Work-Day AI: The Current Offerings

Platforms

Claude Cowork

Anthropic. Ready with caveats. Ready for a small team that works from files and email and is willing to keep approvals on. Watch usage limits on long jobs and pick the plan tier accordingly.

Cowork handles tasks across connected folders, email, calendars, and the web, producing spreadsheets, slide decks, and formatted documents. Cloud execution is now the default; local execution remains available in existing desktop deployments. In September Anthropic began merging Cowork and chat into one Claude app, starting with Pro and Max plans, and on October 6 the setting that keeps new tasks only on your computer is removed for those plans. Scheduled cloud work continues with the laptop closed when its inputs and tools are available remotely. Access to files or browsers on your computer requires the desktop app to stay online.

What to like

What to plan for

ChatGPT Work

OpenAI. Ready with caveats. Ready if you already live in ChatGPT and want to work across connected apps. Check the connections you need, set usage limits, and keep consequential actions behind approval.

ChatGPT Work is OpenAI's agent for substantial tasks across files and connected apps. Web and mobile work runs in the cloud; the desktop app also supports local work. It can produce documents and presentations, use a browser, and run scheduled or supported app-triggered tasks. A broad plugin catalog makes it worth considering when your work spans several services, but check that the connection supports the actions you need. In late September OpenAI also launched dots, separate always-on agents with their own cloud computers; the first is included on Pro and Business Premium, and tasks a dot starts in ChatGPT Work count against the usual limits.

What to like

What to plan for

Grok Bot

xAI. Ready with caveats. Go for a small, supervised pilot on Cursor Teams. Keep each user’s bots within one trusted area of work; stronger organization controls require Enterprise. Build confidence before making it a daily dependency.

Grok Bot gives each user named bots that work on a persistent computer in Cursor's cloud. They can use connected apps, browse, create files, and coordinate on tasks while the laptop is closed. All of one user's bots share that user's files and browser sessions; other users have separate computers. Optional local commands extend its reach with approval. Teams can also publish a shared Team Bot that colleagues use from the app or Slack; each person's use counts against their own allowance.

What to like

What to plan for

Gemini for Workspace, Spark, and Workspace Studio

Google. Go for Workspace; wait for Spark. Go for document work and defined automations inside Workspace. Use Studio for repeatable steps such as inbox sorting and reports. Wait for Spark if you need its broader agent on a business account.

For a Workspace business, the useful starting point is Gemini in Docs, Sheets, and Slides, plus Workspace Studio for repeatable automation. Studio builds flows from triggers and steps: sort incoming mail, process a list, draft or send a reply, or update a document. Since mid-September it can also start from events in other apps, call outside services through webhooks, and connect to tools such as QuickBooks, HubSpot, Salesforce, and Slack in beta. Gemini Spark is a separate, broader agent that can work across apps and browsers, but still requires a personal Google account; it now has a Windows app as well as Mac.

What to like

What to plan for

Microsoft 365 Copilot and Copilot Cowork

Microsoft. Ready with caveats. Ready for Microsoft 365 shops that have someone to administer it. If you have no IT person, the enable-and-govern step is the real cost.

Copilot is the assistant built into Word, Excel, Outlook, and Teams; Copilot Cowork (generally available since June 2026) is the agent layer on top. Give it a long task and it works in a cloud sandbox across your email, meetings, and files, showing a plan with checkpoints so you approve changes before they are applied. It inherits your existing Microsoft 365 permissions, audit logs, and sensitivity labels, making it a good fit for organizations already administering Microsoft 365. Scheduled and supported email- or Teams-triggered tasks can run in the cloud, and Microsoft lets you choose among several underlying models, including Claude and GPT. In late September Microsoft announced a reorganized Copilot with Home (chat and Cowork together), Code, and Autopilot, rolling out first through its early-access program.

What to like

What to plan for

Muse

Meta. Wait. Wait. The connections suit a small business, but there is no business plan or published admin control yet, training on your conversations is on by default, and the product is weeks old. Try it on low-stakes tasks; keep business accounts out of it until Meta publishes team controls.

Muse is Meta's personal agent, launched in September in the US. It works in a virtual machine Meta runs for each user, keeps going in the background, and reaches your email, calendar, and apps through connectors, falling back to a browser where there is none. The Mac app can also work with files, mail, and calendar on your computer. On September 29 Meta added Muse for Small Business, with connections to Shopify, Stripe, QuickBooks, Canva, Slack, Zoom, Notion, Dropbox, and Asana, plus Instagram business accounts, Facebook Pages, and Meta ad accounts; you can also build your own connectors. A separate monitoring agent pauses high-risk actions for approval, and Meta says Muse will not post, send messages, or buy anything without the owner's approval.

