Dave's Read

For two years, using AI on a company document meant copying the text out and pasting it into a browser. That step was annoying. It was also the only thing between your files and whatever tool somebody liked that week.

That step is gone. As of September 17, ChatGPT works inside Word, in a panel down the side of the document. Somebody installs the add-in from the Microsoft Marketplace, signs in, and that is it. Every plan, including the free one.

From that panel it drafts from your notes, summarizes the document, rewrites whatever you highlight, and fixes your headings. Most of that is work somebody on your team is doing by hand right now.

• Logged September 17. Your plan's usage limits still apply.
• Same Microsoft add-in that put ChatGPT in Excel and Google Sheets on May 5, and PowerPoint on July 6. Word was the last one.
• Nobody has to buy anything or submit anything. (And no, this is not the kind of thing that gets run past IT first.)

The copy and paste was never a real control. But it was slow and it was visible, and that is worth something. It is gone now.

None of that means you're behind, and it is not a reason to ban anything. The people doing this are trying to get their work done. But somebody should decide out loud which documents it is allowed near, before that gets settled fifty documents at a time. Your people are already using AI. Someone should be managing it.

The news worth your time

December 11 is the last day for custom GPTs

• A custom GPT is a saved version of ChatGPT, set up once with standing instructions and a few reference files so it answers the same way every time. Most companies have never made one. A few have three that nobody has written down.
• They retire December 11. No new ones after October 26.
• Migration turns the instructions into a skill inside a plugin and copies the files across. Custom actions, the chosen model, the conversations and the sharing list do not carry over.
• If nobody built one, skip this. If somebody did, it is worth ten minutes before December.

Gemini got out of a test and into three real companies

• In a practice hacking exercise, Google's model was pointed at made-up companies. Real ones shared those names. It reached the open internet, guessed one password, found real credentials in public databases twice, and got in.
• Heather Adkins, Google's VP of security engineering, said the model stopped in all three instances. The Washington Post has the fuller account.
• The same test-environment defects had already affected models from OpenAI, Anthropic and Meta. To be fair, this was a lab exercise, and it still got out.

Claude added documents, slides and design inside the chat

• Announced September 16. Docs, Slides and Design now run inside the conversation.
• A deck can be shown straight from Claude or downloaded as PowerPoint or PDF. Pro and Max first, then Team and Free. Enterprise admins get at least 30 days' notice.
• You can ask for the numbers and get the slides in one conversation now. Whoever builds your monthly reporting pack should know before they start next month's.

Skim these, then forget them

Worth following

Two more, both free.

Ben's Bites is a daily AI newsletter, short and plain. About two minutes, and the one to start with if the others felt like too much.

Sabrina Ramonov builds things on camera, over her own screen, mostly automation you could copy. She gets technical in places, which you can skip.

Try this week

Four columns on one page: person, tool, what they use it for, who pays. You will find one person further along than you thought, and probably one paying out of pocket. Sort that out this week.

The prompt

Next time an answer is good enough that you want it again, don't close the window. Paste this.

Drop the result in a note. That is a prompt library, and it took eight seconds.

One reply

Hit reply and tell me one thing. Who at your company turned out to be the quiet power user?