Biwak compared: six AI agents for office work in 2026

Key takeaways

• Biwak leads and fully meets all seven criteria. Like Codex or Claude Code, it works directly on your computer, in the folder you share with it: it reads your files, analyses them and creates new files. The language model runs in the EU Data Zone of Microsoft Azure, and before every task Biwak creates a checkpoint, a saved state of the folder.
• Mistral Vibe, Langdock and Microsoft Copilot follow, level with one another. According to the vendors, all three process data in the EU as a rule, with exceptions for individual features or models, and they work with files in the browser, via upload or in cloud storage such as OneDrive. None of the three works as an office agent with files in a folder on your computer; at Mistral, only the coding variant Vibe Code can do that.
• ChatGPT Work and Claude also work on your computer, but without a commitment to the EU. Both edit files in shared folders in their desktop apps and are connected to Google and Microsoft services. For the plans we assessed, OpenAI and Anthropic do not commit to processing in the EU.
• Test with your own documents. Four of the six agents can be tried without a credit card. Give all of them the same task on copies of real files, then compare the files in the folder, not the answers in the window.

What an AI agent for office work is (and what it is not)

An AI agent for office work carries a task through to the finished file. You write in one sentence what should be there at the end, for example “Compare the three quotes in the folder and create a table with prices and delivery times”. The agent opens the files itself, calculates, checks its interim result and saves the result. The language model provides the understanding; the agent is the program around it that reads files, uses tools and asks for your approval before sensitive steps.

In this sense, an AI agent is not a chat window that you copy text into and then copy the answers back out of. The difference shows up with tasks involving several files: a chat answers questions about what you upload, while an agent works through the folder and leaves files behind for you to open and check. How this plays out with the same task is shown in the comparison ChatGPT answers, Biwak works.

An AI agent also does not replace specialist software such as DATEV (Germany's standard accounting software) or the beA mailbox (the electronic mailbox of German lawyers), and it does not make decisions you are liable for. It prepares what a person then checks and approves. So when choosing one, it is not only what an agent can do that matters, but also what it is allowed to do: which folders it can see, where the language model processes data, and how you can undo a change that does not fit. The overview What is an AI agent? covers use cases by department.

How we ranked the AI agents

We selected six agents that handle office tasks with files and can be bought by businesses in Germany: three from Europe and three from the USA. We measured each one against seven criteria that you can check on the vendors' websites. Each criterion earns a full tick, a half tick or none; for the order, a full tick counts as two points and a half tick as one. The details come from the vendors' websites, retrieved on 14 October 2026; where a page was blocked, we read an archived copy, whose date is given with the source.

In each case we assessed the cheapest paid plan that includes the agent and is meant for businesses with a data processing agreement (DPA). For Biwak, that is Base Camp. For Langdock and ChatGPT, it is Business; for Mistral Vibe and Claude, the Team plan; for Microsoft Copilot, Business together with the cheapest Microsoft 365 plan it requires. Individual plans such as ChatGPT Plus or Claude Pro are cheaper, but they belong to the offers for private customers; OpenAI and Anthropic only name a DPA for their business plans.

Prices are taken at monthly payment and excluding VAT. We count dollar amounts one to one as euros; at exchange rates between 0.81 and 1.19 euros per dollar, this does not change a single tick. If a plan requires another one, as Copilot Business requires a Microsoft 365 plan, we assess price and support for the cheapest package of both. If two agents have the same rating, processing location decides first, then undo, working with files, price and the free trial; if these are also equal, the lower monthly price decides.

Whether training on your content takes place is shown in the comparison as a separate column without ticks, because in all the plans assessed it is either excluded or can be switched off. We do not rate integrations with services such as Microsoft 365 or Google; Biwak has no such integrations, and the ones the other agents offer are listed in their entries. We mention ratings on OMR Reviews and G2 for context, but they earn no ticks; OMR marks reviews there for which the reviewer received something in return, such as a voucher. We did not assess the quality of the results: we did not run the other vendors' agents on the same tasks, and a ranking does not replace your own test.

The six AI agents in detail

The order is based on the ratings alone; in a tie, the rule from the previous section applies. Each entry gives the plan assessed with its price, the strengths and limits as stated by the vendor, and who it suits best. All ticks per criterion are shown in the side-by-side comparison below.

1. Biwak, best for office tasks with files in your own folder

Biwak is a desktop app for macOS and Windows with an AI agent for office tasks, developed by a sole trader in Munich. You choose a working folder and describe what should be there at the end. The agent reads PDFs with a text layer, Word, Excel, PowerPoint and CSV, and saves the result as a new file next to your files. Before every task, Biwak backs up the folder; with “Undo” and a confirmation, the earlier state is back. Files outside the folder that are changed by a command you have approved are not covered by this backup.

The language model processes data in the EU Data Zone of Microsoft Azure, tasks go through our relay in Frankfurt am Main, and from version 0.2.32, web search runs on Amazon Web Services in Ireland. Your content is not used for training. So that we can help when something goes wrong, we keep every task from the app in encrypted form until 30 days after the conversation's last task.

