Using AI safely at work: Copilot, ChatGPT and agents
People already work with AI: they use it to write quotes and summarise email.But you often don't know which tools these are or what data goes into them. We'll help you set rules, monitor data and sort out permissions in Microsoft 365, so the assistant doesn't show people more than they should see.
The initial consultation is free: we'll go through which AI tools your people use and what for, and we'll recommend where to start, even if you end up choosing someone else. We'll get back to you by the next working day at the latest.
endpoint protection that recognises, in the browser, sensitive data being pasted into AI tools
Copilot
we review Microsoft 365 permissions beforehand
Applies to Copilot, ChatGPT and other AI tools
Where company data goes
A salesperson pastes a price list and a customer's terms into a chat so that the AI writes them a quote. Nobody looks into which tool it was or who can get to that data.
The company turns on Microsoft 365 Copilot. A colleague from production asks about the salary spreadsheet, and the assistant finds it, because someone once shared it with the whole company.
An email arrives with hidden text that a person can't see. But an assistant that summarises email may read it as an instruction.
What happens if nothing changes
An AI assistant works with everything it has access to. How much that is depends on the settings.
First:
a customer's contract sits in a tool the company doesn't know about and doesn't manage.
Then:
Microsoft 365 Copilot may show a colleague a document they weren't supposed to see. The fault lies in sharing that someone set up wrongly long before. The assistant merely exposes it.
Finally:
an assistant or agent comes across an instruction hidden in an email or on a website. If it is allowed to send or delete things without a person's approval, it may carry the instruction out.
According to the Czech Statistical Office, 35% of businesses in Czechia with ten or more employees used AI in 2026, compared with 18% a year earlier. Among small businesses with 10 to 49 employees, the figure was 30%.
AI can help attackers too: according to the National Cyber and Information Security Agency (NÚKIB), scam emails are often written in flawless Czech, and AI is probably contributing to their quality.
When permissions are in order and data is monitored, the assistant won't show people what they shouldn't see, and less sensitive data goes into AI tools.
What we'll do and what you get
We start with an initial consultation and a written quote, so before you order anything, you'll know exactly what we'll do.
1. Which tools, and with what data.
At the initial consultation, we'll go through which AI tools people use, what for and with what data.
You get: a recommendation on where to start and which tools are suitable for company data.
2. Rules and training.
We'll discuss rules for the whole company with management: what doesn't belong in AI, which tools are allowed and who decides on new ones. We'll train your staff.
You get: training for staff and management, three hours, online or at your premises.
3. Before you turn on Copilot.
We'll go through permissions and sharing in Microsoft 365 with you, so the assistant doesn't show people more than they should see. The same applies to any assistant that gets access to company files or email.
You get: an overview of what was shared more widely than necessary and how it is set up now.
4. Data monitoring.
We'll deploy protection on your computers that, based on the configured rules, recognises sensitive data in the browser, such as personal data, and catches or blocks it when someone pastes it into an AI tool. It won't catch everything, which is why we start with rules and training.
You get: a monthly overview of what the protection caught, if we also manage it for you.
5. Assistants and agents.
When an AI assistant or agent is given access to email, files or other systems, we'll give it only the permissions it really needs. High-impact actions, such as sending something outside the company or deleting, should be approved by a person.
Want to know where else you have weak spots, for example in accounts, updates or backups? The Security assessment examines your company from the outside and the inside.
What we use
The foundation is rules and people who know what doesn't belong in AI. On top of that, technology monitors the data that goes into AI tools.
Check Point Endpoint Security, Elite edition.
Protection in the browser recognises sensitive data in files and in text that people paste into web-based AI tools, and catches or blocks it according to the rules. Its GenAI Protect component shows which AI tools people use in the browser, including unapproved ones. Endpoint protection
Fortinet FortiDLP.
It recognises sensitive data in text that people enter into the web versions of ChatGPT, Claude, Microsoft Copilot and Google Gemini, captures it as evidence and can block it from being sent.
Permissions in Microsoft 365.
Microsoft 365 Copilot shows only the company data that the user is allowed to open themselves. Before you turn it on, we'll sort out permissions and sharing.
