"How much will it cost?" is the first question we hear on almost every AI project, and the honest answer is "it depends." That's not a dodge. AI work varies enormously: a simple email classifier and a custom assistant connected to five internal systems are different universes.
What we can do is show you where the money usually goes, because it's rarely where people expect.
The model is often the cheap part
Many people assume the main cost is the AI itself. Using an existing model through an API is usually inexpensive to start with, since you pay per use. For a pilot with modest traffic, the fees can be small.
The bigger costs are everything around it.
Where the real cost sits
Data preparation. This is the one that surprises people most. Your documents may be scattered across folders, out of date, duplicated or stuck in scanned PDFs. Cleaning and organising them so an AI can use them takes real time, and it's often the largest part of the work.
Integration. An assistant that can't reach your systems is just a chat window. Connecting it to your CRM, helpdesk, database or internal tools means building and testing those links, plus handling permissions properly.
Design and development. The screens people use, the rules about what the AI may do, the fallback when it's unsure, the review steps. This is ordinary software work, and it costs what software work costs.
Testing. AI doesn't behave the same way every time, so testing means checking hundreds of real examples, not just clicking through a few screens. Skimping here is how teams end up with an embarrassing public mistake.
Usage fees. Per-request pricing looks tiny until you process a million requests. Estimate your volume before you decide, and check how the price changes if people use it more than you expected.
Maintenance. Models get updated, your documents change, new edge cases show up, and someone must keep watching quality. Plan for ongoing time every month, not a one-off build.
People and change. Training staff, writing usage rules and handling the "will this replace me?" worry takes effort. Projects that ignore this often get built and then not used.
Why estimates swing so much
Two things move the number most: how messy your data is, and how many systems the AI must touch. A tidy knowledge base and a single integration can be a small project. A decade of inconsistent records across five tools is a big one, whatever the AI itself costs.
Be wary of anyone who quotes a firm price before looking at your data and your processes.
How to keep the risk low
Start with a pilot. Pick one task, one team and one number to improve, such as minutes saved per ticket. Run it for a few weeks on a small scale. A pilot costs a fraction of a full build, and it answers the questions that actually matter: does it work on our data, do people use it, and is the saving real?
Set a stopping rule before you start. If the pilot doesn't hit its target by a certain date, you stop. That's a successful outcome too, since you spent a little and learned a lot.
Only after that do you build the proper version.
Questions to ask any vendor
- What will you need from us (data, access, people's time)?
- What's included in the price, and what isn't?
- What are the ongoing monthly costs once it's live?
- How do you measure whether it's working?
- What happens to our data, and who owns the result?
Vague answers to these tell you something.
The bottom line
Budget for the data and the integration first, the AI second, and maintenance forever. Start small, measure honestly, and grow the spend only when the numbers justify it. A cheap pilot that teaches you the truth is worth much more than an expensive launch built on hope.