The subscription costs €19 a month for one person or €99 for the whole team, cancellable monthly. Biwak charges no VAT as a small business under § 19 UStG (German small-business VAT exemption). The plan assessed is Base Camp, with 1,500 credits a month; there, a person replies by email on the next working day. For testing, you get a one-off allowance of 100 credits, without payment details, and the DPA applies with the subscription, without a signature. Biwak fully meets all seven criteria. For law firms, tax practices and medical practices: client and patient data should only go in after your own review, because Biwak's model providers have not committed to confidentiality under § 203 StGB (the German criminal provision on professional secrecy). We have not yet published case studies or customer testimonials.

1:01

Best suited for small offices and SMEs that want to hand over recurring tasks involving PDF, Word and Excel files, and that need the model processing in the EU. If you mainly want your agent to work in Microsoft 365 or Google Workspace, you are better off with one of the vendors that has ready-made integrations.

2. Mistral Vibe, best for a low-cost start with European models

Mistral AI is based in Paris and develops its own language models. Since 12 August 2026, its assistant has been called Vibe instead of Le Chat. Vibe Work completes tasks across several steps and services in the browser and in the iOS and Android apps. Only Vibe Code works with files on your computer. It is the version for programmers that runs in the command line and in VS Code, and it asks for approval before writing files and before running commands, which can be configured per agent. We did not find a desktop app for macOS or Windows on the pages we checked.

Team costs €24.99 per user per month excluding VAT, €29.74 including tax in Germany, and applies from two users; Pro for individuals costs €14.99 net. The free plan includes Vibe with limits and, according to the pricing page, more than 100 connectors, which are marked as beta; according to the documentation, these include Gmail, Google Calendar and Google Drive. Data is kept in the EU by default, but depending on the feature, it may be processed temporarily outside the EU. Check the setting for training: it is on by default; on the Team plan, the administrator can switch it off for the whole organisation, and only in Enterprise is it off from the start. The data processing agreement (DPA) in the version of 27 July 2026 is public, and in September 2026 Mistral raised €3 billion in capital. On an equal rating, Vibe ranks ahead of ChatGPT Work because of processing location, ahead of Microsoft Copilot because it can be tried without a credit card, and, with otherwise identical ticks, ahead of Langdock because of its lower monthly price.

Best suited for individuals and small teams that want to work affordably with models from a European model developer and mainly hand over tasks in the browser or on a smartphone. According to Mistral, Vibe is not intended for working in folders on your computer outside programming.

3. Langdock, best for many models under one administration

Langdock, based in Berlin, brings together models from Anthropic, Google, Meta, Mistral, OpenAI and DeepSeek in one interface with chat, agents and workflows, in the browser and as a desktop app for macOS and Windows. Files come in through uploads or synced folders from SharePoint, Google Drive, OneDrive and Confluence. Langdock does not describe working in a folder on your computer on the pages we checked. The application runs on Microsoft Azure in the EU, and so do most models according to the vendor, depending on the setting.

Business costs €29 per user per month with models included when paid monthly and €23.20 when paid annually, or €22 and €17.60 respectively without models, each plus VAT. The trial runs for seven days without a credit card, with €5 of model credit. For agents and workflows there are 57 integrations with 754 actions, plus connections via MCP, an open standard for such links. According to Langdock, training never takes place, and the data processing agreement (DPA) is on the security page. Langdock cites more than 10,000 customers and, for August 2026, annual recurring revenue of $50 million. Langdock does not mention approval before changes, but documents in the Document Editor have a version history: every change, including those made by the AI, can be viewed and reset to an earlier state. That earns a half tick for undo; Langdock ranks behind Mistral Vibe only because of its higher monthly price.

Best suited for businesses that want to provide models from several vendors centrally, with agents and workflows, and need sign-in via SSO. If agents need to edit files in a folder on your computer, a desktop agent is the better choice.

4. Microsoft Copilot, best for businesses that work in Microsoft 365

According to Microsoft's privacy page of 18 August 2026, Microsoft 365 Copilot is now called Microsoft Copilot. In the paid version, it works with its own agents for creating documents in Word, Excel and PowerPoint, and Copilot Cowork carries out tasks on your behalf: it edits Office files in OneDrive and SharePoint, sends emails and schedules meetings; before important actions such as sending an email, it asks for your approval. For this, Microsoft does not mention files on your computer itself. For EU customers the EU Data Boundary applies; according to Microsoft, models that Anthropic supplies as a sub-processor are currently excluded from it.

Copilot Business is an add-on to a Microsoft 365 Business plan and is for up to 300 users. On its own, it costs €21.84 per user per month in Germany when paid monthly, €18.20 with an annual commitment, and €15.60 on offer until 31 December 2026. In a package with Microsoft 365 Business Standard, Microsoft lists €24.43, or €21.10 in the version without Teams, each excluding VAT. The prices come from an archived copy of Microsoft's page from 29 September 2026, because the page was blocked when we retrieved it. Copilot Business can be tried for one month, with a credit card. Phone and web support at any time in the packages and the privacy addendum in the product terms earn full ticks there; more than 100 connectors are available. On 29 July 2026, Microsoft reported more than 30 million paid Copilot seats. On an equal rating, Copilot ranks behind Mistral Vibe and Langdock, which can be tried without a credit card, and ahead of ChatGPT Work because of processing location.

