Blackfort Technology
AI agents

A vendor-neutral overview for companies

AI agents for business: Grok Bot, ChatGPT, Claude and Copilot compared

Agents that operate browsers and applications on behalf of your team are now offered by xAI, OpenAI, Anthropic, Microsoft and others, and Google has announced an agent of its own. They differ in where they work, how they sign in and what their business editions control. We support the rollout, independent of the vendor.

Book an initial callSee the comparison

Four types of agent, four kinds of risk

A chatbot answers, an agent acts: it opens websites, fills in forms and works through several steps, often on a schedule. For a rollout, what matters most is where it does that work. It determines where your sign-ins are kept and which rules apply.

Cloud computer at the vendor

The agent runs in a cloud environment operated by the vendor or its service provider and keeps going when your laptop is closed. Browser sessions for your applications may be stored there, connected-app access on separate vendor systems.

Examples: Grok Bot, ChatGPT Work in the cloud, Perplexity Computer, Manus

Who operates the computer, which country is it in, and who can get at the sessions stored there?

Browser on your computer

The agent controls a browser on the employee's device, with the sign-ins that already exist there or with sessions of its own. Websites see traffic from your device.

Examples: Claude in Chrome and the Claude Cowork browser, Perplexity Comet, Manus Browser Operator

Which websites may the agent use, and which sign-ins stay excluded?

Agent on your own computer

The agent operates programs and files directly on the PC, via mouse and keyboard, commands or shared folders.

Examples: ChatGPT Work in the desktop app, computer use in Claude (beta, currently Pro and Max only), local commands from Grok Bot via the desktop app

Which folders and programs do you share, and what still ends up in the cloud?

Central automation platform

Agents are built and run centrally, with stored credentials, allow lists and billing per step.

Examples: Microsoft Copilot Studio with computer use

Who builds the agents, where are the credentials kept, and who checks what they do?

Seven AI agents compared

The information comes from the vendors’ public documentation, as of 30 October 2026. We do not rate which product is better. The overview shows which questions arise with which agent. Where a vendor does not document something, we say so.

Two people at a meeting table pointing at printed comparison tables
Start with the workflow, then choose the tool. (Illustrative image, AI-generated)

Grok Bot

xAI, run in Cursor's cloud

Cloud computer at the vendor

Where it works
One persistent computer with browser, files and command line per account, shared by all Bots of the account. Can also learn workflows by demonstration ("Teach a task").
Signing in to websites
With logins from the Vault, from 1Password (personal or Family accounts, own vault only) or through a form a person fills in. For single sign-on or passkeys a person takes over. The session stays on the cloud computer and is available to every Bot on the account.
For companies
On Cursor Enterprise, an admin enables Grok Bot. Network controls, audit logs, SCIM and checking model requests on your own server are only available there. Connected apps fall under the network controls only with an additional switch, and the model request checks start in shadow mode.
Data location according to the vendor
The computers run in the United States. Enterprise customers can restrict inference, processing and storage to the US through a Cursor program, at a surcharge and with exceptions. Data residency in the EU is not documented.

Source: docs.x.ai, as of 30 October 2026Grok Bot in your company →

ChatGPT Work

OpenAI

Cloud computer at the vendor or agent on your own computer

Where it works
On the web and mobile in the cloud, with its own browser on a separate computer. In the desktop app optionally local; messages and context may still be stored in the cloud.
Signing in to websites
Through a dedicated sign-in form. According to OpenAI, credentials go directly to the cloud browser and are not visible to the model; the session persists until it expires.
For companies
In eligible Enterprise and Edu workspaces, Work is on by default for the default role until an admin turns it off. Cloud and local execution can be controlled separately per role. By default, the agent asks before visiting a new website.
Data location according to the vendor
Data residency in Europe is available for new Enterprise and Edu workspaces. Work and the cloud browser are not on the list of covered features, and according to OpenAI new features are only covered once they are explicitly listed.

