Enterprise AI is a machine connected to your systems. It works with your enquiries, drawings, price lists and mail, not with whatever someone pastes into a chat window. It comes in two forms. The on-premise one runs on your server and the data never leave the company. The cloud one runs with me, and only the question and the answer go out, not your database. Both include training your people, not just an installation.
One decision that can be changed later. Both forms have the same training for your people, the same morning report on whether things in your company happen on time, and the same rules on where the machine may act on its own. They differ in where they live and what leaves your network.
| On-premise | Cloud | |
|---|---|---|
| Where the data lives | With you. The database, the documents and the model run on a server in your network. | Your systems stay with you. The model and its memory for the given task run on infrastructure I manage. |
| What leaves your network | Nothing. The server does not need the internet to do its work. | The question and the answer. The one enquiry or document it is working on right now. The whole database and its history stay with you. |
| What it costs | Hardware once, bought separately and yours. Implementation from approx. EUR 3,700. Oversight from approx. EUR 330 per month. No fee per question or per user. | Implementation from approx. EUR 3,700. Operation from approx. EUR 330 per month including models. No hardware, no upfront investment. |
| Who has access | Your people, according to a list of accounts you get on one page. Me for the time of the implementation, and after that only remotely, when you switch it on for me. The machine sees only what the person asking is allowed to see. Permissions are set on the data, according to your company list of accounts, not according to what the machine specialises in. | Your people, according to the same list. Me, as the one who runs the infrastructure. The model provider sees the question and the answer; it does not see into your systems. The machine sees only what the person asking is allowed to see. Permissions are set on the data, according to your company list of accounts, not according to what the machine specialises in. |
| What the law says | The AI Act rules on labelling (Art. 50, from 2 August 2026) and on training people (Art. 4, from 2 February 2025) apply the same way. Under GDPR no one new processes your personal data; it stays with you. | The same AI Act rules. Under GDPR one more processor of personal data is added, the model provider, together with a contract with it. I prepare that contract as part of the implementation. |
| When it makes sense | Your contracts with customers say the data must not go to a third party. You process hundreds of documents a day and the volume does not change. You want one cost now instead of a bill every month. Do not do it for the price alone: at a small volume the cloud comes out cheaper. | You want a result within four weeks and do not want to buy hardware. The volume of work is smaller or it fluctuates. The task does not work with data that must not leave, or the data can be redacted before sending. |
Most companies are still waiting. 18 percent of Czech companies use AI, and 54 percent of companies with more than 250 people. The technology is the same for everyone. The difference is that the large companies had people who answered the three questions below.
Czech Statistical Office, ICT usage in the business sector, data for 2025, published 22 May 2026.
True: Most offers on the market are cloud, and the data really do leave. Many companies stayed away from AI because of it: in the Czech Republic 80 percent of companies name data security as a concern, and in the EU almost half of the companies that considered AI and did not adopt it.
Not true: That there is no other way. The on-premise form runs with you and nothing leaves it. And with the cloud form you can say exactly what goes out: the question and the answer, not the database. That is the difference between "AI in the cloud" and "your data in the cloud".
Randstad HR Trends 2026, 210 Czech companies, 1 April 2026. Eurostat, ICT usage in enterprises 2025: data protection 49 %, unclear legal consequences 53 %.
True: The law exists and two of its parts already apply. Since February 2025 a company has to make sure that the people who work with AI on its behalf know how it works and where its limits are. From August 2026 it has to be recognisable when a machine is writing.
Not true: That it is getting stricter. The strictest part, the rules for high-risk systems, was postponed in July 2026 to December 2027 and August 2028. And tasks like mail, enquiries, drawings or price lists are not high-risk systems. Training your people is part of the implementation with me, and labelling messages is the job of whoever supplies the system, which is me.
Regulation (EU) 2024/1689, Art. 4 in force from 2 February 2025, Art. 50 from 2 August 2026; amendment in force from 27 July 2026 (Czech Ministry of Industry and Trade, 29 July 2026).
True: A per-user subscription is the form you will meet most often, and for a company with thirty people it adds up to a decent monthly sum for a tool that three of them use.
