How much does it cost to build an AI agent?

At Reflekt Lab, scoping one workflow costs $2,300 to $4,700, a full AI audit $7,000 to $14,000, and an audit plus a first agent in production starts at $17,500. Running it afterwards starts around $470 per month. Model usage, the API calls themselves, is a separate line and is usually small next to the build.

How much do the AI model calls cost to run?

Usually less than people expect when the model is chosen per task. We read an insurer's full claims book, about 13,000 documents, for under $25 in model costs: $12.68 to turn some 6,500 invoices into structured lines, $5.96 to read about 5,500 handwritten forms twice, and $1.60 to classify every claim. The expensive part of an agent is building, testing and maintaining it, not the tokens.

How long does it take to build an AI agent?

Scoping takes days. A first agent in production, with a human review step, typically takes six to ten weeks after the audit. The biggest variable is how clean and reachable the data is.

What makes an AI agent more expensive to build?

Four things: the number of systems it has to read from and write to, the number of exceptions in the process it automates, the quality of the data it works on, and whether it is allowed to change records on its own. An agent that only drafts for a person to approve is far cheaper to make safe than one that writes to your ERP directly.

Is it cheaper to build an AI agent with no-code tools?

To start, yes. A no-code workflow can be running in a day. The cost shows up later: when the process has exceptions, when the agent needs your users' own permissions, or when something breaks at 2 a.m. and nobody owns it. No-code is the right answer for a simple internal workflow. A process your business depends on usually needs code, tests and someone who runs it.

AI Agent Development Cost: Our Real Prices (2026)

Three blocks of rising height topped with increasingly complex robot arms, next to a stack of coins and a meter connected to the tallest block

Short answer: at Reflekt Lab, an AI agent costs $2,300 to $4,700 to scope, from $17,500 to put the first one into production, and from about $470 a month to run. The model calls themselves are usually the smallest line: we read 13,000 insurance documents for under $25. What makes an agent expensive is not the AI. It is the number of systems it touches, the exceptions in the process, the state of the data, and whether it is allowed to change records without a person approving. Below: our full price ladder, what moves the number, the running cost nobody budgets for, and how to spend the least before you know it works.

Our prices

These are the prices published on our AI integration page, converted to US dollars from our euro prices.

Step	What you get	Price
Express scoping (one workflow)	A half-day workshop on one process, a written summary, a clear answer on feasibility and the size of the pilot	$2,300 to $4,700
Full AI audit	Your use cases ranked by return and effort, the model choice for each, and a scoped pilot with success criteria	$7,000 to $14,000
Audit plus a first agent in production	The audit, then the first workflow built through to production, with your team trained on it	from $17,500
Running it	Hosting, monitoring, re-testing when models change, and a monthly allocation of changes	from about $470 per month

The audit is credited against the pilot if you build it with us. Model usage is a separate line in the quote, because it depends on your volume, not on our work.

For comparison, the cost guides that rank for this question mostly quote $20,000 to $60,000 and up for a first agent, often behind a calculator that ends in a sales form. Our first step is smaller on purpose: you should know whether the agent works on your real data before you pay for the whole thing.

What the AI itself costs to run

Less than most people fear, when the model is chosen per task rather than "the best one for everything".

When we read an insurer's entire claims book, about 13,000 documents, 96% of them photos of paper, the model bill came to under $25 (the full write-up):

$12.68 to read some 6,500 invoices into structured lines;
$5.96 to read about 5,500 handwritten claim forms, twice each;
$1.60 to classify every claim.

That is possible because each step uses the cheapest model that is accurate enough for it. In our benchmark of 9 models over 1,260 tests, routing each task to the right tier cut cost by 10 to 100 times without losing reliability. An agent that sends every request to the most expensive model is not more capable. It is just more expensive.

What moves the price

Four things, roughly in this order:

How many systems it touches. Reading from one spreadsheet is cheap. Reading from your CRM, writing to your ERP and sending through your email platform means three integrations, each with its own authentication, limits and failure modes.
How many exceptions the process has. "Approve every invoice under $500" is a rule. "Unless it is from a new supplier, or the PO is missing, or the amount changed since the quote" is where the work is. We count exceptions during scoping, because they drive the build more than anything else.
The state of the data. An agent working on clean, structured records is a few weeks. One working on scanned PDFs, free-text emails and three spellings of every customer name needs a cleanup step first.
Whether it can act on its own. An agent that drafts and lets a person approve is far cheaper to make safe than one that writes to production directly. When it does write, we prefer making every change visible with a one-step undo to an approval button people click without reading: we explained why here.

The cost nobody budgets for: after launch

The build is a one-time cost. An agent is not a one-time thing.

Models change every few months. A new version can be cheaper, better, or subtly different on your exact task. Someone has to re-test the workflow before switching, and sometimes because the old one is retired.
Providers go down, and bills run out. Two of the Claude incidents in early October 2026 were billing faults that refused paying customers for hours. We had a similar one on our own account in September. An agent needs someone who gets paged, and code that stops instead of retrying forever. The full story is here.
Your process changes. A new product line, a new supplier, a new field in the ERP. The agent has to follow.

That is what the monthly running cost pays for. If a quote has no line for it, ask who will do it, and at what price.

How to spend the least before you know it works

Start with one process, the one that costs your team the most hours and has the fewest exceptions. Not the most impressive one.
Pay for scoping first. $2,300 to $4,700 tells you whether it is feasible, what it will cost to build, and what the agent should not be allowed to do.
Ship it with a person in the loop. The agent drafts, a person approves. Remove the approval only for the cases where the data shows it is never needed.
Measure the hours, not the demo. The question after six weeks is how much time it saves on real work, not how impressive it looked in a meeting.

Also: OpenAI Agent Builder is shutting down: where to move your agents.

How Much Does It Cost to Build an AI Agent? Our Real Prices (2026)

Most cost guides give you a range wide enough to be useless and a calculator that ends in a contact form. These are the prices we actually quote, what moves them, and the part nobody budgets for: running the agent after launch.