Michał runs a seven-person design office. On Monday he spots on a designer's screen a client email pasted into a public chatbot - complete with a name, a site address and a figure from a cost estimate. The designer just wanted to shorten three paragraphs into two sentences. Michał needs to establish what the team shouldn't paste into ChatGPT at work, and where they can ask AI about company matters instead.
A ban alone can push AI use onto private accounts
A ban on using AI alone doesn't give the team a way to do the work they previously used a chatbot for. Without a designated alternative, an employee may turn to a private account or their phone, and the company won't know what content was sent there.
Using AI tools this way, outside the company's knowledge, is called shadow AI. It's the name of a phenomenon, not a judgement of Michał's designer. In this story, she just wanted to edit an email faster; nobody had told her beforehand what she could paste in, or where she should do it.
So start with a conversation about actual tasks. If the team asks AI for draft replies, translations and summaries, define acceptable content for each of these activities. Only then point them to a tool.
Text sent to the model leaves the company
Text sent to the model reaches its provider, even when you're using the AI Assistant in Task Manager. Having the application on your own server doesn't change where a query with context goes.
When the team uses a company tool instead of a random private account, the administrator chooses the provider and sets access, and the company pays for the model's usage. In Task Manager, the administrator enters the company's API key. An API key is the data that lets the program connect to the provider's service; the company is billed directly by OpenAI, Google Gemini, Anthropic Claude or xAI Grok, with no middleman's margin. Check the terms of your plan with your chosen provider.
That doesn't mean you can send anything to the AI Assistant. Even content encrypted within the program has to be decrypted before it's passed to the model. We've covered the scope of data shared separately in our article on what reaches the AI provider.
ChatGPT at work: what's OK to paste in, and what should give you pause?
You can send a public chatbot content whose disclosure outside the company wouldn't cause any harm. With customer data, commercial terms and internal documents, pause before you hit “Send”.
| Usually fine for a public chat | Pause before sending |
|---|---|
| Your own email draft with no customer data | An email with a customer's name, address or case details |
| Brainstorming notes with no confidential findings | A cost estimate, an offer with commercial terms, or a contract |
| A question about how to phrase a sentence | An HR document or internal personnel decisions |
| Marketing copy intended for publication | A password or API key - don't paste these into a chat |
The rule for the team fits into one question: would you send this same content by email to a stranger outside the company? If not, don't paste it into a public chat. This is a practical working rule, not legal advice; the topic of regulations and data storage location is covered in our article on task software, GDPR and your own server.
For the email Michał saw, the answer is simple. The designer can ask a public chat how to shorten neutral text, but she shouldn't send the whole correspondence containing customer data for that purpose. She also needs a place where she can work on the actual thread itself, without copying it into a foreign window.
Three decisions give the team one clear path
The team needs a designated place for questions about company matters, a company account with the model provider, and a defined scope of access. So Michał works out with the team where to direct questions about company threads, who manages the account, and who can use AI.
First, name the place: “we ask AI about task threads in our work program, not on private accounts.” Then decide who manages the key and billing with the provider. Finally, define who can use AI and what company content they can access. Simply asking people to “be careful what you send” won't replace these decisions.
Also write down what to do with a borderline case. On Tuesday, Marta wants to translate a customer message containing a quote, while Krzysiek wants to translate a public service description. They don't need to get the same answer just because both tasks are called translation.
In Task Manager, the key and AI settings belong to the company
In Task Manager, the administrator chooses the model provider and enters the API key in Settings, under the AI tab. The program encrypts the key in the database and, once saved, no longer displays it in full.
Michał can choose OpenAI, Google Gemini, Anthropic Claude or xAI Grok. After entering the key, he refreshes the list of models fetched from the provider's account and picks a model; one for each provider is marked "recommended". The company pays the provider directly for AI usage, rather than through a per-user subscription. The billing method and configuration are described in more detail in the article on using your own AI key in a company.
In the settings, the administrator also decides whether conversation history with the AI Assistant should be saved in the database; this is enabled by default. They can set the maximum response length from 1024 to 16384 tokens, with 4096 as the default. Tokens are the units the provider uses to bill processed text; a longer response can increase the cost. Task Manager itself, along with the application data, runs on the client's server - ordinary shared hosting is enough.
Permissions limit what the AI Assistant can access
The AI Assistant only accesses content available to the user asking the question. The administrator enables AI separately for tasks, chats and the AI Assistant window; a person without the relevant permission does not see that feature in the interface.
Marta can select her own tasks, Calendar, Workspace, a specific task, chat or note as sources. In the Pro plan, where the Files and Mail modules are available, she can also select a folder or file, as well as her own mailbox. A usage bar shows how much of the available context has been filled; if some material doesn't fit, the AI Assistant states what it left out. You can read about who sees which tasks and how permissions work in a separate article.