What to like

What to plan for

What about running one yourself?

If you have heard of "local models" and wondered whether you could keep all of this in-house, the honest answer is: partly, and only if someone on your side enjoys the plumbing.

We run gpt-oss-20b, OpenAI's open-weight model released under the Apache 2.0 license, on a small Linux box in our office and on our own laptops. It can run in roughly 16 GB of memory, with additional headroom useful for longer tasks. There is no vendor fee per call; hardware and electricity still cost money. We configure these workflows to keep their data on our own systems. We use it for the work where that matters: nightly health analysis and weekly summaries of our own systems, semantic search across months of operational history, and first-pass estimates on internal project cards. Nothing customer-facing, nothing that needs the reasoning depth of the frontier models, and nothing we would want to explain to a client if it had left our network.

What you give up is everything the six cards above package for you: the connectors, the approval gates, the scheduler, the audit trail. You assemble those, or someone does it for you.

Our verdict. Wait, unless privacy is the constraint that decides the question, or you already have the person. For a specific, well-bounded workflow that must stay in-house, a local model is a good answer, and it is one we set up for clients. If you think local inference might be right for part of your workflow, we can help with that.

What changed since September

Until recently, "using AI at work" meant typing into a chat window and copying the answer somewhere else. The current generation is different. You give it a task ("reconcile these two spreadsheets and write me a summary," "watch my inbox and draft replies to vendor questions"), it opens your files and apps, works through the steps, asks permission at the risky moments, and hands back a finished spreadsheet, document, or email draft. This guide compares six major vendors’ agent offerings. They combine app connections, file handling, and automation in different ways; the useful question is whether the combination fits your work.

The differences that matter for a small business are not the model's IQ. They are where your data goes, how the bill behaves, how much can go wrong without you noticing, and whether the product is actually available on a business account today. Our recommendation this month is to start with a bounded workflow on the business platform that fits your existing tools. "Ready with caveats" means go with the stated conditions; "Wait" means do not depend on it for that job yet.

Since September, the changes have been mostly in packaging, defaults, and billing rather than in what the agents can do. Anthropic is merging Cowork into the main Claude app and making cloud the default for new tasks. OpenAI added a second, always-on agent called dots alongside ChatGPT Work. Grok Bot added a shared bot a whole team can use. Workspace Studio can now start from events in outside apps and call other services. Microsoft will turn on usage billing by default for new Copilot Business licenses in November and is rolling out data-loss prevention for Cowork. One new entrant: Meta launched Muse for Small Business on September 29, with connections to tools such as Shopify, Stripe, QuickBooks, and Slack. It is weeks old, so we rate it Wait; its profile is below.

Five questions before you adopt any of them

  1. Where does the work happen, and is that acceptable for this data? Cloud execution is convenient and lets tasks run overnight, but the data used by the task is processed on the vendor’s infrastructure. Local app access does not mean local AI processing. Decide per workflow, not per vendor.
  2. What is the bill going to look like in month three? Compare the seat price, included usage, and any extra usage charges. Run a few representative jobs and set spending limits before rollout; a low seat price does not establish the cost of an automated workday.
  3. What happens when it reads something malicious? An agent that reads your email and browses the web can be tricked by text hidden in a web page or a message ("prompt injection"). Every vendor here acknowledges the risk; none claims to have solved it. Keep approvals on for anything that sends, deletes, or pays.
  4. Is it actually available on a business account? Gemini Spark still requires a personal account; Workspace Studio supports business automation, and Grok Bot is available through Cursor Teams. Meta's Muse has a small-business edition but no business plan. Check the specific product and plan.
  5. Who administers it? Microsoft's and OpenAI's business tiers assume an admin who enables features, sets spend, and reviews audit logs. If that person is you, budget the time.

The ratings are recommendations for starting useful work with human review. They do not promise that every task can run unattended.

Comparison notes

Cloud execution can continue while you are away, but only with data and tools reachable from the cloud. A local file connection or browser may still need your computer awake. For any platform, confirm that it can read the source, produce the result, and put it where you need it in the same workflow.

Memory helps carry context forward; saved templates and instructions are the more dependable home for your procedures. Approval settings, memory controls, and audit access differ by plan, so the business tier matters as much as the product name.