Best suited for businesses whose documents are already in OneDrive and SharePoint and that want the agent where work already happens: in Word, Excel and PowerPoint. If you do not have a Microsoft 365 Business plan, or want to edit files on your computer, one of the desktop agents is cheaper.

5. ChatGPT Work, best for teams that already use ChatGPT

ChatGPT Work is OpenAI's agent for longer tasks with several steps and finished results. In the desktop app, it opens a local folder or a project that you give it access to. Reading, changing and running commands all follow the approval settings of the task. In the browser and on a smartphone, Work does not access files on your computer, but you can start tasks on your computer and continue them on the go. According to OpenAI, even when working locally, messages and task context may be stored in the cloud, and file excerpts may be sent to OpenAI.

Business costs $25 per user per month when paid monthly, and $20 when paid annually, from two users. You can try Work without a credit card on the free plan, where it is limited and only available on your computer. Integrations include Google Workspace, Slack, GitHub and Microsoft 365, among others; tasks can also be started by new Gmail messages or Slack posts. OpenAI names EU data residency only for Enterprise and Edu, not for Business, so Business gets no tick for processing location. Business data is not used for training by default, and the data processing agreement (DPA) is public. ChatGPT as a whole scores 4.7 out of 5 from 80 reviews on OMR Reviews and 4.6 from 3,008 on G2. On an equal rating, Work ranks behind Mistral Vibe, Langdock and Microsoft Copilot because of processing location.

Best suited for teams that already use ChatGPT and want Work on their computer, in the browser and on a smartphone, with integrations to their services. If processing in the EU is required, check Enterprise or one of the European vendors.

6. Claude, best for tasks that keep running in the cloud

Claude comes from Anthropic, based in San Francisco; the contracting party in the EU is Anthropic Ireland. Since 16 September 2026, Anthropic has been merging the chat and the Cowork agent mode, first in Pro and Max; according to Anthropic, Team will follow soon. In the desktop app for macOS and Windows, Cowork reads and writes files in folders you connect; depending on the mode you choose, it asks for approval before actions, and it always asks before deleting files for good. On the Team plan, tasks also keep running in the cloud when your laptop lid is closed; this is currently in beta. In that case, local files that are open are also processed on Anthropic's servers.

A Standard seat on the Team plan costs $25 a month when paid monthly and $20 when paid annually, for 2 to 150 seats; Premium seats with more usage cost $125 and $100 respectively. The agent is only available on paid plans; the free plan includes the chat. Integrations include Google Drive, Gmail, Google Calendar, GitHub, Microsoft 365 and Slack, and scheduled tasks run in the cloud without any device having to be switched on. According to Anthropic, data is stored in the USA; by default, Anthropic may route requests to the USA and to countries in Europe, Asia and Australia. Support is available in writing via messenger and email; according to Anthropic, it does not offer phone support. The data processing agreement (DPA) applies with the commercial terms, and Team data is not used for training by default. Claude is rated 4.3 out of 5 from 9 reviews on OMR Reviews. Claude ranks behind ChatGPT Work because the agent can only be tried on a paid plan.

Best suited for teams that want tasks to keep running in the cloud and want to open the same session on their computer, in the browser or on a smartphone. For the Team plan, Anthropic does not commit to processing in the EU.

All six AI agents side by side

The first table sets the plans assessed side by side, using the vendors' details, and the second shows the ticks per criterion. Prices apply to monthly payment, excluding VAT.

Which AI agent fits your use case (and when another one is better)

The ranking counts all seven criteria equally. For a particular task, a single criterion often counts for more than the overall rating, and in that case another agent is the better fit.

Results and case studies: what the vendors document

We have not found any independent measurements that compare all six agents on the same tasks. What does exist are figures and case studies from the vendors themselves. They show how widespread a tool is and who uses it, not how well it solves your task.

30 million

paid seats for Microsoft 365 Copilot, according to Microsoft, on 29 July 2026

98%

Use of Microsoft 365 Copilot at KOHLER, according to Microsoft's case study

10 000+

Customers that Langdock names in its press area

• KOHLER at Microsoft: rollout of Copilot with its own programme to support employees; more customer stories come from DLA Piper and Globo.
• Customer stories from Mistral: including CMA CGM and Stellantis; according to the homepage, ASML finds faults in subsystems 70% faster.
• Langdock customer page: logos such as Merck, Personio and SumUp, and a report on the rollout at HERO Software.
• Customer stories from Anthropic: 284 stories, including a conversation with HubSpot about using Cowork in marketing.

1. Choose a task

   Pick a task that comes up every week and whose result you can check in five minutes, for example a table built from ten PDF invoices.

2. Measure the time by hand

   Write down how long the task takes you today, including your check at the end.

3. Give every agent the same task

   Put copies in a folder, redact names and account numbers, and give each agent the same task word for word.

4. Count the rework

   Count the errors and the minutes until the finished result. The article Introducing AI in SMEs gives the framework for a trial over several weeks.

Frequently asked questions