Not true: That it is the only form. Here you do not pay per head, you pay for running one task. With the on-premise form you pay for the hardware once and then for oversight. With the cloud form for operation including models, itemised and with a cap on spending. And before you pay for any of it, the AI plan tells you within 48 hours whether there is anything to build on at your company at all.
Randstad HR Trends 2026: 15 % of companies name implementation costs as a barrier.
Enterprise AI is not one product. It is a decision about where the data lives (see above), and under it three different pieces of work. It always starts with the smallest one.
It answers a single question: what AI will help my company with, and what it won't. Within two days you have a list of tasks ordered by how many minutes a day they take today, and for each one whether it is a job for a machine, for a person, or for cleaning up the data first. It comes before everything else, and its price is deducted from the next step. What is in the plan
A program with its own access to your systems. It runs non-stop, looks for what needs doing by itself, and every morning reports at the top what is stuck in your company: an enquiry without a quote, an order without delivery, an invoice past due. Only below that, what it did itself and where it did not know what to do. One task from start to finish, for example the mailbox with enquiries. How the agent works
It takes an assignment, produces something new and stops. A drawing from an enquiry, a list of companies to approach based on your brief, an article from notes, a quote from a price list. It does not run until you start it, and it has no access anywhere except where you send it.
You can tell 2 and 3 apart by one question: should it work even when nobody has asked it to do anything? If yes, it is an employee. If not, it is a robot.
Both machines read from your systems. When a customer is in four systems under four names, or enquiries live in one person's head, there is nothing to read from. That is why company preparation and integration (priced per job) come first. If the AI plan shows that this is missing, we start there.
What it looks like in practice: see the working demo of a fictional company.
Two forms, two different numbers. Here is what is in them.
Hardware, separate. Bought separately and yours. A server with a graphics card sits at your place and you own it, even if we stop working together. I choose it, and you order it in your company's name.
Implementation from approx. EUR 3,700, included. Installing the model and the agent on your server, connecting it to your systems, setting the limits of where it may act on its own, training your people with a written record, and the first month under supervision.
Oversight from approx. EUR 330 per month, included. I make sure it runs, fix it when something changes at your company, and replace the model when a better one comes out. No fee per question, no fee per user.
Electricity, separate. Paid by you.
Implementation from approx. EUR 3,700, included. The same as on the left, only without your server: the agent runs on infrastructure I manage.
Operation from approx. EUR 330 per month including models, included. Spending has a cap that is set during implementation.
Nothing else. No hardware, no electricity, no upfront investment.
Both forms include training the people who will work with it, and a written record of that training. The AI Act (Art. 4) requires it, so this obligation is covered. And in both forms you get a list of access rights, a description of the setup and a data export, in case it should ever end.
The most common reason AI falls asleep in a company is not the technology. It is that people do not know what they may put into it, what to expect from it and when not to trust it. Most companies know the subscription that management bought and nobody uses.
That is why both forms include training the people who will work with it. Not a lecture about AI. Two hours over their own tasks: what the machine does on its own, where it stops and waits, how to read the morning report and what to do when it is not sure. The report does not measure who uses the machine and who does not. It measures whether things are ordered, delivered and paid on time.
And a written record of it. Since February 2025 the AI Act requires a company to make sure that the people who work with AI on its behalf understand it. There is no certificate and no required number of hours. The training record covers that obligation for you.
After the training, people can tell what the machine did on its own and what it left to them.
They know what must not be put into the machine, and why this matters less with the on-premise form.
They can stop it when it does something other than they expected, and write to me about it.
For some companies the AI plan within 48 hours ends with the sentence "not yet". These are the four reasons that come up most often.
Enquiries in people's heads, a customer under four names, a price list in two files. The machine cannot read anything reliable from that. Company preparation comes first, and the price of the plan is deducted from that instead of from the AI.
A good thing, and it costs a small monthly fee per person. Buy it, give people rules on what they may put into it, and buy nothing from me. Enterprise AI makes sense only once the machine should work with your systems on its own.
When your systems do not talk to each other and things are retyped by hand, AI only speeds up the retyping and repeats its mistakes. Integration comes first. Sometimes it turns out that AI is no longer needed.