We supply and manage both Check Point and Fortinet as their official partner. We won't hide the conflict of interest: we sell both vendors' products ourselves.
If rules, training and the Microsoft 365 settings you already have are enough for you, we'll tell you.
How the price is set
The price depends mainly on how many people work with AI, whether you're turning on Copilot and whether you also want data monitored technically.
In the quote, we always itemise:
training for staff and management, depending on the number of people, the content and the format
a review of permissions and sharing in Microsoft 365 before Copilot or another assistant is turned on
a Check Point endpoint protection licence in the Elite edition, paid monthly, with no annual commitment
a FortiDLP licence and its terms, if you choose Fortinet
ongoing management of the protection, if you want it
You pay separately:
Microsoft 365 and Copilot licences
responding to an attack, if it isn't included in the price of managed protection, at the hourly rate in the price list Hourly rates in the price list
Do we manage your IT? Work beyond the package, such as cleaning up permissions in Microsoft 365, is billed at the rate in the price list. If we don't manage your IT, we'll still go through your permissions and sharing, and itemise the price in the quote.
Your rules will then reflect what is really happening in the company.
Say one thing out loud: contracts, price lists, personal data and source code don't belong in public AI tools.
By a public tool, we mean one your company doesn't manage, even if someone signs in to it with their work email address. What may go into tools the company manages is set by your rules.
Check what in SharePoint and OneDrive is shared with the whole company.
Whatever everyone can see, Microsoft 365 Copilot can find too.
What businesses ask us
We've received a suspicious email, or one of our accounts has been compromised. What now?
Call +420 777 809 802 straight away, even if we don't manage your IT or it's outside your package's working hours. On working days from 7:00 to 17:00 we answer the phone; at other times, we'll try to accommodate you. If we don't pick up, email us and we'll get back to you on the first working day. If you can, change the password of the compromised account from another device. Don't delete messages in the mailbox: they are used to trace what happened. If we provide managed email, computer or mobile protection for you, its price includes routine intervention (quarantine, blocking, temporary isolation of a device) on working days from 8:00 to 16:00. Extensive investigation and recovery are paid separately. We bill other interventions at the hourly rate in the price list, with a surcharge outside your package's working hours (without a package, outside 7:00–17:00), at weekends and on public holidays.
Is Copilot Chat included in Microsoft 365? Can it see our documents?
Copilot Chat, which is included in selected Microsoft 365 licences, doesn't go into your documents by itself. It works mainly with the web and with what a person pastes in, uploads or has open in Outlook or Teams. It can reach other company data only through a paid agent. Microsoft 365 Copilot, with an add-on licence, finds documents by itself, but only those the user has permission to access. With a work account, company data protection, which Microsoft calls Enterprise Data Protection, applies in all variants.
What is a hidden instruction for an assistant (indirect prompt injection)?
An instruction hidden in content the assistant reads: on a website, in a document or in an email. A person may not see it, but the assistant may take it as a command. According to OWASP, it isn't clear whether a foolproof defence against this exists, and Microsoft itself says that some attacks may get past its protection. But the impact can be reduced: an assistant should carry out high-impact actions only after a person has approved them.
Under Article 4 of the EU AI Act (as amended with effect from 27 July 2026), companies that use AI systems in their work must take measures to support AI literacy among their staff and other people who work with AI on their behalf. The Act doesn't prescribe a specific level of knowledge or a certificate.
In web-based tools, yes. Check Point Endpoint Security, Elite edition, monitors in the browser the files people upload to the web and the text they paste into web-based AI tools, and catches or blocks them according to the rules. With Fortinet, FortiDLP does something similar, so far for the web versions of ChatGPT, Claude, Microsoft Copilot and Google Gemini. This protection doesn't extend to AI tool apps outside the browser or to personal devices.
Has NÚKIB warned against any AI tool?
Yes, on 10 July 2025, against products of the company DeepSeek, meaning its apps, websites and services. It advises the public to consider carefully whether to use them and, if they do, what information they put into them. For companies that don't fall under the Cybersecurity Act, the warning isn't generally binding, but NÚKIB recommends that they also deal with the threat.
Book an initial consultation
A few sentences are enough: how many people at your company work with AI, which tools they use and whether you're turning on Copilot.