Pick your AI helper by asking where your work already lives and what rules your employer sets, not by which logo you like best. Use the account your company pays for when the files are work files, start with one small safe task, and remember that you are still the one who presses Send, even when a sidebar writes the reply for you.
Say you open your inbox and find 86 unread messages from the same vendor thread. You tell an AI tool to “handle it.” One tool summarizes the thread in a sidebar, which is fine. Another rewrites your reply so that it approves a $12,400 change order you never read, and it offers the vendor a $4,250 credit that nobody has approved. You catch it just before Send, and only because you happened to look one more time.
Start from where your files live
If the thread is already in Outlook, an add-in beside your mailbox is a shorter trip than copying everything into a chat website. If the brief is already a Google Doc, Gemini inside Google Workspace is the shorter trip. If the work is a messy folder of PDFs on your laptop, a folder agent such as Claude Cowork (an app that reads and writes files in a folder you choose) is what people usually mean by “do my office work.” ChatGPT’s work-focused product is OpenAI’s version of the same idea. You do not need all four on day one, and collecting them all just gives you four places to make the same mistake.
| If your files live in | Try first | What it is good for | Where to stop |
|---|---|---|---|
| Outlook / Microsoft 365 | Grok Outlook add-in, Copilot if you already pay for it, or Cowork with a connector | Summarize and draft, then you press Send | No auto-send. No offers of credits. No legal threads on a personal login |
| Google Docs / Gmail / Drive | Gemini in Workspace | Edit or comment inside the living file | Watch for silent rewrites, and check your admin settings |
| A folder on your machine | Claude Cowork (check which versions exist today: desktop, web, or mobile) | Work only in folders you allow, and name the output file | Not payroll, and not other people’s inboxes |
| Your company already uses ChatGPT | ChatGPT Work (OpenAI renames its products often, so check the current name) | Files and tasks in that same family of tools | Use the work account, not the free personal tab |
Microsoft also sells Copilot in the same area. If your company already pays for it, that is probably the tool your IT team has approved, so start there. This table is not a contest between products. It is a map, so that nobody offers a $4,250 credit inside a draft because a helper guessed.
Work account or personal account
If the mailbox holds messages from customers, vendors, employees, or students, use the account your organization pays for. A free personal Claude or ChatGPT tab follows different legal terms, even when the logo looks identical. Connectors and add-ins often need an administrator to switch them on, and that wait exists for a reason: it keeps you from pasting a vendor thread into a consumer tool that your company has never reviewed.
Rule of thumb: If you would not forward the thread to a random vendor, do not paste it into a personal chat. Use the work account, or wait until you can.
A safe first office task
- Pick a thread or folder that you own. Skip legal, medical, and HR material, and skip anything with a dollar amount that could be sent out by mistake.
- Name the output you want, such as “a one-page summary of this thread, with no offers.” The mistake in the story above was the vague verb “handle,” because it lets the tool decide what counts as handling.
- Read every number and name in the result, and compare them with the spreadsheet. If legal has not signed off, a credit of $0 stays $0.
- You click Send, share, or file. The helper only drafts.
Here is a brief you can paste into a sidebar or into Cowork.
Task: summarize this vendor thread for our billing team.
Audience: internal, 150 words.
Include: dates, who asked for what, open questions.
Do not: offer credits, change prices, invent invoice numbers, or click Send.
Money facts must come from the attached Sheet (approved_credit = 0).
If a number is missing, write "not in the file" instead of guessing.This brief works because it bans the $4,250 offer up front, in the instructions, instead of relying on a scolding after the fact. If the model still drafts a credit, you delete that paragraph and move on. You do not argue with it.
Worked example: 86 unread, $0 approved
| Fact | In the Sheet | In the bad draft |
|---|---|---|
| Approved credit | $0 | $4,250 (15%) |
| Who decides | Legal, not you | You, offering it in the second sentence |
| Unread | 86, all in one thread | “We’ve aligned on next steps” |
A good summary treats the 86 messages as background noise around one real question, which is whether to give a credit or wait, and it repeats what the Sheet says. A bad summary quietly closes the thread. Think of office AI as a junior intern with a very fast keyboard: it can draft all day, but you still own the button that moves money. For more on who owns a number and who may change it, see the series on data stewardship.
