Marek runs an 11-person agency. On Monday he opens Task Manager's settings and sees four AI providers: OpenAI, Google Gemini, Anthropic Claude and xAI Grok. Which one should he choose, given that the team will be asking the AI Assistant about tasks and company documents? The answer depends mainly on whether you need search in the knowledge base and which model you want to connect - and you can change your choice later.
AI provider for your company: what are you actually choosing?
You're choosing the company whose model Task Manager will send questions to, using your API key. You're not choosing a new place to store tasks, chats, notes or files - application data stays on your own server.
The boundary, however, is clear: when you use AI, the query along with the necessary context is sent to the model provider. This is not a promise that content used in the question will stay solely on your hosting. If you want to establish what gets sent to the AI provider before connecting a key, start by reviewing the sources your team will be using.
Task Manager also works without AI. Marek can therefore first organise the agency's tasks and set up an account with a provider only once he has a specific use for the AI Assistant.
Your own key for the chosen provider, billed directly by them
To connect your chosen provider, you need an account with them and your own API key. An API key is a confidential identifier that lets the provider recognise which account the application is using and whom to bill for usage.
You can generate a key on the chosen provider's website:
- OpenAI: https://platform.openai.com/api-keys
- Google Gemini: https://aistudio.google.com/apikey
- Anthropic Claude: https://console.anthropic.com/settings/keys
- xAI Grok: https://console.x.ai/
The administrator pastes the key into Settings → AI. Task Manager encrypts it in the database and, once saved, does not display it in full; API key encryption works in both the Standard and Pro plans. You pay for model usage separately, directly to the provider. There is no token package or margin added to your AI bill in the Task Manager licence.
If this is your first company API account, read how your own AI key works and how to control costs. In the advanced settings you can also limit the maximum response length: the available range is 1024-16384 tokens, with a default value of 4096. A token is the unit the model uses to measure processed text. Longer responses increase the cost with the provider, so you don't have to set the highest limit straight away.
The choice comes down to semantic search and image handling
Semantic search in the Pro Company Knowledge base requires OpenAI or Gemini. Visual image analysis, in turn, depends on the capabilities of the chosen model, not on the provider's name alone. Choosing a provider does not change the AI Assistant's core features in everyday conversation.
On Tuesday, Kasia from the agency is looking for the mock-up delivery deadline but can't remember the name of the company document with the client's arrangements. In the Pro Company Knowledge base, semantic search helps find content by meaning, not just by exact words. It requires OpenAI or Gemini and is enabled by default for these providers. The Standard plan has no Company Knowledge base or this toggle. There, the Assistant answers based on the current conversation and the sources indicated in its window. The way of working with sources is described in the article on the AI Company Knowledge base built from company documents.
The second difference is easy to miss. With OpenAI and Gemini, when semantic search is enabled, document fragments are sent to the provider as vectors - numerical representations of text used for searching by meaning. These requests are billed separately. With Claude and Grok, indexing the knowledge base alone sends nothing to the provider. This doesn't mean later questions to the Assistant stay on the server: queries with context are still sent to the chosen model. If you want to limit what content is sent during indexing, compare these two modes rather than looking for a general "private AI" label.
On Thursday, Marek attaches a photo of a damaged worktop to a task. The Assistant can visually analyse JPG, PNG, GIF and WebP images, but only if the chosen model supports images; large photos are resized before being sent. So check the capabilities of the model from the fetched list, rather than assuming the provider's name alone determines whether it can read a photo.
Everyday Assistant tasks don't need a different provider
A conversation with the AI Assistant, a private question in a thread and editing text all work regardless of which of the four providers you connect. The same goes for asking it to prepare a task, a reminder or a file to download; choosing a provider doesn't add a separate set of these features.
Ola needs to summarise a conversation about an offer. She turns on the private AI Assistant mode in the thread and asks about what was agreed; she gets a reply marked "FOR YOU ONLY". She can also use "Organise" and choose one of the options: Original content, Polish, Summarise or Expert. In the editor, "Polish text" improves her draft while keeping the existing layout.
After a request for a new task, the Assistant opens a filled-in "Add task" form. Nothing is created until you click "Save" yourself. Saving a reminder also requires confirmation. You can also request a Word, PDF, Excel, CSV or HTML file: the model prepares the content, and Task Manager puts the file together. What you can ask the AI Assistant to do remains part of the application, not a subscription with a particular provider.
The Assistant also reads text from TXT, CSV, MD, JSON, XML, HTML, DOCX, XLSX and PDF attachments with a text layer. Reading these formats doesn't require changing provider. For very long documents, check how you're providing the file: an attachment from a message is read up to about 50 KB of text, while a file set as a knowledge source is read up to 300,000 characters.
How to choose a model and provider without a made-up ranking?
