"How much does AI automation cost?" is the first question every business owner asks, and most agencies answer it with "it depends" and a discovery call. Here is the honest version: the cost has three parts, and you can estimate each one before you talk to anyone.
// Part 1: The Build
The build is the one-off engineering: mapping your workflow, connecting the AI to your systems (inbox, CRM, document store, accounting), writing the logic that decides what the AI may do on its own and what a human must approve, and testing it against real data.
A single well-scoped workflow, such as lead qualification that reads enquiries and drafts responses, or invoice data extraction into your accounting system, is typically a low-thousands project. An orchestrated system where several agents hand work to each other across departments is a bigger engagement. The multiplier is integration count, not intelligence: every extra system the automation touches adds scoping, permissions and failure handling.
// Part 2: The Running Cost
Automations run on API calls, billed per token (roughly, per chunk of text processed). This is the part most quotes hide and the part you should demand transparency on. The good news: for most business workflows it is small. Reading a thousand enquiries a month through a mid-tier model costs less than a coffee run. The trap is unmonitored volume: an automation that reprocesses entire document histories on every run can burn hundreds of dollars doing nothing useful.
We covered the mechanics in the API token economy: the short version is that model choice matters more than model quality. Bulk stages run on cheap fast models; judgement stages run on frontier models; nothing runs on a bigger model than the task needs.
// Part 3: The Human Layer
Every automation worth deploying keeps a human on the approvals that matter: sending an email, changing a record, paying an invoice. Budget for a short bedding-in period where a person reviews the AI's work daily, tightening the rules until trust is earned. Skipping this is how businesses end up with horror stories; including it is why the successful ones quietly save twenty hours a week.
// What the Ranges Look Like
As a rule of thumb: one workflow, one system, human approval on every action sits at the entry level. Multi-system automation with selective autonomy is a mid-range project. A department-scale agent architecture with custom routing and on-premise components is enterprise territory, and at that point the conversation is about return, not price: what does the hour of admin time cost you now, multiplied by every hour the automation gives back?
See our automation services, read what an AI automation agency actually does, or start with a free visibility audit.