All insightsAI Consulting

What determines what an AI project costs?

Why two AI projects that sound the same can cost very different amounts. The five things that drive the price, and what to clarify before any quote.

· 6 min read · Stefan Parge

“What does an AI agent like that cost?” is one of the first questions in a first call. The honest answer: I can’t tell you after 30 minutes. Not because I’m dodging, but because two projects that sound the same can be very different amounts of work.

“An agent that processes invoices” can be running in a short time at a company with five regular suppliers and a clean accounting system. At another company with 400 suppliers, three currencies and an old industry system with no API, it’s a different project. The AI is the same in both.

The five things that drive the price

1. The exceptions. The normal case is quick to build. The effort goes into the cases your team groans about: collective invoices, customers who write in two languages, enquiries that are really three enquiries. The more of them there are, and the less clear it is how to handle them, the more it costs.

2. The data. Are there clean examples that show how your team has decided so far? Or is the knowledge in people’s heads and scattered emails? In the second case it has to be written down first. That’s work, and it usually pays off even without AI.

3. The integration. A modern system with a well-documented API is quick to connect. An old database, a program with no API, or a portal where you have to click through forms can make up most of the project. I always clarify this early, because it’s the one most often underestimated.

4. The testing. How good does the agent have to be before it goes live? A draft that a person checks anyway needs less safeguarding than a booking that goes straight into the system. The higher the bar, the more real cases have to be run and compared beforehand.

5. Running it. Who takes care of it when an API changes, a new model comes out or a rule needs adjusting? Either I run the system, or your team takes it over with documentation. Both work, but it belongs in the calculation.

What barely drives the price

Surprisingly often, the AI itself. For most business processes the model cost per case is small compared with the time a person spends on the same case. And the question “Claude or ChatGPT?” rarely changes the project price. The cost comes from the work around the model, not the model.

Why I bill for the result, not by the hour

With hourly billing you carry the risk. If the integration takes twice as long as expected, you pay twice. And you don’t know beforehand what the final invoice will say.

So after the first call you get a written plan from me: what gets built, what it costs, how long it takes and where the risks are. The price covers the described result. If something bigger comes up along the way, it’s a separate job with its own price, not a surprise on the final invoice.

What to clarify before any quote

Whoever you talk to, these questions save you money:

  • Which process exactly, and how often does it happen?
  • Which exceptions are there, and who decides them today?
  • Which systems does the agent have to talk to, and do they have an API?
  • Where does a person have to approve?
  • Who runs the system after launch?
  • What are the running costs per month?

Anyone who gives you a price without these answers is either estimating very generously or will recalculate later.

If you want to know where your project sits on this scale, the first call is free. More about how I work: AI consulting.

Short answers

How much does an AI project cost?

That depends less on the AI than on five things: how many exceptions the process has, how good the data is, how hard it is to connect to your systems, how much testing is needed and who runs it afterwards. A serious price only exists once those points are clear.

Why don't you bill AI projects by the hour?

Because you want to pay for a result, not for time. With hourly billing the customer carries the risk that it takes longer. With a fixed price for a clearly described result, I carry it.

What are the running costs of an AI agent?

The language model per processed case, hosting, and maintenance, for example when an API or a rule changes. You should know these costs before you start, not when the first invoice arrives.

Read something that raised more questions?

Good. That's usually where the interesting work starts. Book a free 30-minute call and we'll talk it through, without a sales pitch.

Book a call