Mail as a source of AI knowledge is disabled by default. The administrator has to enable it deliberately, since the content of the selected correspondence will be sent to the model provider; once enabled, a user can only select their own mailboxes. The AI Assistant's access to the internet requires a separate permission and lets it read a specified page, rather than search the web like a search engine.
An employee's questions remain private from the rest of the team. No one else, including the administrator, can see conversations with the AI Assistant. When Marta asks for a summary directly within a task or chat, only she sees the question and the answer, and the answer is marked "FOR YOU ONLY".
A company thread doesn't have to move to another window
An employee can ask the AI Assistant for help without copying a task or chat thread into a public chatbot. The content needed for the answer is pulled from within the program, but the query with its context is still sent to the chosen model provider using the company's key.
On Wednesday, Marta opens a conversation about the project and privately asks, "Summarise this thread". If she just wants to read a long exchange of messages in a different way, she uses Organise mode with the "Summarise" option; that result only lasts for the current session. She chooses a private question when she wants to follow up on decisions made, and Organise mode when a shortened version of the conversation on screen is enough.
While drafting a reply, Marta can use "Enhance text" in the editor. The feature improves the language and style, keeps the text layout, and the change can be undone. She can also ask the AI Assistant to translate or read an attachment in a thread she has access to, for example a DOCX file or a PDF with a text layer.
On Friday, Michał needs a document with the agreed decisions. The AI Assistant prepares the content, and the application compiles it into a Word or PDF file ready to download. Michał can discuss these uses of AI in office work with the team and designate one place for routine questions. These uses don't override the rule that the scope of company content must be assessed before sending it to a model.
Know the program's limits before turning AI on
Task Manager doesn't have its own AI model. Without an account and key with a provider, AI features won't work, although the program itself still handles tasks and team communication.
The AI Assistant is no substitute for a search engine: even with internet permission, it only reads the page address you specify. It also won't recognise text in a PDF that is just a scan, since it has no OCR, i.e. the feature that reads letters from an image. Nor does it read old DOC files. So check external questions in a browser, and prepare any document meant to be read by AI in a supported text-based format.
It's also worth distinguishing between plans. In Standard, the AI Assistant uses the current conversation and any sources you specify. In Pro, Mail, Files, the AI Knowledge Base and AES-256 content encryption are also available; the AI Knowledge Base lets the Assistant find information within the company's knowledge. Content encryption doesn't stop the context from being sent to the model provider - the program decrypts it before sending. You'll find help choosing in the comparison of Standard and Pro plans.
Tomorrow, write down five things your team sends to AI
Tomorrow, ask your team about five types of content they've recently sent to AI. For each one, note whether it could safely leave the company, and alongside it, point to a place for company-related questions.
For Michał's company, Task Manager means tasks and conversations on their own server, AI on the company's own key, and access limited by permissions. The licence is perpetual, with no per-user subscription: the Standard plan in the Starter tier, for up to 8 people, costs PLN 690, and in the Classic tier, for up to 20 people - PLN 1,490. The price includes 12 months of updates; after that, renewal is optional. The company pays for AI usage separately, directly to the model provider.
Once you've written down the rules, see what working in the program looks like: Demo.
Frequently asked questions
Can employees use ChatGPT at work?
Yes, you can allow ChatGPT for specific tasks and set out what content your team shouldn't paste into it. Also designate a place for questions about company matters, rather than leaving employees with nothing but a ban.
What shouldn't you paste into a public chatbot?
Don't paste passwords, API keys, customer data, cost estimates, commercial terms, contracts or HR documents. Apply the rule that you don't send content you wouldn't send to a stranger outside the company.
Does the AI Assistant in Task Manager send content outside the company?
Yes, a query with context goes to the chosen model provider using the company's key. The fact that the application's data sits on your server doesn't mean the model runs on that server.
Can the administrator see employees' questions to the AI Assistant?
No, conversation threads with the AI Assistant are private, even from the administrator. The administrator does, however, decide who has access to the AI features.
Can you limit what the AI in Task Manager can access?
Yes, the AI Assistant uses content available to the user asking the question, and the administrator enables AI separately for tasks, chats and the AI Assistant window. Mail as a knowledge source is disabled by default, and internet access requires a separate permission.
Sources
The Task Manager Team - we write about teamwork, AI in business and keeping tasks organised, using examples from Task Manager, software installed on your own server. Task Manager has been developed by Blue Creation Group (Piotr Łącz) since 2013 and is used by organisations including Neonet, the Military University of Technology and Polpharma - read customer reviews. We check our articles against the user guide and the latest version of the software.
