In firms, AI works best where the input is clear, the output is clear and somebody checks the result: customer support, work with documents and text, code. Autonomous agents on complex tasks still need supervision. The Czech Statistical Office found that in 2025, 18% of firms with more than 10 employees used AI, half as many again as a year earlier. The Czech Republic is below the EU average, which is 20%. What an agent actually is and whether you need one is explained in the article What is an AI agent. This one is about what runs in practice.
- In 2025, 18% of Czech firms with more than 10 employees used AI; the EU average is 20%.
- It works best where the input and output are clear and somebody checks the result.
- In support, AI raised issues resolved per hour by 14% on average, but it was one firm.
- Autonomous agents handled about 35% of multi-step conversations in the CRMArena-Pro test.
How many firms use AI and for what
In January 2026 the Czech Statistical Office (ČSÚ) published the figures for 2025. 18% of Czech firms with more than 10 employees use AI, and 54% of large firms. The EU average is 20%; Denmark has the most firms with AI (42%), Romania the fewest (5%). By sector it ranges from 64% in IT to 6% in transport and warehouses.
What matters more is what firms use AI for. ČSÚ asked only those that use it:
| Used for | Share of firms with AI |
|---|---|
| Marketing, sales or customer support | 59% |
| Management, administration or HR | 40% |
| Accounting or finance (for example invoices processed by a machine) | 24% |
| Protection against cyberattacks | 24% |
| Research, development and new methods | 23% |
| Logistics or warehouse | 7% |
Among technologies, writing text, code or voice leads, used by 13% of all firms. Images, video and music are created by 10% of firms, and advanced text analysis, such as reading data out of documents or summarising long texts, is used by 7%. A quarter of firms with AI (23%) use RPA with AI elements. ČSÚ describes it as software robots that copy data or download attachments from e-mails.
The number that is missing: ČSÚ doesn't measure how many firms use AI agents in the narrower sense, or how much time it saved them. That's why this article shows only measured figures, no estimates.
What works
Customer support with a person in the loop
A well-measured case is support staff who get advice from AI. A study by Brynjolfsson, Li and Raymond (NBER, April 2023, revised November 2023) followed 5,179 support workers and found that an assistant with generative AI raised the number of issues resolved per hour by 14% on average, and by 34% for novices. It was one firm and one assistant, so it is an average from one deployment, not a promise.
In its guide of 19 December 2024, Anthropic explains why support suits agents: it combines a familiar chat with tools that can pull customer data and order history, and success is easy to measure.
An example for a team of 5 people: fourteen percent of five people is 0.7 of a person. That is how to read the number. Your result will differ.
Documents, e-mails and text
Advanced text analysis is used by 7% of firms, which is 41% of firms with AI. ČSÚ lists reading data out of documents and summarising text, and in finance invoices processed by a machine (24% of firms with AI). These tasks have one thing in common: the input (e-mail, PDF) is messy, but the output (a record, a category, a draft reply) can be described exactly and checked. Lumenax offers such work: it pulls the data out of e-mails, invoices and contracts and passes it on.
Code
Anthropic writes that code solutions can be verified with automated tests, so an agent can tell from the test results what to fix. Here too it adds that a person has to check whether the solution fits the whole system.
Questions over your own documents
An internal assistant that answers from company documents is simpler than an agent. Anthropic writes that for many tasks one model query with a lookup in the documents is enough. How to build one on your own server is shown in the article on Ollama. In xOffice such an assistant answers from the data in the system, sees only what the asking user may see, and every change to the data has to be confirmed by a person.
What doesn't work yet without supervision
Agents are weaker where a task takes more steps. The CRMArena-Pro benchmark (arXiv:2505.18878), which tests agents on business tasks around CRM, reports that the best agents handled about 58% of simple queries and about 35% of multi-step conversations. The test is from May 2025, and today's models may beat it. We know of no newer comparable measurement that we could cite.
