Monday, 8:50am. Marek, the owner of an 11-person design studio, pastes a task description and three emails from a client into Claude to remind himself of the delivery arrangements. A fourth email stays in the inbox. AI Connections in Task Manager let you ask the model about data you have access to, without manually gathering it into the conversation.
Why does copying emails into the model stop being enough?
When you copy content yourself, you decide what the model sees, so it's easy to miss a message that matters for the answer. Each time, you also have to find the task, gather the correspondence, and bring it into a new conversation.
Marek asks about the delivery date, but doesn't copy the email where the client changed the arrangements. Claude can answer based on the three messages it received, not knowing a fourth one exists. The answer sounds convincing, yet Marek still has to go back to the inbox and check the result.
A connection doesn't free you from checking the answer against the source. It changes how the data is delivered: instead of pasting in fragment after fragment, you let the model search and read the items available to you in Task Manager. This still means sharing the read content with the AI provider, so the decision to enable the feature should belong to the company, not follow from one person's convenience.
How do AI Connections find the answer via MCP?
ChatGPT or Claude searches for items available to the user who connected their account, and then reads the content it finds. AI Connections rely on MCP, an open standard for linking models with other applications.
The model has four tools at its disposal: it checks account information, searches, reads the item it finds, and checks the calendar. Search covers tasks, chats, messages, notes, file names, mail threads from the user's own inboxes, events, and the AI Knowledge Base - that is, summaries, decisions and conclusions from the tasks, chats, notes and files it has access to. It searches by text, including partial words.
Questions about time are answered by the calendar. For a chosen period, the model gets the due dates of visible tasks, its own events and reminders, and the team's absences. So Marek can ask Claude: "What have I got next week?".
Marek asks: “What did we agree with the Nowak company about the delivery?” Claude can search for “Nowak”, read the found task and the thread from Marek's inbox, and then put together an answer. If the text of a file is needed, the read tool can return it; long items are delivered in parts. The model reads mail only from the inboxes connected to Marek's account.
More on the scope of the feature is in the description of AI Connections in Task Manager. The difference from manual copying is simple: the model searches for sources, but the question and the judgement of the answer are still yours.
Who enables AI Connections, and how do you connect your own account?
An administrator must first enable AI Connections before each user can connect their own account to ChatGPT or Claude. You copy the connection address in Settings, under the Profile tab, and add it with the model provider as your own connector. The feature belongs to the Pro variant; simply having an account with the model provider doesn't activate it in Task Manager.
The administrator opens Settings, the Advanced tab, and the “AI Connections (MCP)” section, then uses the “Enable AI connections” toggle. When data encryption is active in the company, a warning appears before enabling: content read by the model will be decrypted and sent to the AI provider in unencrypted form. They must knowingly tick the consent box; the app records who gave it and when.
Now it's your turn. In Settings, under the Profile tab, find the "AI Connections (ChatGPT, Claude)" section and copy the address using the "Copy" button. In Claude, you'll add it in Settings via Connectors and "Add custom connector". In ChatGPT, go through Apps and connectors, and in developer mode select "Create". Provider menu names may change; look for the option to add a custom connector.
Once you paste the address, your Task Manager login page will open. Sign in with your email and password, and on the consent page check the application name and the return address. Choose the connection's validity period and click "Allow". By default, a connection lasts 90 days. Leaving the validity field empty lets you create a connection with no expiry date, but entering a number of days limits the access time.
Can the model see more than you, or change anything?
No: the connected model sees data according to your account's permissions and has read-only access only. It won't add a task, change its content, send an email, or delete an item in Task Manager.
If Marek has access to a task in a given Workspace, meaning the team's work area, he can ask about it; if he can't see it, he won't see it through Claude either. The global administrator has a wider view, as they see all workspaces even within the application itself. The model, however, doesn't get passwords, API keys, or settings. Other people's private conversations with the AI Assistant are also excluded.
On Tuesday, Marek asks Claude for a draft reply for a client. He gets the text in the conversation with the model, but he still has to check it, paste it, and send it himself. You can't instruct the model to close a task or send a message. This way, reading company data stays separate from acting on its behalf.
Messages from clients also require attention. An outside sender could write text in an email that reads like an instruction for the AI. The model might misinterpret it, so check the reply at the source. Even then, the connection won't let it make a change in Task Manager.
What does a Mail inbox set to "Read-only" give you?
An inbox set to "Read-only" lets you read correspondence in Task Manager but blocks sending messages from it through the application. You won't reply from it manually or select it for sending in Automations; a connected model can still read its correspondence if the inbox belongs to your account.
On Wednesday, Marek connects an inbox containing invoice-related correspondence on his own and sets it to "Read-only". He can ask Claude about available threads, but Task Manager won't send a reply from that inbox. If he wants to reply, he does it in another email program. With this setting, the SMTP server, meaning the server used for sending emails, is optional.
Mail in Task Manager is a Pro module. An inbox belongs to the person who connected it: the fact that Marek reads its threads through Claude doesn't share them with other employees or their connected models.
How do you cut off access and check reads?
You disconnect your own connection in Settings, under Profile, with the "Disconnect" button; the AI application loses access immediately. The administrator can disconnect any user's connection and check the "Connection Log" without viewing the content that was read.
The log shows who used the connection and when, which application and tool was used, and which item the read concerned. Entries older than 180 days are deleted automatically. If Marek sees a read on Friday that he doesn't recognise, he has a starting point for checking the connection, although the log won't show him the content of the model's reply.
A connection also ends when the chosen validity period expires, the password is changed, or the administrator blocks the account. The administrator can also disable the feature for the entire company. In that case, all connections stop working immediately, but they aren't deleted and will return once re-enabled. If you want to end a specific connection, use "Disconnect".
Where does the read content go, and what does the connection require?
Content read by ChatGPT goes to OpenAI, and content read by Claude goes to Anthropic, including outside the European Union. Task Manager must be accessible from the internet via HTTPS, meaning an encrypted connection to the site; on the provider's side, you also need a plan that supports custom connectors. The application's data still resides on your company's server, but having your own server doesn't mean that content passed to the model stays on it.
Before connecting, check your company's rules on client data and correspondence. If your company doesn't allow sending such content to an AI provider, don't connect the account. The distinction between where data is stored and where the question is sent explains the article what goes to the AI provider.
ChatGPT and Claude connect from their own servers, not from your computer. This is a different mechanism than the AI Assistant in Task Manager, which runs within the application using the company's own API key. AI Connections don't use this key.
Tomorrow, write down one question where you usually paste company emails into ChatGPT or Claude, and check with your administrator whether this content may be sent to the model provider. If the answer is "yes", check out AI Connections in the Pro variant: the model can read according to your permissions, while the company controls the connections. See Demo.
Frequently asked questions
Can ChatGPT or Claude change anything in Task Manager?
No. AI Connections give the model read-only access: it can't add, change, send or delete anything in Task Manager.
Are AI Connections included in the Standard variant?
No, AI Connections belong to the Pro variant. An administrator still has to enable them before users can connect their own accounts.
Will an employee see more through ChatGPT than in the app?
No. The model uses that person's account permissions and sees only the data available to them; it reads mail only from their own inboxes.
How long does a connection with ChatGPT or Claude last?
90 days by default, and you choose the validity period on the consent page. You can leave the field blank so the connection never expires on its own, but it will still end if disconnected, if the password changes, or if the account is blocked.
Is AI Connections the same as the AI Assistant in Task Manager?
No. The AI Assistant works in Task Manager using the company's API key, while AI Connections let you link your account in the app to ChatGPT or Claude through your own connector.
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.
