This guide compares products from their official documentation, pricing pages and published comparisons as of November 2026. It is not a hands-on test of each tool, so capabilities and prices are attributed to their sources, and you should confirm current plan details before buying.
AI desktop agents are finally good enough to take real work off your plate, but only the right work. ChatGPT Work, Claude’s desktop agent and Microsoft’s Copilot Cowork can all now open your files, build documents and run multi-step jobs while you do something else. The mistake most people make is starting with the most impressive task. Start with the most boring one.
My rule after reading every plan and permission page: hand off work that is tedious, reversible and easy to check. Keep anything that sends, pays or deletes in your own hands until you have watched the agent get the boring stuff right for a few weeks. Here are the five tasks I would delegate first, which tool fits each one, and where the real limits sit.
Table of Contents
- What AI Desktop Agents Actually Do
- The Three AI Desktop Agents Worth Knowing
- AI Desktop Agents Pricing Compared
- Task 1: Clean Up and Organise Folders
- Task 2: Turn Messy Inputs Into Finished Documents
- Task 3: Recurring Reports on a Schedule
- Task 4: Research Packs From Many Sources
- Task 5: Inbox and Calendar Triage, With Approval
- How to Brief AI Desktop Agents
- What Never to Give AI Desktop Agents
- Which of the AI Desktop Agents to Choose
- FAQ About AI Desktop Agents
- Verdict
- Sources and Further Reading

What AI Desktop Agents Actually Do
A chatbot answers. An agent acts. AI desktop agents sit on your computer or connect to it, read the files and apps you allow, and carry out a sequence of steps towards a goal: open these spreadsheets, combine them, build a summary deck, save it here. You describe the outcome, approve what needs approving and come back to a result.
The category grew out of coding tools. TechTarget notes that OpenAI’s desktop agent grew out of Codex, its command-line coding agent, and usage took off once it was built into the desktop app to manage several agents at once. The same pattern, an agent with tools and a workspace, now applies to spreadsheets, documents and email.
Where the work runs matters more than it first appears. Some agents run in the vendor’s cloud and reach your machine only through an open desktop app. Others stay inside a corporate tenant. That decides what they can touch, whether work continues when your laptop is closed and what your IT team will allow.
The Three AI Desktop Agents Worth Knowing
ChatGPT Work launched on 9 July 2026 alongside a new ChatGPT desktop app for Mac and Windows. It works with your files, plugins and approved tools, and creates documents, presentations, spreadsheets and PDFs that the app previews beside the chat. On desktop it can use local files, apps and the browser, and you can start a task on an approved folder and follow it from your phone. It asks before visiting a new website.
Claude’s desktop agent, launched as Cowork in research preview on 12 January 2026, was merged into the main Claude app on 16 September and reached Enterprise customers on 23 October. It creates Excel files with working formulas, PowerPoint, Word and Claude’s own slide format, and connects to Gmail, Calendar, Drive and Microsoft 365. Work runs in Anthropic’s cloud, so it continues with the laptop closed, while local folders are reached through the open desktop app on macOS, Windows and, in beta, Linux.
Copilot Cowork from Microsoft was announced on 9 March and became generally available on 16 June 2026. It lives inside Microsoft 365: Copilot in the browser, Outlook, Teams, the desktop app and mobile. It sends email, schedules meetings, posts in Teams and edits Word, Excel and PowerPoint files in OneDrive and SharePoint. It does not document local folder access or Google Workspace support.
There are others. OpenAI’s Codex desktop app suits developers who want parallel coding threads, and ServiceNow previewed its Project Arc desktop agent in May. For most knowledge workers, though, the choice is between these three AI desktop agents.
AI Desktop Agents Pricing Compared
Here are US list prices as published on 1 November 2026, compiled by Deckary.
| Agent | Plans | Notes |
|---|---|---|
| ChatGPT Work | Free (limited, desktop only), Go $8, Plus $20, Pro from $100, Business $20 per user annual or $25 monthly | Long tasks draw on shared Work and Codex credits |
| Claude | Pro $20 monthly ($17 annual), Max $100, Ultra $200, Team $20 annual or $25 monthly per seat, Enterprise by seat plus usage | Paid plans only |
| Copilot Cowork | Needs a Microsoft 365 Copilot licence: $30 per user enterprise, Business add-on from $18 promotional | Usage billed at $0.01 per Copilot Credit |
At the individual level, $20 a month gets you a capable agent from either OpenAI or Anthropic. In a Microsoft shop, Copilot Cowork’s price is mostly the Copilot licence you may already be paying for, plus credits that are hard to forecast because Microsoft publishes no per-task rate.