The second warning is organisational. Anthropic advises testing agents in a sandboxed environment and setting guardrails so that the agent waits for a person at a blocker or at chosen points. The same goes for you: test on a copy, not on live data.
The rest is simple. What can't be described by an input and an output can't be given to an agent. A process that lives only in people's heads has to be written down first.
How to pick the first process
Five questions follow from what we know. The answers are our judgement, not quotes.
- Are the input and the output described? An invoice comes in as a PDF and leaves as an entry in the accounts. That works. "Improve communication with customers" doesn't.
- Who checks the result? If nobody does, start elsewhere.
- What does a mistake cost? A wrongly sorted e-mail can be fixed. A wrongly approved payment can't.
- How many cases a month is it? The more, the faster a pilot pays off.
- How will you measure the benefit? Before the pilot, measure how many minutes the current procedure takes. Without that you can't tell whether anything improved.
According to its page on AI, Lumenax works in four steps: scoping, prototype, operation, measurement. It usually promises the first working prototype within two to four weeks and a fixed price after a short scoping.
What running AI costs
Lumenax doesn't publish a price for a custom agent; it sets it after scoping. The running cost of a model depends on where it runs. For xOffice, Lumenax's page lists three routes, verified on 12 October 2026:
- Your own account with a provider (for example Anthropic or OpenAI): 0 CZK, you pay the usage directly to the provider according to its price list. Monthly limits are set in the system.
- A model on Lumenax's infrastructure in the EU: from 4,900 CZK a month. The page states that the data goes to no other firm and nobody trains on it.
- An AI server in your network: from 5,999 EUR one-off plus installation. Lumenax recommends it only for the highest security, when data must not leave the firm.
More about each route is on the xOffice page. The prices are starting prices; Lumenax confirms the final amount after a fit-gap analysis.
Common questions
How many Czech firms use AI?
According to ČSÚ, 18% of firms with more than 10 employees in 2025, half as many again as a year earlier. Among large firms it is 54%, and the EU average is 20%.
Does AI work in customer support?
As a helper for people, yes. An NBER study of 5,179 support workers measured a 14% average rise in issues resolved per hour. It covers one firm, so your result will differ.
How reliable are agents on business tasks?
Not on everything yet. The CRMArena-Pro benchmark from May 2025 reports that the best agents handled about 58% of simple queries and about 35% of multi-step conversations. Newer models may be better. You'll find reliability out on your own cases.
Where to start?
With one process that has a clear input and output, where somebody checks the result and a mistake destroys nothing. Before the pilot, measure how many minutes the current procedure takes.
What does AI in xOffice cost?
Your own account with a provider costs 0 CZK and you pay the provider for usage. A model on Lumenax's infrastructure is from 4,900 CZK a month and an AI server in your network from 5,999 EUR one-off plus installation. The xOffice page marks the prices as starting prices.
Sources
- Czech Statistical Office, ICT use in the business sector 2025, chapter 10 Artificial intelligence, January 2026 (file dated 13 Jan 2026): https://csu.gov.cz/docs/107508/b09756a7-a39c-9eb9-b845-b69c4d89bd8d/06200525a10a.pdf
- E. Brynjolfsson, D. Li, L. Raymond, Generative AI at Work, NBER Working Paper 31161, April 2023, revised November 2023: https://www.nber.org/papers/w31161
- Anthropic, Building effective agents, 19 Dec 2024: https://www.anthropic.com/engineering/building-effective-agents
- CRMArena-Pro, arXiv:2505.18878, v1 24 May 2025: https://arxiv.org/abs/2505.18878
- Lumenax, xOffice ERP system, verified on the page 12 Oct 2026, viewed 25 Oct 2026: https://www.lumenax.cz/erp-system
- Lumenax, page on AI in operation, viewed 25 Oct 2026: https://www.lumenax.cz/ai-implementace
Ollama: how to run local AI over your own company data
What is an AI agent and when a firm really needs one
Risk analysis for NIS2 and ISO 27001: a step-by-step guide
