What the market actually pays
Start with real invoices, because they show AI development cost as buyers actually paid it. Clutch’s pricing guide for AI development, built from verified client reviews, puts the average project at $120,594. The most common budget is $10,000–49,999, a typical project runs ten months, and the monthly spend averages $11,553.
Read the average with care. The most common rate on Clutch is $25–49 an hour, so in our reading the average leans toward small and mid-sized projects at those rates. We would not use it to budget a large enterprise build in the US or the EU.
Rates explain much of the spread. On the same Clutch page, AI developers in the US charge $50–99 an hour, while teams in India, the Philippines, and Ukraine charge $25–49. An in-house hire in the US costs more again: the US Bureau of Labor Statistics reports a median of $140,300 a year, or $67.45 an hour, as of May 2025, for the occupation closest to AI research. That figure is pay alone. How that compares with hiring a partner is covered in our cost comparison of an in-house AI team and an external one.
Our own numbers, for the record: projects start at $25,000, and our rate is $25–49 an hour. That places us in the most common band on Clutch.
Why the model is the cheap part
The model is rented by the token, and the token is cheap. Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens. OpenAI lists gpt-6-luna at $0.10 and $0.50. At Haiku prices, a thousand short answers that each read a page of your knowledge base cost about three dollars in model fees.
The model is rarely what separates two quotes.
So when two quotes for “an AI assistant” differ widely, the model is rarely the reason. The money goes into everything around it:
What moves the price
-
Data, and how messy it is
-
Integrations, one per system
-
Scope, one process or many
-
Model setup, API or fine-tuning
-
Support after launch
- Data. Where your records live, who can grant access, and how messy they are. If requests sit in an inbox, a web form, and a manager’s personal messenger, part of the budget goes to merging them before any model sees them.
- Integrations. Every system the AI reads from or writes to is separate work: CRM, email, task tracker, calendar, ERP, a partner’s API.
- Scope. One narrow process costs far less than an end-to-end flow across several systems that also checks its own output.
- Model setup. An off-the-shelf model through an API is the inexpensive option. Fine-tuning for a narrow domain costs more, and so does private hosting when the data cannot leave your perimeter. Not every project needs either.
- Support. The AI lives inside processes that change. Someone has to adjust it when they do.
Four kinds of AI build
The same word, “AI,” covers projects that differ by an order of magnitude. Here is how we sort them on a first call.
| Kind of build | What it does | What moves the cost |
|---|---|---|
| Scripted or FAQ bot | Answers common questions with buttons or generated replies, touches no systems | Number of scenarios |
| Generative assistant | Answers in free text from your knowledge base | State of the knowledge base, one or two lookups |
| Single-process agent | Takes a request, searches, drafts a result, updates records | Number of systems, rules for exceptions |
| Multi-system platform | Classifies a stream of documents or requests and extracts the data | Volume, number of systems, fault tolerance |
When a task fits under our $25,000 floor, we usually point people elsewhere. A packaged tool or a plain automation often does the job, and we say so on the call.
Two halves: build once, run every month
Every AI budget has two parts, and in our experience buyers often compare only the first.
The build is paid once. It covers connecting systems, writing the logic, setting up the model, and testing on real cases.
Running costs arrive every month. They include model calls, which add up with volume, plus the support that keeps the system in step when the process shifts. A team that compares quotes on build price alone can get surprised by the monthly bill once volume grows. Working out whether those two halves pay back is its own exercise, and it should happen before signing.
Our two products at opposite ends of the scale
Wizard is our agent for inbound requests. We built it for ourselves first: requests arrived faster than managers could answer, and nights and weekends got no reply at all. Wizard takes a request around the clock, searches the knowledge base, drafts a reply for the manager, and opens a card in the task tracker. It is a single-process agent, the middle of the table above.
DocuFlow sits at the other end. It reads incoming documents, classifies them, and extracts the data, about two hundred a day. Several systems, a steady stream, and fault tolerance put it in a different budget, although both products get called “AI agents.”
How a vague request becomes a number
A company asks for “an AI that handles our incoming requests.” It sounds like one task. It usually turns out to be three.
First question: where do the requests arrive? Often in three places at once, so some of the budget goes to pulling them into one stream.
Second: what should happen to each request? Finding the client in the CRM, checking the history, and drafting a reply for a manager makes it a single-process agent with several systems.
Third: who keeps it running? A process that changes monthly needs more support than a stable one.
Three short answers turn the vague question into a range we can defend. The stages after that call are laid out in how AI agent development works and what we need from you.
What a good quote should show
A useful quote splits AI development cost into the parts you can check. If a vendor sends one figure for the whole project, ask for the breakdown before comparing anything.
Look for four things in it:
- The build and the monthly run, priced separately. Model calls should appear as an estimate tied to your volume, so you can see how the bill grows when usage does.
- What the vendor has already seen. A price set before anyone opened your data is a price on risk, and it carries a cushion you pay for whether or not it gets used.
- A narrow first piece with a fixed price. One process with a visible finish line can be scoped honestly. The wider rollout gets estimated once there is something real to measure.
- What happens after launch. Who adjusts the logic when your rules change, how fast, and on what terms.
Fixed prices for a whole AI project look safe from the buyer’s side. In practice they lock the scope before the data has been opened, and every discovery after signing becomes a negotiation. What works better for us is fixing the price on the first narrow step and keeping the rest open.
Where AI software development cost leaks
Three habits we see on review calls push the cost up, and each has a simple fix.
- Starting too big. If the task fits a script, an agent adds cost without adding value. A smaller tool does the same work and breaks less often.
- Skipping the data work. Until the data is gathered and accessible, any AI hits the same mess your staff does. Sometimes the right first spend is on the data, with the model second.
- Automating everything at once. One process with a clear success measure is a safer bet than a broad rollout that quietly stalls. How to pick that first process and sequence the rest is covered in our guide on how to implement AI in business.
The fourth factor shows up later: who builds it. What to ask a vendor before you sign is in how to choose an AI development partner.
Two more things push the price up in our projects. A deadline of “yesterday” means a team working at a forced pace, and that costs more. Dirty data turns a clear project into an open-ended one, so ask yourself honestly whether the data is ready before you set a budget.
Where to start
The price of AI is the price of the work around the model. Describe one process, list the systems it touches, and check the state of the data, and the cost question gets a concrete answer.
If you want that answer for your case, book a free review of one process. It takes 15–30 minutes, and you leave knowing whether you need an agent, a simple bot, or a fix that involves no AI at all.