Task 1: Clean Up and Organise Folders
This is the perfect first job for AI desktop agents. It is tedious, it is reversible if you work on a copy and you can check the result in a minute. “Sort these 400 downloads into folders by project, rename the invoices to date and supplier, and list any duplicates” is the kind of request that takes a person an afternoon and an agent a few minutes.
Claude and ChatGPT Work both handle local folders through their desktop apps. Copilot Cowork works on OneDrive and SharePoint instead. Whichever you use, point it at a copy first, not your only archive, and ask for a list of proposed changes before it moves anything.
The best prompt pattern here is two-step: “Show me the plan as a table, then wait.” You review the table, fix the two rules it got wrong and only then let it run.
Task 2: Turn Messy Inputs Into Finished Documents
The second job is turning raw material into a finished file. Meeting notes into a tidy memo. Three exported CSVs into one spreadsheet with working totals. A rambling brief into a ten-slide deck.
This plays to what AI desktop agents do well: reading lots of input, structuring it and producing an editable output. Claude’s agent builds Excel files with real formulas rather than pasted numbers, which matters if anyone will update the sheet later. ChatGPT Work lets you annotate part of a file to ask for a focused revision, which is handy for iterating on a draft.
Check the numbers. Agents are good at structure and occasionally careless with arithmetic or with which column they summed. Open the spreadsheet and click into two or three totals before you send it to anyone.
Task 3: Recurring Reports on a Schedule
Once a one-off task works, make it recurring. A Monday summary of last week’s sales export, a monthly roll-up of expenses, a Friday list of open tickets older than seven days. These are the tasks where AI desktop agents start paying for themselves, because the effort to set one up is spent once and the time saved repeats every week.
Claude supports scheduled tasks, and because its work runs in Anthropic’s cloud, a scheduled job can run while your laptop is shut, as long as it does not need a local folder. Keep the first scheduled tasks read-only: they gather and report, they do not send or change anything.
If your recurring work spans several apps and needs strict, repeatable steps, a workflow tool may still be better than an agent. Our guide to n8n AI agents explains when deterministic automation beats a model making choices.
Task 4: Research Packs From Many Sources
The fourth job is research: “Pull the pricing pages for these six competitors, put plans and prices in a table, and note anything that changed since the version in this folder.” AI desktop agents are good at the gathering and comparing, and the output is easy to verify because every claim should point to a source.
ChatGPT Work can browse on desktop and asks before visiting a new site. Claude can use connectors and browser tools. Either way, ask for links next to every figure. An agent that cannot tell you where a number came from has probably guessed it.
This is also the task where prompt injection becomes a real risk. A web page can contain hidden instructions aimed at whatever agent reads it. Keep research tasks separate from tasks with write access to email or files, and read our guide to prompt injection and agent security before you combine them.
Task 5: Inbox and Calendar Triage, With Approval
Only after the first four are working would I let an agent near email. Start with triage: “Label everything from the last three days as needs reply, FYI or ignore, and draft replies for the first group.” The agent drafts. You send.
Copilot Cowork is strongest here if you live in Outlook and Teams, and Microsoft shows an approval dialog before it sends, posts or schedules anything. Claude’s Microsoft 365 connector is read-only until an Entra administrator approves write tools, which is a sensible default. Claude also offers three permission modes: Manual, which asks before actions and is the Enterprise default; Auto, where Claude reviews each action and blocks unsafe ones; and Skip.
Stay on Manual for anything that sends on your behalf. Drafting saves most of the time anyway. The last click is cheap, and it is the click that prevents the embarrassing email.
How to Brief AI Desktop Agents
Most disappointing results come from vague briefs. An agent cannot ask the questions a colleague would, or at least it often does not, so the brief has to carry that information up front.
A good brief has four parts. The outcome: what the finished file or folder should look like. The inputs: exactly which files or sources to use, and which to ignore. The rules: naming conventions, the currency, the date format, what counts as a duplicate. And the stopping point: “show me the plan before changing anything” or “save the draft, do not send it”.
Write the brief once, test it, then save it. Every one of these AI desktop agents supports some form of reusable instruction, whether that is a project, a skill or a saved prompt. A brief you have refined over three runs is worth far more than a clever one-off request, and it turns a personal trick into something your whole team can use.
Finally, ask for a short log at the end of each run: what the agent did, what it skipped and anything it was unsure about. That one line in the brief makes checking the work far quicker.
