What does an AI agent project actually cost?
Four things move the number. The model licence is almost never the biggest one.
Answered in short
5 things that decide this
- 01Data preparation is routinely 40% to 60% of an AI agent timeline, which makes it the largest single cost driver on most builds.
- 02Every system the agent touches adds cost, because each integration needs its own error handling, its own permissions and its own test.
- 03A build with no evaluation suite is cheaper to ship and far more expensive to keep, since every prompt change becomes a manual regression hunt.
- 04The model licence is usually a small share of the bill next to the engineering around it.
- 05Anyone quoting a firm price before reading your data is guessing, and the guess moves once they see the data.
The price list everybody wants does not exist
Ask ten agencies what an AI agent costs and you get ten ladders of numbers. Those ladders are marketing. They describe a project nobody has scoped, using data nobody has opened.
We publish no price ranges. That is a deliberate choice, not a gap. A range honest enough to cover the real spread is too wide to be useful. A range tight enough to be useful is one we would have to break later.
Here is what actually happens. A team quotes against a demo, then meets the data. Three spellings of the same customer name. A field the ops team repurposed two years ago. Exports nobody can regenerate. The build did not get more expensive. It was always this expensive, and the quote was wrong.
- Ask any vendor which of these four they priced. Most priced the first and hoped for the rest.
The four drivers that set your number
Data preparation dominates. Industry practitioners have put data work at 40% to 60% of an AI project timeline for years, and nothing about agents changed that. Your agent is only as good as what it can read.
Integration count multiplies. One agent reading one database is a small build. Point that same agent at your CRM, your ticketing tool and your billing system and you have four builds. Each connection needs permissions, retries, and a decision about what happens when it fails at 3am.
Evals decide the running cost. An agent without a test suite cannot be changed safely, so every model update becomes a week of manual checking. Skipping evals looks like a saving right up until the first model retirement.
Ownership after launch is the line most quotes omit entirely. Someone maintains this. If that is not written down before you sign, it becomes an argument later.
- Scoping calls are free. Where we have to go into an existing codebase, we run a paid two-week diagnostic first, so the number we give you is one we can hold.
- Read the dataProfiling what you have. Usually the biggest block.
- Clean and mapFixing what the demo never saw.
- Build the agentThe part everyone pictures. Rarely the largest.
- Wire the systemsCost multiplies per integration.
- Evals and gatesCheap now. Very expensive to add later.
- Run itSomeone owns the pager.
The block most quotes price is the third one. The first two usually cost more.
Related questions
01Why will you not give me a ballpark?
A ballpark given before we read your data would be a number we invent, and inventing it helps neither of us. We give you a fixed price after a diagnostic, so the certainty arrives when it is real rather than when it is convenient. Scoping conversations cost nothing.
02Does the model licence matter at all?
Model spend matters most for high-volume agents that run constantly, and it is usually small next to the engineering around it. A model that answers a few hundred internal queries a day costs little to run. The work of connecting it safely to your systems is what you are buying.
03Can we cut cost by skipping the data work?
You can defer data work, and it comes back with interest. An agent reading messy records produces confident wrong answers, which costs more to unpick than the cleanup would have cost. Narrowing scope to one clean data source is the honest way to spend less.
04What makes a project come in cheaper than expected?
Clean data in one system, a narrow task with a clear right answer, and one owner who can make decisions. Little Tree Confections is close to that shape: transcripts arrive in a consistent form, and the automation routes them into tools the team already ran.
