost into four parts. The first is model usage: what the AI provider charges each time the workflow sends it work. The second is tool charges, such as web searches, on top of that usage. The third is the software around the model: the form, inbox, CRM or automation platform it connects to, each with its own subscription. The fourth is people: the staff time spent checking what the system prepared, fixing exceptions and keeping the workflow working as your tools change.

Model usage is the part that's hardest to guess, because it isn't a flat fee. The rest of this guide shows how it's priced, how to estimate it from your own numbers and how to stop it growing without anyone noticing. Keep the other three parts in the same estimate, because a workflow that's cheap to run but slow to check isn't saving you much.

Keep readingPlan who operates the workflow after launch

How AI usage is priced

AI providers charge by the token. A token is a small piece of text. Anthropic says that for its Claude models "a token approximately represents 3.5 English characters", though the exact number varies. Everything the workflow sends counts as input: your instructions, any examples, the customer's message and any documents attached. Everything the model writes back counts as output.

Prices are quoted per million tokens, with input and output priced separately. At both providers we checked, output costs more than input. When we checked on 5 October 2026, Anthropic listed Claude Haiku 4.5 at US$1 per million input tokens and US$5 per million output tokens, and Claude Opus 5.5 at US$4 and US$20. Anthropic states that all of its prices are in US dollars, so your Australian dollar cost also moves with the exchange rate. OpenAI's pricing page is set out the same way: prices per million tokens, with separate columns for input, cached input and output.

Prices change often, and per-token prices aren't directly comparable between models. Anthropic notes that its newer models use a tokenizer that "produces approximately 30% more tokens for the same text". So the same email can cost a different number of tokens on two models, before you apply either price. Estimate with the model you'll actually use, and recheck the provider's pricing page before you rely on a figure.

Estimate one workflow on paper

You don't need to build anything to get a useful estimate. Pick one workflow and take the numbers from your real work, not from a demo. If you haven't chosen a workflow yet, start with a task that's frequent, has a clear owner and has a clear start and finish.

Count the requests the workflow will make each month, not just the jobs. One job can involve several model calls if the workflow drafts, checks and then summarises. Estimate tokens from real examples: run a few typical emails or forms through the provider's token counter. Anthropic says its token counting endpoint lets you count a message's tokens before you send it, and that the count is an estimate. Then work through these steps:

• Monthly volume: how many items arrive in a normal month and in your busiest month.
• Model calls per item: every draft, check, summary and retry counts.
• Input tokens per call: instructions, examples and the item itself, measured on real samples.
• Output tokens per call: the length of a typical prepared record, reply or summary.
• Price: input tokens times the input rate, plus output tokens times the output rate, for the model you'll use.
• Extras: tool charges, software subscriptions, staff checking time and ongoing support.

Keep readingChoose the first workflow to cost

A worked example

This example is hypothetical. The figures are invented to show the method, not taken from a Demandify client or a measured workflow. Imagine a business that gets 400 job requests a month through its website and inbox. An AI step reads each request, pulls out the service, location and timing, flags anything missing and prepares a record for a staff member to approve.

Say each request sends 3,000 tokens of instructions and examples plus 1,000 tokens for the customer's message, so 4,000 input tokens. The prepared record is about 500 output tokens. Over a month that's 1.6 million input tokens and 200,000 output tokens. At an example rate of US$1 per million input tokens and US$5 per million output tokens, the month costs US$1.60 plus US$1.00: US$2.60. At example rates of US$4 and US$20, the same month costs US$10.40. If the workflow adds a second checking call for every request, roughly double those figures.

Now add the people. If a staff member spends two minutes checking each prepared record, that's 800 minutes, more than 13 hours a month. In this example the checking time costs far more than the model usage, so it's the number to work on: clearer instructions and better flags that cut checking time do more than switching to a cheaper model. The original manual process belongs in the comparison too. If staff spent eight minutes on each request before, the saving comes from the difference, after checking and exceptions are counted.

Costs that grow quietly

The estimate above covers a simple workflow. Costs grow when the workflow does more for each request, often in ways nobody sees on the first invoice.

Tools add tokens. Anthropic says the tool names, descriptions and schemas sent with a request count as input tokens, and that using tools also adds a special system prompt of several hundred tokens on many models. Some tools charge on top of that: Anthropic lists web search at US$10 per 1,000 searches plus the tokens the results add. It estimates an average web page at about 2,500 tokens once fetched, and a research paper PDF at about 125,000. A workflow that browses or reads long documents can use far more tokens per job than one that reads a short email.

Conversations add up as well. Earlier messages, documents and results are usually sent again with each new turn, so a long back-and-forth costs more per reply than a single request. Reruns do the same thing: a step that runs again because its first answer was rejected is paid for every time it runs.

Where your data is processed can change the price. Anthropic lists a 1.1x multiplier when you specify US-only processing for its newer models on its own platform, and a 10% premium for regional endpoints on Amazon Bedrock and Google Cloud. OpenAI lists a 10% uplift for regional processing on models released from 5 March 2026. If you need processing in a particular region, price it from the start.

Ways to keep usage costs down

Match the model to the task. A smaller model may handle routine field extraction well, while a harder judgement may need a stronger one. Test both on the same real examples and compare the checking time as well as the price. A cheaper model that produces more corrections can cost more overall.

Use batch processing for work that can wait. Anthropic's Batch API gives "a 50% discount on both input and output tokens", with "most batches completing within 1 hour", but a batch can take up to 24 hours. OpenAI also describes "50% lower costs" for its Batch API, with "a clear 24-hour turnaround time". That suits an overnight summary of the day's job notes. It doesn't suit a reply to a customer who's waiting.

Prompt caching can help, but check whether your workflow qualifies. Caching lets repeated content, such as long fixed instructions, be read back at a fraction of the normal input price. Anthropic's cache has "a 5-minute lifetime" by default, refreshed each time it's used, with a dearer one-hour option. It also has a minimum length: 4,096 tokens on Claude Haiku 4.5, for example, and shorter prompts are processed without caching and without an error. In the worked example, the 3,000 tokens of instructions are too short to cache on that model, and requests that arrive hours apart would miss a five-minute cache anyway. Caching suits steady, high-volume traffic better than an occasional enquiry.

Finally, send less. Trim instructions to what the task needs, avoid attaching whole documents when a section will do, and limit how much fetched web content goes into each request.

Set limits and watch the bill

Put a ceiling on spending before the workflow goes live. Anthropic says spend limits "set a maximum monthly cost an organization can incur for API usage", and that you can set your own limit below your account's monthly cap. When usage reaches that limit, requests return an error until the limit is raised or access resumes. That protects the budget, but it also means the workflow stops. Decide in advance what happens then: work should fall back to a person's queue, not disappear.

Run the API account under your business, with named people who can see usage and change limits. Review usage each month against the estimate, alongside the measures you set before launch: volume, checking time, exceptions and errors. If usage is well above the estimate, find out why before raising the limit. The cause is often a retry loop, a longer prompt than planned or more calls per job than anyone expected.

If you'd like help estimating and testing a first workflow, Demandify's AI Systems & Automation service can scope it with you, including the build, the ongoing operation and the third-party costs. Tell us about your business and the task you have in mind.

Keep readingApprove the AI tools and accounts your team uses