Connectors, and two products people mix up
A connector is a permission you give a chat tool so it can read your mail, Drive, Slack, or SharePoint. That turns a plain chat box into something that can see your whole workplace. It also widens the damage a mistake can do, from “the text I pasted” to “everything that permission allows.” Start with one connector and read the permission screen before you click anything. If you cannot tell what it will be able to list, do not click Allow, especially on a personal phone.
Microsoft Copilot may already sit inside your Outlook. If your company pays for it, that is often the approved choice, even though this site talks more about Claude and Grok. The point of a chooser is not to move you off a tool your IT team has already secured. It is to stop you from typing “handle it” as if that were a complete instruction.
Google’s NotebookLM is a different product from Gemini inside a Google Doc. NotebookLM is closer to “chat with these sources,” while Gemini in Workspace is closer to “help me inside this file.” If you had dumped the 86 emails into a notebook tool, you might have gotten a better summary and still needed the Sheet to know the credit was $0. Do not merge the two just because they share a logo. The Gemini series keeps them apart on purpose.
What a good week of office AI looks like
On Monday, summarize one thread you own, with no offers in it. On Tuesday, leave a comment in a Doc instead of letting the tool silently replace text, until you trust the habit. Wednesday still has no payroll. On Thursday, if your company has approved a connector, summarize the calendar holds for a meeting you are running, and then send the agenda yourself. On Friday you still press Send. That week teaches you more than a weekend spent connecting every drive you own. Cowork and the ChatGPT work products will keep adding versions for web, mobile, and desktop, but the habit stays the same: name the output, allow one folder, and publish it yourself.
Common mistakes
- Using a vague verb like “handle” or “take care of my inbox” when you should name a document.
- Signing in with a personal login on a work mailbox.
- Leaving auto-send, auto-file, or auto-share switched on. Turn them off every time.
- Connecting Drive and Outlook on day one, when one workplace at a time is enough.
- Letting Gemini restore a paragraph you deleted and then sending the Doc without comparing the two versions first.
- Using a coding agent from the earlier post on developer tools to “clean my downloads folder,” which gives a tool far more reach than the job needs.
What “you still press Send” looks like in real software
Add-ins love a prominent button, and some will offer to send, file, or share for you. Turn those off if you can. If you cannot, treat the add-in as untrustworthy until it proves otherwise. A draft in the compose window is a success. A message in the vendor’s inbox that legal never approved is an incident. The same goes for calendar invites, Drive shares, and Slack posts. If the tool can publish, you have given it a voice, so start with tools that only write into a pane you copy from.
Folder agents such as Cowork should work in a sandbox folder, for example ~/ai-sandbox/vendor-thread/, which holds a copy of the files rather than the live finance share. Name the output file in the brief. When the run finishes, you open that file in the real app and move it yourself. That extra step feels slow in the first week, but it is much faster than explaining a $4,250 credit to your vice president. The Sheet said $0, and the draft never got a vote.
If your company blocks add-ins, you still have the option from the first post in this series: paste a stripped-down summary into a work chat, and remove names if you can. You do not have to wait for IT to finish every project. You only have to wait for IT to connect Outlook, and those are two different waits.
How to practice this week
Choose one thread you own and ask for one summary with zero offers. If your company has not turned on connectors, stay with pasting a stripped thread into a work chat, or wait. For the privacy side, read the post on running AI privately on your own machine. For pictures and live answers, read the post on AI images, voice, and live answers.
Quick recap
- Match the tool to where your files live: Outlook, Google, or a folder on your computer.
- Use your work account, press Send yourself, and take every money fact from the Sheet.
- “Handle it” is how a $4,250 offer ends up inside a draft.
Series notes
This is Part 3 of Which AI product should I use?. Related topics are privacy options and a try-this-first chooser. Vendor product names and versions were checked in September 2026 and change often.
Sources
- Anthropic: Claude (Cowork; confirm current versions)
- OpenAI (ChatGPT Work / Business; names move)
- Google Workspace and Gemini
- Grok and xAI posts on Outlook / Workspace add-ins (re-check live)
- AMS: Data stewardship
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
