How much does AI automation cost is four questions, not one: the one-off build, the monthly platform fee, the model cost per use, and maintenance after handover. Only the model cost has a published price list you can check to the cent. The other three depend on your process, your stack and your own hourly cost.
This is written for the person who has to defend the number internally, in a company of 5 to 200 people, before the first supplier conversation. Not a range to quote, but a bill you can split.
What does an automation actually cost?
Four line items. A one-off build, a monthly platform fee for the tool the workflow runs on, a model cost per use if the workflow calls an AI model, and maintenance once it is live. Three of the four are specific to your company. One is a public price list.
| Line item | Where the amount comes from | Who decides the size | One-off or recurring |
|---|---|---|---|
| Build | hours multiplied by the rate of whoever builds it | the scope of your process: steps, systems, exceptions | one-off |
| Platform | the vendor’s own pricing page, after the cloud or self-hosted choice [S-006] | you, by choosing where it runs | recurring, monthly |
| Model use | the published price per million tokens [S-040] | your volume and your model choice | recurring, per use |
| Maintenance | hours multiplied by the rate of whoever owns the workflow | you, at handover | recurring |
One row in that table has real numbers in this article. That is not laziness, it is the state of the evidence: Anthropic publishes its prices, and everything else in the bill is either your number or a vendor page you should read on the day you decide.
Whether your automation has a model cost at all depends on what you are building. A workflow that moves data between two systems never calls a model; a workflow that reads an email and decides what it is about does. The difference between automation and AI automation covers that line, so this article does not repeat it.
What does the model itself cost per month?
Lower than most estimates, and it is the one line item you can calculate before you speak to anyone. Anthropic publishes the cost per token as a price per million tokens for every Claude model: Haiku 4.5 at $1 input and $5 output, Sonnet 5 at $2 and $10, Opus 5 at $5 and $25 [S-040]. Those are dollars, so leave them in dollars.
| Model | Input per million tokens | Output per million tokens |
|---|---|---|
| Claude Haiku 4.5 | $1 | $5 |
| Claude Sonnet 5 | $2 | $10 |
| Claude Opus 5 | $5 | $25 |
Source: the Claude API pricing page, checked on 27 September 2026 [S-040].
Turning your own volume into tokens
A token is roughly 4 characters, or about 0.75 words in English, and the exact count varies by language and content type [S-045]. That rule of thumb is enough for a budget: a thousand words of text sits in the region of 1,300 tokens, and both what you send and what the model writes back are counted.
Anthropic works one example out on the same page. Ten thousand customer conversations, averaging about 3,700 tokens each, handled by Claude Haiku 4.5, come to about $37 in total [S-043]. Ten thousand conversations, one lunch. Per conversation that is under half a cent.
Model choice moves that number more than volume does. Both Sonnet 5 prices are exactly double the Haiku 4.5 prices, and both Opus 5 prices are five times them [S-040], so the same 10,000 conversations land near $74 and near $185. If a supplier cannot tell you which model a workflow runs on, they cannot tell you what it costs.
Two published discounts, and when they pay off
The Batch API processes large volumes of requests asynchronously at a 50 percent discount on both input and output tokens [S-041]. Anything that does not have to answer within seconds belongs there: overnight enrichment, a weekly report, a backlog of documents. Applied to the example above, the same 10,000 conversations halve to roughly $18.50 [S-041], [S-043].
Prompt caching is the second one, and it has a break-even point Anthropic states itself. A 5 minute cache write costs 1.25 times the base input price, a 1 hour write costs 2 times, and a cache hit costs 0.1 times the base input price; caching pays off after one cache read for the 5 minute duration, or after two reads for the 1 hour duration [S-042]. In practice: cache the instruction block that every run shares, and do not bother caching a prompt that is different every time.
The tool cost nobody adds up
Server side tools are billed on top of the tokens. Web search on the Claude API costs $10 per 1,000 searches, plus the standard token costs for the content the search returns; web fetch carries no additional cost [S-044].
Put that next to the worked example. If each of those 10,000 conversations runs one search, the searches come to about $100 while the model comes to about $37 [S-043], [S-044]. The tool is the larger number, and it is the one missing from every price range on every agency page. A model budget without a tool budget is not a budget.
