Almost nobody publishes real numbers on this, which is why the question keeps getting asked. Here is a straight answer, with the caveat that costs in this market move quickly.
There are four costs, and most people only budget for the first.
1. The software itself — usually the smallest number
Model usage is metered and, for most small-business workloads, cheap. A system handling a few hundred interactions a month typically costs tens of dollars, not thousands. Off-the-shelf tools with agent features generally run somewhere between a modest monthly subscription and a few hundred dollars a month depending on volume.
If a vendor's licence fee is the biggest line in your budget, ask what the other three are.
2. The build — where the real money goes
This is the cost that varies most and gets underestimated most.
A narrow, single-purpose agent using an off-the-shelf platform, connected to tools you already run, is measured in days. Something that touches your own data, needs custom integration, or has to behave reliably in a regulated context is measured in weeks.
The variable is almost never model capability. It is integration — how cleanly the systems you already use expose their data, and how much cleanup that data needs first.
3. Verification — the cost nobody budgets
Every deployment needs a way to check the output. Not once at launch, continuously.
That is a person's time, and it is the line most businesses discover after go-live rather than before. We would budget it explicitly at ten to twenty per cent of the ongoing cost of running the system. Businesses that skip it do not save the money — they defer it until the first undetected error costs more than the checking would have.
4. The failed first attempt
Assume one. Ours failed more than once and we build these systems.
The realistic budget for a first deployment includes an abandoned version, because the first attempt is how you discover what the requirement actually was. Treating that as waste rather than as the cost of learning is why so many programmes stop after one try.
A realistic shape for a first project
For a lower-middle-market business automating one well-defined process, expect the software to be the cheapest component, the build to dominate the initial cost, integration to be the thing that overruns, and verification to be the permanent line you did not plan for.
The number that actually determines payback is none of these. It is how much time the process currently consumes, multiplied by how reliably the system removes it. A task consuming ten hours a week at a fully-loaded cost is a very different case from one consuming ten hours a month, and the build cost is roughly the same either way.
The question to ask a vendor
Not "how much does it cost." Ask: what happens in month six? Who checks it, who fixes it when the underlying system changes, and what does that cost. A vendor without a clear answer has quoted you a build, not a system.
More on evaluating vendors in How to Tell If an AI Vendor Knows What They're Doing.