What AI Automation Actually Costs — And How to Think About ROI

Most AI agency content is written for people who have already decided to buy. This is written for people who are still deciding — who want to know what this actually costs, what they actually get, and whether the numbers make sense for their business.

The two cost categories

AI automation projects have two types of cost: build cost and running cost.

Build cost is what you pay to design and implement the system. For a focused automation project — one process, well-scoped — this typically ranges from $3,000 to $15,000 depending on complexity. A simple single-process automation (like a document routing system) is toward the lower end. A multi-channel pipeline with AI-generated content and a custom dashboard is toward the higher end. The 4–8 week timeline maps to this range.

Running cost is the ongoing cost of operating the system after it's built. This includes API usage costs (OpenAI, Anthropic), infrastructure (a VPS or cloud hosting), and any SaaS tools used in the pipeline. For most small business automations, this is $30–150 per month. It's not zero, but it's rarely significant relative to the value being generated.

How to think about ROI

The cleanest way to calculate ROI on an automation is to start with the cost of the manual process it replaces.

If a process currently takes 3 hours per day, five days a week, and the person doing it costs $25/hour fully loaded — that's $375 per week, or roughly $19,500 per year. A system that eliminates that process has a clear economic case even if the build cost is $10,000.

This is the honest version of ROI. Not "potential revenue unlocked" or "value of insights gained" — actual staff time currently being spent on something that a well-built system can handle.

The cases where the math doesn't work

Not every automation has a clean ROI case. Here are the situations where we tell clients to reconsider:

The process is too infrequent. If something happens five times a month, the automation pays back slowly even if it works perfectly. The economics get better as volume increases.

The process requires judgment that AI can't reliably provide. Some decisions genuinely require human context, relationship knowledge, or ethical judgment that LLMs don't have. Automating these creates risk, not value.

The process is about to change. If you're reorganising how you do client intake, or switching CRMs, or restructuring your team — automate after the dust settles, not before. A system built to a process that changes in six months is a system that needs to be rebuilt.

Realistic payback timelines

For most of the projects we've delivered — based on the actual time saved and the actual build cost — payback is between 3 and 9 months. The legal intake system we built took 4 weeks and cost a mid-range amount; the firm recovered that in saved staff time in about four months. The cart recovery system had a faster payback because it was recovering direct revenue.

Three to nine months is a reasonable expectation. Anyone telling you weeks is either selling you something or working on a problem with unusually clear economics.

What to ask before you commission anything

Three questions worth answering before you talk to any vendor, including us:

1. How many hours per week does this process currently take, and what does that cost?
2. What happens if the automation makes a mistake — what's the failure mode?
3. Will we be able to maintain this ourselves, or does it require ongoing vendor involvement?

If the vendor can't answer question three clearly, that's a problem. Every system we build comes with documentation, a handoff session, and a 30-day support window. After that, the client owns it and can operate it independently. That should be the standard.