If you need semantic search in the AI knowledge base in the Pro plan, choose OpenAI or Gemini; if you want indexing itself not to send content to the provider, check Claude or Grok. Then check which model is available on your own account. The list of providers alone doesn't prove that one model writes better than another; Task Manager fetches the current list of models once you enter a key and click refresh.
| Situation in your company | What to check |
|---|---|
| You have the Pro plan and want to search the knowledge base by meaning | Choose OpenAI or Gemini and check the semantic search setting. |
| You want indexing the knowledge base itself not to send its content to the provider | Check Claude or Grok; questions with context still go to the model. |
| The team wants to ask about photos and screenshots | Choose a model from the list that supports images. |
| The company already has an API account with one of the four providers | You can start with that one and check the available models. |
| You're still testing how you work with AI | Start with the model marked "(recommended)" and the default reply length limit. |
The "(recommended)" label is given to one model per provider: gpt-4o in OpenAI, gemini-2.0-flash in Gemini, claude-opus-4-8 in Claude and grok-3 in Grok. The other models from the fetched list also work - they simply don't have this label. Model names given in the instructions don't replace the current list fetched from your account.
Changing provider requires a new key and fetching models
You change the provider in the same place you chose it the first time: Settings → AI. The administrator selects a different item in the "AI provider" field, enters a new key, refreshes the list of models, chooses a model and clicks "Save AI settings". Task Manager then tests the connection and shows any warnings.
Tasks, Chats, Notes, Files and Mail are application data on your server - changing the model provider doesn't move them. The administrator decides, with a separate switch, whether the history of conversations with the AI Assistant is saved in the database; by default, saving is enabled. In the Pro plan, they can also use the "Update AI knowledge base" button, which processes tasks and chats without a summary or changed since the last update.
If the model doesn't respond after the change, check the message in the thread. The application points out, among other things, an invalid key, an exceeded query limit or an unavailable model. The message "Model does not support chat mode" means you need to go back to settings and choose a different model from the list.
No provider can read a scan without text
Changing provider doesn't add OCR, i.e. text recognition from an image, to Task Manager. A PDF that is a scan without a text layer won't be read by any of the four providers; the program also doesn't read old, binary DOC files.
If the company receives contracts as a PDF with a text layer or as DOCX, the Assistant can read their content. If you only get a scan, you need a text version or you have to put the key information in the task's content. Simply swapping Claude for Gemini won't change that.
For long files, the path the text takes to reach the Assistant matters. An attachment from a message is read up to about 50 KB of text, a file set as a knowledge source up to 300,000 characters - in chunks of about 24,000 - and the knowledge base index stores up to 54,000 characters from a single file. Above these limits, the AI only sees the start of the document. For a lengthy report, point to the specific file as a source and check whether the answer covers the part you need.
Application settings also keep an eye on the budget. The daily limit of AI queries from Automations is 200 per day by default (entering 0 disables the limit); once it's used up, AI recipes are skipped for the rest of the day with an entry in the history, and the counter resets at midnight. AI summaries of attachments default to "Manual" mode, so they aren't created automatically. This matters more for the bill than choosing a provider by name alone.
For tomorrow, prepare one question and your own key
Tomorrow, write down one question your team actually asks about their tasks. Choose a provider based on your team's needs, not on a ranking of models. For a company like Marek's agency, Task Manager keeps the one-off, perpetual, subscription-free, per-user licence separate from the bill for AI on your own key; the application's data stays on the company server. If you need semantic search in the knowledge base, check the Pro plan and OpenAI or Gemini. If not, you can start with any of the four supported providers.
Create an account with the provider you've chosen, generate a key, fetch the list of models in Settings → AI and, after saving, check the Assistant's reply. Then take a look at the billing on your provider account and decide whether this choice suits you. If you'd like to see the layout of the application first, open the Demo.
Frequently asked questions
Which AI provider should I choose for my company if I'm just starting out?
Start with the provider you already have a company API account with, or with the model marked "(recommended)". If you need semantic search in the Pro Company Knowledge base, choose OpenAI or Gemini.
Can I change the AI provider in Task Manager later?
Yes, the administrator changes the provider in Settings → AI, enters the new key, fetches the list of models and saves the settings. Tasks, chats, notes, files and mail remain application data on your own server.
Can the AI Assistant do the same with every provider?
Conversation, private questions in threads, text editing and file preparation are available with all four providers. Semantic search in the Company Knowledge base requires OpenAI or Gemini, and visual image analysis depends on the chosen model.
How is AI usage billed?
You pay for AI usage separately, directly to the provider on your own API key, according to their usage-based billing. Task Manager does not add a margin or include a token package in the licence; your bill depends, among other things, on the chosen model, the number of requests and the length of the responses.
Will the AI Assistant read a scanned PDF?
No, Task Manager has no OCR and cannot read a PDF that is a scan without a text layer. It reads PDFs with text, as well as, among others, DOCX and XLSX files; it does not read old DOC files.
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.
