Which AI Jobs Need Your OK?
The assistants your staff use can now send, move and delete, not just suggest. Sort every job into four piles before they do.
By John Martines · October 13, 2026 · 6 min
Claude, ChatGPT and Gemini can now act inside your email, and each one asks permission in a different way. Here is a four-pile sort (look, draft, ask first, never) that any small business can set up in an afternoon.
It is Friday afternoon at a remodeling company. The office manager has forty unread emails and a site visit at four. She connects her AI assistant to the company inbox and types: "Clean this up and answer the easy ones."
By the time she gets back, it has done exactly that. It archived six supplier threads, including the one with Monday's delivery window. It answered a homeowner who asked about a start date, using the date from an old estimate. It sent a polite "thanks, we'll be in touch" to a sales rep, which was fine.
Nothing was hacked. Nobody did anything wrong. The assistant did what it was told, with the permissions it was given, and nobody had decided which of those jobs it could finish on its own.
That decision is now yours to make, and the tools are finally asking you to make it.
For two years, AI assistants mostly wrote things for you to copy and paste. That is changing. Connected to your email, files and calendar, they can now take the action themselves: send the reply, move the file, delete the thread.
Every major assistant handles this with some kind of permission setting. The settings just don't match.
Claude lets you set each email action to "Always allow," "Needs approval" or "Block." By default, reading is allowed and anything that sends, replies, forwards or deletes needs approval. On Team and Enterprise plans, the account owner decides whether staff may turn the prompts off.
ChatGPT's Gmail plugin has four levels: "Always ask," "Allow read actions," "Allow low-risk actions" and "Allow all actions." The default is the middle one, "Allow low-risk actions."
Gemini's Gmail connection is on or off, with no settings in between, according to PCWorld's October review.
Claude Cowork, the desktop version that works on your files, has three modes: manually approve, automatically approve (Claude checks each step and blocks what it thinks is unsafe) and skip all approvals, where in Anthropic's own words "nothing checks its actions." It always asks before permanently deleting a file, in any mode.
Notice what those settings have in common. Each one is a sorting decision: which jobs the assistant may do alone, and which wait for a person. A new product built for IT departments, Cuttlefish, makes the sorting explicit with four tiers: read, prepare, require approval and block. You don't need that product to use the idea.
How it got here:
- September 8, 2026: Engadget published a how-to for letting Claude send email, noting that by default it asks before sending, replying or forwarding.
- October 9, 2026: PCWorld compared the Gmail permissions in ChatGPT, Claude and Gemini and found three different designs.
- October 2026: Anthropic's Cowork help page described its three approval modes and recommended manual approval when "the task touches sensitive files, accounts, or sites."
- October 12, 2026: Cuttlefish published its pricing, $499 a month for five people, for a control layer that sorts every AI action into read, prepare, approve or block.
In a big company, someone in IT sets these switches. In a small business, they get set by whoever connects the assistant first, usually on a busy day, usually by clicking through.
Three things make that risky.
Defaults differ by product. If one person on your team uses Claude and another uses ChatGPT, the same request ("answer the easy ones") can stop and ask in one tool and go straight out in the other.
"Low-risk" is the vendor's call, not yours. A reply to a customer may be low-risk to the software and high-risk to you.
Some actions can't be undone. A sent email stays sent. A permanently deleted file is gone. An approval prompt is the last chance to catch a mistake, so it belongs in front of the actions you can't take back.
A free account opens the rest of this article and the whole library.