Review: AI by text message, files inside chat, and a simple test for every new AI feature

Every week brings another wave of AI features, and most of them land in the same place: a chat panel behind a sparkle icon. This week a few launches tried something different, and one essay offered a test I plan to use on all of them. These are first assessments from the announcements, not test results.

The reviews

ChatGPT Dot over text message. Verdict: promising, but wait for the team version. OpenAI now lets Dot agents take instructions over iMessage and RCS. You can send a task, a photo or a voice note, and get proactive updates without opening ChatGPT. Two details matter. You can limit texts to urgent information only, and approvals still happen inside ChatGPT. That second rule is the right instinct: the agent can reach you anywhere, but it cannot get a yes from a hurried reply. The beta is limited to ChatGPT Pro users in the US, with one-to-one messaging and no group chats. For marketing teams that work through shared channels, that limit decides it for now.

Files inside Gemini. Verdict: use it this week for first drafts. Gemini can now generate, preview and edit documents, PDFs, presentations and spreadsheets directly in chat. PDF previews and edits work on web and mobile, and the other formats work on the web only. It is available to all users wherever the Gemini app is. The value is fewer steps between the idea and a file someone can review. The risk is the same as with any first draft: it looks finished before it is. Treat what comes out as a starting point for your editor, not a deliverable.

Anthropic's AI spend reporting. Verdict: overdue, and the right question. Anthropic launched reporting tools that show companies what their spending on products like Claude Code is buying, with insights on productivity, costs and recommendations. Whatever tools you use, this is the question finance will ask next year. If you cannot say what your team's AI spend produced, the budget conversation will be short. Start tracking it yourself, even in a spreadsheet, before a vendor dashboard tells the story for you.

Instagram long-form video. Verdict: skip it for now, unless your short videos already hold attention. One argument this week made a case I agree with. Long-form is not a priority for brands that have not mastered short-form. First make Reels over 60 seconds that earn strong watch time, turn the formats that work into a consistent series, and explore long-form only once short videos are persuading people to watch more. A new format does not fix weak content. It just gives weak content more room.

The test that matters more

The most useful piece this week was a critique. Products from note apps to banks are shipping the same AI chat panel behind a sparkle icon, often on by default and hard to turn off. The essay cites a study finding that 95% of companies saw no measurable return from their generative AI pilots. Its fix: build features with AI, not AI features.

It offers three questions, and I would put them in front of every AI tool your team considers buying or building:

• Who asked for this? If no customer or teammate asked, be careful.
• What is it named after? A feature named after the job it does beats one named after the technology.
• Would it survive being off by default? If nobody would turn it on, it is not solving a problem.

Run this week's launches through them. Dot by text and files in chat both answer a real request: fewer steps to get work done. A sparkle button on a tool you already use often does not.

What I would do

• Pick one launch from this list and test it on a real task from last month.
• Ask the three questions before you add any AI feature to your own product or site.
• Start a simple log of what your AI spend produced, so you are ready when finance asks.

The tools worth keeping are the ones people would miss if you turned them off.