What is the one-off build cost made of?
Hours, and the rate of whoever works them. There is no published price for this, so any figure here would be invented. What can be stated is the shape: scope decides hours, and scope is the number of steps, the number of systems and the number of exceptions the workflow has to survive without a human.
Two numbers on WeAdapt’s own side are public, so they belong in the open. The Quickscan starts at €995 and takes 1 to 2 weeks [A-006]. From discovery call to live is four weeks [A-007]. Neither is a price for your build. They bound the conversation: what the mapping costs, and how long a first workflow takes before it earns anything.
You can judge scope yourself before anyone quotes. A single notification is not the same job as a chain that reads, decides and writes back; the five levels of marketing automation puts that on a scale, and automating management reports shows what one concrete build contains.
Then do the arithmetic with your own numbers, not a market average:
- Hours the task takes now, per month. Count them for two weeks and double it. Estimates from memory are always low.
- Your fully loaded hourly cost for the person doing it: gross salary plus employer costs plus overhead, divided by productive hours. Your finance colleague has this figure. A national average is not a substitute, and this article will not supply one.
- The share of the task that can run without a human. Rarely all of it, and the exceptions are where the hours hide.
Hours times rate times share is the monthly saving. The build cost divided by that saving is the number of months before it is back. Both inputs are yours, which is what makes the answer defensible when someone in the room asks where the number came from.
Why is there no monthly platform cost in this article?
Because it could not be sourced, and that is a deliberate omission rather than a gap. The platform line starts with a choice, not a price. n8n runs as n8n Cloud, with no installation needed, or self-hosted, where you manage your own deployment [S-006]. Cloud puts the cost in a subscription. Self-hosted puts it in a server plus the person who keeps it running, which is a cost with no invoice attached.
The vendor pricing pages were not reachable when this article was researched on 27 September 2026, so there is no primary source to quote for the monthly platform cost. Repeating a figure from another blog would make this page look more complete and be worth less. Read the vendor’s pricing page on the day you decide, and ask which plan the supplier assumed.
What are the hidden costs of automation?
They are not hidden. They are unbilled, which is worse, because nobody owns them until something stops. Six, in the order they usually surface.
- Maintenance after handover. Someone has to notice a failure, make small changes and keep connections current. That is a couple of hours a month per small set of workflows, and it is real work. Who owns an automation after launch covers how to assign it before it becomes a problem.
- The error path. An automation that fails silently produces work that surfaces later as missing data somebody rebuilds by hand. Building the alert route costs hours in the project and saves days afterwards.
- Access and credentials. Connections expire, people change roles, and a workflow that ran for eight months stops on a password nobody owns.
- Model cost that scales with success. The cost of workflow automation on the model line is per use. Triple the volume and you triple that row, which is the good problem, provided the budget expects it.
- Cleanup before the build. If two systems disagree about what a customer is called, that gets decided during the project, by your people, in hours nobody planned.
- Switching it off. Every workflow that no longer matches the process still runs, still costs and still produces output someone half trusts. Deciding to retire one is a task, not an event.
Why do the published prices differ so much?
Because each page is adding up a different set of line items and none of them says which. One quotes the platform subscription. The next adds the build. A third quotes a retainer with maintenance inside it. Put three of them side by side and the ranges do not even overlap, which is the tell: they are not measuring the same thing, and none of them names a source or an hourly rate.
So do not go looking for the right range. Ask any supplier four questions instead: which of the four rows is inside your number, in which currency, what remains per month once the build is paid off, and whose hourly rate you assumed. A supplier who cannot split the bill has not built enough of them to know.
What to do with this before you call anyone
Three steps, in order. Calculate the model line from the published price list and your own volume, because that one needs nobody’s help. Write down your hours, your fully loaded hourly rate and the share that can run unattended. Then open the vendor pricing page for the platform you would actually use and add the plan you would actually need.
What is left is the build and the maintenance, and those are the two you negotiate. WeAdapt builds the system, your team runs it, and the Quickscan from €995 maps what there is to build and what it will save before anyone commits to a scope [A-006]. Request an audit to get the four rows filled in for your own processes, or read what a project costs and whether there is a monthly subscription in the FAQ.