A model running on a server at your place is smaller than the largest one in the cloud. It is enough for mail, enquiries, price lists and drawings. It is not enough for complex legal analysis or long reasoning. If you want both at once, I will tell you in the plan and we will not force it.
A list of tasks, and for each one whether it is a job for a machine, for a person, or for cleaning up first.
For the on-premise form, counted from the moment the hardware arrives.
You pay for running the task, not for the number of people who look at it. In both forms.
About training the people who work with it. An obligation under the AI Act that this record takes care of.
48 hours and 4 weeks are deadlines from the offer, not measured averages. Zero per user is a pricing rule. The record is part of the delivery.
I go through the tasks that take up time with you, and for each one I say whether it is a job for a machine. What I need from you: two hours with the people who do that work. Output: a plan and a recommendation, on-premise or cloud.
Based on the plan and your contracts you choose the form. For on-premise I choose the hardware and you order it in your company's name. What I need from you: the decision and the order. For the cloud form this week is skipped.
For on-premise, counted from the delivery of the hardware. Installation, connection to your systems, setting the limits of where it may act on its own. What I need from you: access and one hour with the person who does the task today.
Two hours with the people who will work with it, over their own tasks. What I need from you: those people in one room.
A report every morning: at the top, what is stuck in your company (an enquiry without a quote, an order without delivery, an invoice past due), below it, what the machine did. For the first month I read it with you. What I need from you: five minutes a day. After that, only oversight.
With the cloud form it takes roughly five weeks from the plan to operation. With on-premise, the delivery of the hardware comes on top, usually two to four weeks; I cannot influence that and I say so upfront.
No. That is the whole reason there are two forms. Your own server makes sense when your contracts say the data must not go to a third party, or when you process hundreds of documents a day. If not, start in the cloud: no hardware, running within four weeks. And if it later turns out you want the data at home, we move it. The setup, the limits and the training stay; only the place where it runs changes.
It will; it happens roughly twice a year. With the cloud form I replace it as part of operation, and nothing changes at your end. With the on-premise form I download it to your server and test it on your task before I switch; that is part of oversight, not an extra charge. The hardware does not change because of it: a server that is enough today is enough for a new model of the same size. If a new model needed a bigger machine, I would tell you and you would decide.
With the on-premise form, your people according to a list of accounts you get on paper, and me for the time of the implementation. After that only remotely, when you switch it on for me. Nobody else; the server does not need the internet to work. With the cloud form I am added as the operator of the infrastructure, and so is the model provider, which sees the question and the answer: one enquiry, one document. It does not see into your systems and it does not have your database. I prepare the data processing agreement with it as part of the implementation. The machine sees only what the person asking is allowed to see. Permissions are set on the data, according to your company list of accounts, not according to what the machine specialises in.
It is not, and I will not claim it is. A model running on one server in a company is smaller than the largest one in the cloud. For tasks like mail, enquiries, price lists, drawings and minutes it is enough, and you will not notice the difference there. For long reasoning or legal analysis it is not. In the AI plan I tell you for each task which group it belongs to, and I recommend the form accordingly.
No. Two things apply: since February 2025 the people who work with AI should know how it works (that is the training and the record, part of the implementation), and from August 2026 it has to be recognisable when a machine is writing (that is the job of whoever supplies the system, which is me). The rules for high-risk systems were postponed by the July 2026 amendment to December 2027 and August 2028, and tasks like mail or drawings are not among them anyway.
For what it does, it is enough, and keep it. It helps the person sitting at it. Enterprise AI does a different thing: it works with your systems on its own, even when nobody is sitting anywhere, and in the morning it says what it did. Whether you need that second thing comes out of the AI plan. For some companies the answer is no.
That is the only way I do it. The smallest piece is the AI plan for approx. EUR 400 and two days. Then one task, not five. When it proves itself, the second task does not pay for the implementation again; it runs on the same foundation. And the robot on assignment can be tried on a single assignment before the size of the job is agreed.
Write to me about three things that are done over and over at your company and that nobody wants to do. Within 48 hours I will tell you for each one whether it is a job for a machine, and whether it should live with you or in the cloud. Until then we do not talk about hardware or subscriptions.
Get in touch