Source: help.openai.com, as of 30 October 2026

Claude Cowork

Anthropic

Cloud session at the vendor, browser on your computer

Where it works
Cloud sessions run on Anthropic's servers; local sessions remain available for existing desktop deployments. The built-in browser and Claude in Chrome run on the employee's computer; the built-in browser needs the desktop app open.
Signing in to websites
Claude in Chrome uses the sign-ins in the employee's Chrome. The built-in browser has its own sessions: users sign in there or import sign-ins through cookies, where banking, email and single sign-on sites are unchecked by default.
For companies
Cowork itself is on by default on Team and Enterprise. Cloud execution is on by default on the Team plan (in beta) and off on Enterprise until an owner enables it; HIPAA configurations have exceptions. Admins control network access and allowed websites separately, and the network rules do not apply to web search, web fetch or MCP connections, including Claude in Chrome. Computer use is currently not available for Team and Enterprise.
Data location according to the vendor
Not documented for Cowork on claude.ai. With Claude Desktop through Amazon Bedrock or Google Cloud in an EU region, requests go to that endpoint and the history stays on the device.

Source: support.claude.com, claude.com/docs, as of 30 October 2026

Copilot Studio with computer use

Microsoft

Central automation platform

Where it works
Three options: a Microsoft-managed browser, a Cloud PC pool via Windows 365 (Entra joined and Intune enrolled) or your own machine. The first two are previews that Microsoft does not intend for production use.
Signing in to websites
On the machine, the agent uses the maker's credentials by default, optionally those of the end users. Microsoft warns that if the agent is shared, others act with the maker's access. Website passwords are stored separately in Power Platform or an Azure Key Vault.
For companies
By default the agent may operate any website and application. An optional allow list prevents actions outside the list but not opening other sites. Billing is per step in Copilot Credits, and the Cloud PC pool is additionally metered through Azure.
Data location according to the vendor
The Cloud PC pool is hosted in the same geography as the Power Platform environment. No location is documented for the managed browser.

Source: learn.microsoft.com, as of 30 October 2026

Gemini agent

Google Cloud

Cloud agent at the vendor

Where it works
Announced on 8 October 2026. According to Google it runs in the cloud in a sandboxed environment and keeps working after you close your laptop.
Signing in to websites
Whether and how the agent operates a browser is not documented. For external systems, identity is passed on through standards such as OAuth.
For companies
A dedicated identity for the agent, role-based permissions and real-time spend caps.
Data location according to the vendor
Not documented for the agent, nor is its availability status.

Source: Google Cloud Blog, as of 30 October 2026

Comet and Computer

Perplexity

Browser on your computer (Comet), cloud agent at the vendor (Computer)

Where it works
Comet is a separate AI browser for macOS and Windows. With Computer, each task runs in its own isolated container with a browser.
Signing in to websites
Comet uses the sessions in its own browser. How Computer signs in to websites is not documented.
For companies
Comet is deployed via MDM; admins decide on which sites the assistant may control the browser. Computer is enabled by default on Enterprise Pro and Max.
Data location according to the vendor
Not documented. SOC 2 Type II and no training on enterprise data are documented.

Source: perplexity.ai, as of 30 October 2026

Manus

Butterfly Effect Pte. Ltd., Singapore

Cloud agent at the vendor, plus a browser on your computer

Where it works
The agent works in isolated virtual machines. The Browser Operator works in the local browser.
Signing in to websites
The Browser Operator uses existing sessions and sign-ins; according to the privacy policy, the data collected this way is sent to Manus' servers.
For companies
Team plan with admin dashboard, access control and audit logs; SSO is included from 30 seats and available at extra cost below that. Enterprise adds enforced SSO and network controls for agents.
Data location according to the vendor
Based in Singapore, transfers to the United States possible. No commitment on storage location is documented.

Source: manus.im, as of 30 October 2026

Five questions that apply to every agent

Whichever vendor you choose: these points should be settled before an agent works with real data and real accounts.

Contract and data location

Who is the contracting party, is there a data processing agreement, in which country is work done and data stored, and is your data used for training?

Sign-ins

Which accounts does the agent use, where do the sessions stay, who else can use them, and how are they ended again?

Permissions and approvals

What may the agent do without asking, which websites and applications are blocked, and who approves payments, outgoing emails or deletions?

Recording and co-determination

Which logs, screenshots or recordings are created, how long are they kept, and is the works council involved where there is one?

Use without rules

Who already uses an agent today, with a private or a business account? Some features are switched on by default in business plans.

