AI pricing
How much does an AI agent cost for a business in 2026
An AI agent for a business costs a one-off build plus monthly running costs. An agent for one process starts at €5,000 excl. VAT (about 125,000 CZK) at Prahapp, and the AI model bill is usually a few hundred to a few thousand crowns a month, for example about 295 CZK for 800 invoices a month.
An AI agent for a business costs a one-off build plus a monthly running cost. At Prahapp, an agent for one clearly defined process starts at €5,000 excl. VAT, roughly 125,000 CZK, and the monthly bill for the AI model itself usually lands between a few hundred and a few thousand crowns for typical business tasks. Most of the difference between quotes is not model pricing but the scope of the task, the number of connected systems and how much work is left to people.
An agent differs from a chatbot because it does not only answer, it acts: it reads email, extracts data from documents, writes to accounting or a CRM and decides across several steps. How that works is explained in what is an AI agent. Here we focus on money: what the price is made of, what ChatGPT, Claude and Gemini models cost in crowns, what the payback looks like and when a project is not worth it.
What an AI agent price is made of
Quotes for AI agents are hard to compare because each vendor counts something different. Split the price into four items:
- Analysis and design. Choosing the process, the exceptions, the rules and the points where a person must decide. Good analysis is the cheapest way to avoid rework later. Our first AI consulting session is free.
- Build and integration. The agent itself, its instructions, tools and connections to your systems, for example Pohoda, Money S3, Shoptet or email. Every integration adds cost because errors and permissions must be handled.
- Testing and safeguards. Verification on real data, amount limits, human approval for actions with consequences and logging. We describe this in the article on AI guardrails.
- Operation. The AI model bill, hosting, quality monitoring and adjustments as your documents, price lists and rules change.
The one-off items (1 to 3) make up most of the budget. Running costs are surprisingly low for agents that handle hundreds to thousands of tasks a month, as the calculations below show.
How much an AI agent costs by complexity
We set a fixed price after a written scope. Roughly, projects fall into three bands:
| Agent type | Example | Indicative build price |
|---|---|---|
| Single-purpose agent | sorting incoming email and drafting replies | from €5,000 (about 125,000 CZK) excl. VAT |
| Agent with several integrations | invoice processing into an accounting system with approval | priced by scope, fixed after specification |
| Several agents and custom data | agent over a company knowledge base linked to CRM and orders | priced by scope, fixed after specification |
Smaller changes and further development after launch are billed at €25 to €49 per hour excl. VAT, roughly 625 to 1,225 CZK. See the current AI project pricing. We deliberately do not publish one universal price, because two agents both described as "invoice processing" can differ tenfold depending on how many suppliers, formats and exceptions they must handle.
For comparison, foreign vendor blogs quote dollar prices from a few thousand to hundreds of thousands, depending on whether it is a simple rule-based bot or an enterprise autonomous system. Those figures are hard to verify and often come from firms that sell the build themselves, so treat them only as a rough frame.
Running costs: ChatGPT, Claude and Gemini models in crowns
Models are billed by tokens, the pieces of text a model reads or writes. Terms are explained in the AI glossary. An agent uses more tokens than a chatbot because it solves one task in several model calls: it reads a document, calls a tool and checks the result. Anthropic prices, according to the official Claude pricing page, are:
| Model | Input per 1M tokens | Output per 1M tokens | Note |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | for prompts up to 100K tokens |
| Claude Sonnet 5.5 | $2 | $10 | cache reads $0.10 |
Anthropic also lists a 50% discount for batch processing, which suits tasks that need not finish instantly, such as overnight invoice processing. OpenAI and Google use the same token-based pricing, but exact numbers change, so we always check them on the providers' pages before a decision. For conversion we use 21.651 CZK per dollar (Czech National Bank rate of 22 October 2026).
Example 1: accounting firm, 800 invoices a month. An agent from AI invoice processing reads each invoice and proposes a booking. We assume 6,000 input and 500 output tokens per invoice, which is 4.8 million input and 0.4 million output tokens a month. With Claude Sonnet 5.5 that is $13.60, about 295 CZK a month. With Claude Haiku 5.5 the bill is $0.68, about 15 CZK. Building the agent costs orders of magnitude more than running it.
Example 2: e-shop support agent, 2,000 tasks a month. Each task averages 8 model calls with 5,000 input and 400 output tokens each. That is 80 million input and 6.4 million output tokens a month. On Claude Sonnet 5.5 this is $224, about 4,850 CZK. On Claude Haiku 5.5 it is $11.20, about 240 CZK. The practical answer is a mix: a cheap model handles routine questions and passes harder cases to a stronger one, the same principle we use for an AI chatbot for websites.
These calculations are illustrative and rest on the stated assumptions. We measure real consumption on a sample of your data before launch.
Hidden costs a budget often misses
- Integration with older systems. Research firm Gartner predicted in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value and weak risk controls, as Reuters reported. Integration into existing systems is usually the biggest surprise.
- Preparing source material. An agent is only as good as the documents it uses. An outdated price list or contradictory terms lead to wrong answers. Plan for cleanup, possibly a company knowledge base.
- Human oversight. For payments, contracts and outgoing email, human approval makes sense. It costs time but protects against an expensive mistake.
- Maintenance. Document formats, system APIs and model prices change. Plan for a few hours a month. Four hours of changes at €25 to €49 is €100 to €196, roughly 2,500 to 4,900 CZK a month.
- Agent washing. Gartner also warns about ordinary chatbots and automations being relabelled as agents. Ask the vendor for a demo on your data, not a presentation.
Three Czech scenarios: what to choose by company size
A small accounting firm with dozens of clients usually starts with one agent for document extraction and checks how many cases pass without intervention. A mid-sized Shoptet e-shop is best served by customer support first, because questions about delivery and order status repeat and directly affect revenue. A manufacturer with a few hundred orders a month benefits from an agent that reads enquiries from email and prepares quotes for a salesperson to approve. In all three cases the right approach is the same: one process, a measurable goal, a prototype on your own data and only then expansion. Companies that start with three agents at once usually find that none of them has good enough source material or a clear owner.
Payback: when an AI agent is worth it
Payback is simple: hours saved times hourly cost minus monthly running cost. Take the accounting firm above. Processing one invoice by hand takes 4 minutes, so 800 invoices take about 53 hours a month. If the agent handles 70% of cases without intervention, it saves about 37 hours. At an internal hourly cost of 350 CZK (an assumption) that is about 13,000 CZK a month. After the model bill and a few hours of maintenance, roughly 10,000 to 12,000 CZK remains, so an agent costing 125,000 CZK pays back in about 10 to 12 months.
A project is worth it where there is:
- a repeating, high-volume process, ideally hundreds of cases a month or more,
- a clearly measurable result (hours, errors, response time),
- available data and defined exceptions,
- a person responsible for approvals.
An agent is not worth it for a process that runs three times a year or a task nobody can describe. In that case consider simpler process automation without AI, or training your team on ChatGPT, Claude or Gemini.
How to check the price before you pay
Before signing, demand three things: a fixed price for a clearly described scope, an estimate of the monthly model bill based on your data and a working demo. Prahapp, the AI division of Buinsoft Technology in Prague, builds a prototype on your data for free within 3 working days, and if you don't like it, you don't pay. If you are still figuring out where AI belongs, the process map in our article on agentic AI helps.
Want to know what an AI agent would cost in your company? Write to us via the contact page and within 3 working days you get a free working prototype.