An AI automation has three costs. The build is paid once, the running costs every month, and upkeep keeps it working when your apps or AI models change. Most quotes show the build clearly and say less about the other two. To judge an AI automation fairly, put all three side by side and ask for each one in writing.
The three parts of the cost
| Part | What it covers | How it is billed |
|---|---|---|
| The build | Setting up the automation, testing it on your real work and showing your team how to use it | Once |
| Running costs | AI usage, plus any automation platform, hosting or business software it runs on | Monthly, partly by use |
| Upkeep | Monitoring, fixes and updates when a connected app or AI model changes | Monthly, or by the hour when something breaks |
Some providers fold running costs and upkeep into one monthly fee. Others pass the AI usage straight through and charge for fixes as they happen. Neither is wrong, but you can only compare two quotes once you know which model each one uses.
What drives the build price
Three things decide most of the build cost.
- The number of steps. Reading an email and filing it is a short job. Reading it, checking a price list, drafting a quote and booking a follow-up is a long one.
- The number of apps it connects. Each connection to an inbox, calendar, accounting tool or customer list adds set-up and testing. Apps without a ready-made connection cost more.
- How much checking the results need. Anything sent to customers or touching money needs review steps, test cases and clear rules for when a person takes over.
The best first automation is a task that happens often, follows the same steps each time and is easy to check. Sorting new enquiries, logging invoices and drafting replies to reviews fit well. A task that happens once a month rarely repays its build. Our guide to AI automations worth starting with lists good candidates.
What the AI costs to run
The AI provider bills per token, a small piece of text, for what goes into the model and what comes out. Prices are listed in US dollars per million tokens. Claude Haiku 5.5, for example, costs US$0.10 per million input tokens and US$0.50 per million output tokens for prompts up to 100,000 tokens.
Claude API output price per million tokens (US$)
Here is an illustration with round numbers. An enquiry automation handles 1,000 enquiries a month. Each run sends about 2,000 tokens of instructions and message, and gets back about 500 tokens of reply. That is 2 million input tokens and 0.5 million output tokens a month:
- Haiku 5.5: about US$0.45 a month
- Sonnet 5.5: about US$9 a month
- Opus 5.5: about US$18 a month
The model you choose changes the bill more than anything else. Sorting enquiries or reading invoices rarely needs the most capable model. Pick the smallest model that does the task well, then test it on real examples. Our guide to choosing an AI model compares the options.
Two more things lower the cost:
- Batch work: jobs that can wait, such as a weekly report, can use the Batch API at a 50% discount on input and output tokens.
- Short instructions: every run sends the instructions again, so tight instructions and only the documents the task needs keep each run cheaper.
The costs that are not the AI
Automation platforms, hosting and business software, such as an email suite or a customer list, charge their own fees. For a routine task these often cost more than the AI usage itself. Messaging channels can also charge per message. Ask for each of these by name in the quote.
Why a chat plan does not cover it
A chat subscription pays for a person using the chat app. An automation runs on its own through the API, and Claude API usage is billed separately from paid Claude plans. Other AI providers work the same way. Budget for API usage even if your team already pays for chat plans.
How to keep running costs under control
A spending limit is the simplest protection. On Anthropic’s Claude Console, you can set a monthly spending limit on the Billing page. Set it at what you expect to spend, with a little room.
You can also give each automation its own workspace. Workspace spend limits cap that workspace’s monthly spending and send alerts at amounts you choose. One busy automation then cannot use up the budget of the others.
Why upkeep is a real cost
An automation is connected to apps you do not control. When an app changes how it connects, or an AI model is replaced by a newer one, the automation needs checking and sometimes changing. Plan for it in one of two ways: a monthly fee that covers fixes, or an hourly rate written into the agreement. A quote with neither leaves you to find out the cost when something breaks.
How to compare quotes
Use these questions to put two or three quotes side by side.
1. Ask what the build covers
Check what counts as the task, how many changes are included after launch, and for how long fixes are included.
2. Ask for the running costs in writing
A good quote estimates the monthly AI usage and names every platform, hosting and software fee. Ask which AI model it uses, and why that one.
3. Ask how upkeep is charged
Find out whether fixes are included in a monthly fee or charged when they happen, and what happens when a model is retired.
4. Check whose accounts it runs in
If the AI and software accounts are in your name, you see every charge. If the provider hosts the automation, ask how you get your data back when you leave.
5. Check a person approves what matters
Ask which steps run on their own and which wait for your approval. Anything sent to customers or touching money should wait for a person until you trust it.
Our own prices for the AI systems we build and host are on our pricing page. Before you compare quotes, write down the one task that takes your team the most time each week. It is usually the best place to start.