Settle these questions for your company

How we support the rollout

From the first conversation into operation. Each phase can be commissioned on its own, and you decide after each one whether to continue.

  1. Orientation

    Which workflows should agents take over, which tools are already in place, who is already using something?

  2. Vendors and security

    Contract documents, data flows and admin controls of the vendors under consideration, together with your data protection officer.

  3. Rules and co-determination

    Account and permission concept, technical description for the works agreement, a short usage policy for the team.

  4. Pilot

    One non-critical workflow with test data, logging and defined stop criteria, followed by a review.

  5. Rollout and training

    Enabling further teams and workflows, training on your own workflow, settings documented.

  6. Ongoing support

    Reviewing permissions and logs regularly, assessing vendor changes, keeping the policy up to date.

What we do not do

We are not a law firm and do not provide legal advice. We deliver the technical and organisational basis on which your legal counsel, your data protection officer and your works council decide. We do not sell licences of the vendors named.

In detail

Grok Bot in your company

Teach a task, the shared cloud computer and what the Enterprise edition controls: data protection, works council and credentials with Grok Bot.

Grok Bot in your company →

Background

Governance for AI agents

Permissions, approvals, logging and protection against planted instructions: the basics for every agent.

Read the guide →

Rules for AI

AI compliance and governance

Policy, roles and evidence for using AI across the company, with reference to ISO 27001 and the AI Act.

AI compliance and governance →

Frequently asked questions about AI agents for business

What is the difference between an AI agent and a chatbot?

A chatbot answers, an agent acts. It opens websites, clicks, fills in forms, creates files and works through several steps in a row, often on a schedule and while nobody is watching. For tasks in your applications it needs access to them. Accounts, permissions and logs matter for chatbots too; with agents, actions it carries out itself and their effects on other systems are added.

Is there an alternative to Grok Bot with data residency in the EU?

For none of the cloud agents compared here does the vendor documentation establish a complete EU commitment covering execution, model processing and storage (as of 30 October 2026). There are, however, options with a region you choose: Claude Desktop with Cowork through Amazon Bedrock or Google Cloud in an EU region, with the history staying on the device, or the Copilot Studio Cloud PC pool in the geography of your Power Platform environment, currently a preview. For the Cloud PC pool this only establishes where the computer is; where the model runs is not documented for it. Which option fits depends on your workflows and data.

What happened to ChatGPT agent?

According to OpenAI, ChatGPT agent is no longer available. For longer, multi-step tasks OpenAI points to ChatGPT Work, which works with its own cloud browser on the web. If you wrote policies or approvals for ChatGPT agent, carry them over to Work and take the new switches for cloud and local execution into account.

Which AI agent is best for companies?

There is no general answer, and we do not publish a ranking. What matters are the workflow, the data involved and the applications the agent has to sign in to. That determines which type of agent is an option and which business edition comes with the contracts and admin controls you need. Often a classic automation without an agent is enough.

How can we tell whether our team already uses agents?

Personal accounts outside your company workspace are not reliably visible in its admin settings. Indications come from asking the team and from device, network and identity controls. In business plans, the admin settings help: according to OpenAI, ChatGPT Work is on by default for the default role in eligible Enterprise workspaces, according to Anthropic, Cowork and cloud execution are on by default on the Team plan unless a HIPAA configuration applies, according to Perplexity, Computer is enabled by default on its Enterprise plans, and according to xAI, every member of a self-serve Cursor Teams plan gets Grok Bot. Browser extensions and AI browsers on devices add to this. An inventory therefore belongs at the start of every rollout.

Do we need a separate review for every agent?

The questions are the same for every agent: contract and data location, sign-ins, permissions, recording and co-determination. The answers differ per product and change with every new version. A policy you have worked out once can be carried over to other agents, but each vendor's statements have to be checked individually.

As of 30 October 2026. The product names mentioned are trademarks of their vendors. Blackfort Technology is not affiliated with any of the companies named and does not sell their licences. The information is based on the vendors’ public documentation; many of the products are new or in preview and change frequently. It does not replace a review of the current contract terms. This page is not legal advice. The photos on this page are AI-generated illustrative images.

Before the first agent works with real accounts

Tell us which workflows an agent should take over and which tools are under discussion. In an initial call, we clarify what needs to be settled first.

Book an initial call+49 228 29978061