What Never to Give AI Desktop Agents
Some things should stay with you regardless of which agent you use.
Payments and purchases. Visa’s new agentic commerce benchmark found that models failed 42.8% of realistic shopping runs, and over half of those failures still ended in a purchase. Research with an agent. Pay yourself.
Passwords and primary logins. United Airlines now blocks unauthorised AI agents from booking its flights, and Bank of America has warned customers not to share passwords with assistants. Use the connectors and permissions the agent provides, never your raw credentials.
Irreversible deletes. Let an agent propose what to delete, then do it yourself, or have it move files to a holding folder you empty later.
Anything regulated. Check data retention terms first. Deckary notes that ChatGPT Work’s local access with Work Cloud does not provide strict zero data retention, and compliance logs are kept for 30 days. Claude removes deleted tasks from history immediately and from its backend within 30 days. Those details matter for client and patient data.
Which of the AI Desktop Agents to Choose
| If you are | Start with | Why | Link |
|---|---|---|---|
| Living in Outlook, Teams and SharePoint | Copilot Cowork | Native Microsoft 365 data, approval before sending, stays in your tenant | Microsoft 365 Copilot |
| Working with local files and spreadsheets | Claude | Real Excel formulas, cloud runs with laptop closed, fine-grained approval modes | Claude |
| Already paying for ChatGPT | ChatGPT Work | Included across plans, phone follow-up, wide plugin set | ChatGPT Work |
| A developer | Codex desktop or a CLI agent | Parallel coding threads and worktrees | AI coding assistants |
| Wanting everything on your own machine | A self-hosted agent | No cloud, full control, more setup | Local AI agent guide |
If you are torn, pick the one that matches where your files already live. AI desktop agents are only as useful as what they can reach, and moving your work to suit an agent is rarely worth it.
Related on this site: for more context, see Muse Glimmer and Siri AI Two Months In.
FAQ About AI Desktop Agents
What are AI desktop agents?
They are AI assistants that can act on your files and apps, not just chat. They open documents, combine data, create files and run multi-step tasks, usually with approval steps for actions that change or send things.
Which AI desktop agents are available in 2026?
The main three are ChatGPT Work from OpenAI, Claude’s desktop agent from Anthropic and Copilot Cowork from Microsoft. Developers also use OpenAI’s Codex desktop app, and enterprise vendors such as ServiceNow are building their own.
How much do AI desktop agents cost?
For individuals, about $20 a month on ChatGPT Plus or Claude Pro, with higher tiers for heavy use. Copilot Cowork needs a Microsoft 365 Copilot licence and bills usage in credits.
Do AI desktop agents work when my computer is off?
Some do. Claude’s agent runs in Anthropic’s cloud and continues with the laptop closed, but it needs the desktop app open to reach local folders. ChatGPT Work can run in the cloud or on an approved connected computer.
Are AI desktop agents safe for company data?
They can be, with the right settings. Use admin controls, keep approval on for sending and writing, check data retention terms and never give an agent raw passwords.
What is the best first task for an AI desktop agent?
Organising a copy of a messy folder. It is tedious, reversible and easy to check, which makes it the ideal way to learn how the agent behaves.
Verdict
Start using AI desktop agents this week, on the dullest work you have. Folder cleanup, document assembly and a weekly read-only report will save real hours and teach you how the agent behaves before anything important is at stake. Pick the agent that sits closest to where your files already live.
The honest caveat is that these tools still make confident mistakes, and the tasks that sound most impressive, sending, buying and deleting, are exactly the ones where a mistake costs the most. Approval modes exist for a reason. Use them.
Your next step: make a copy of your messiest folder, ask your agent of choice to propose an organisation plan as a table, and review it before letting it run. If that goes well, schedule one read-only weekly report next.
Sources and Further Reading
- Deckary’s three-way comparison, for launch dates, plans, prices, permissions and data retention
- TechTarget on desktop agents, for how OpenAI’s agent grew out of Codex and how Copilot Cowork differs
- Developers Digest, for the ChatGPT Work launch and the Microsoft 365 connector history
- Adapt on Workspace Agents, for how OpenAI’s team agents compare with single-user desktop agents
- CallSphere’s buyer’s map, for ServiceNow Project Arc and the wider enterprise field
More on this site: AI browser agents covers agents that work on the web, no-code AI agents shows how to build your own without code, and OpenClaw explained covers the open-source alternative for self-hosters.

