How much does AI automation cost?

Almost all of the variance comes from integration surface, not from the AI. Automating a process that lives in one system is cheap. The same process spread across a CRM, a mailbox, a spreadsheet, and a messenger costs several times more, because most of the work is making four systems agree on what a record is.

The other three multipliers: data that has to be cleaned before anything can run on it, decisions that need a human approval step wired in, and whether you need monitoring and documentation at the end or just something that works today.

Diagnostics should be free. Any competent vendor can map your manual operations and estimate ROI on the top candidates without charging for it — that is a sales activity, not a deliverable. Paying four or five figures for a discovery phase that ends in a document is the single most common way automation budgets get spent without anything shipping.

What you should expect to pay for: the build, the integrations, and the handover. Those are real engineering.

Ask for the ROI arithmetic behind the number — hours saved per week, multiplied by loaded cost, against the build price. If a vendor cannot produce that on one page, they have not scoped the process, they have scoped a budget.

Then ask what happens after handover: who can change the automation in six months, and what it costs. A cheap build with no documentation is the expensive option on a two-year view.