Stories

Gemini Notebook turns a notebook's files into a website, and Gemini 4 Neon halves the cost of a task

Every Shopify store publishes a file that tells customers' AI agents what they may do, and a campaign to keep customers once made more of them leave.

By , Senior AI ConsultantEdition of

5stories
3minute read
In This Edition

Type "make me a site" into a notebook's chat, describe the page you want, and Gemini Notebook builds a dashboard, tracker or explainer from the files in that notebook. It could already summarise those files and answer questions about them in its chat. What is different is the result: a page that other people open in a browser and click through on their own.

That moves the follow-up questions from the person who writes a report to the people who read it. Picture a hospital group's finance director who sends 12 department heads a monthly pack comparing spending with budget. Each head cares only about their own lines, so the pack starts a stream of emails, such as why agency staff cost more in March. If she builds a site from the same files, each head can choose their department and find the answer without writing to her. At 20 such questions a month and 20 minutes each, she gets back almost seven hours every month.

Check what else is in the notebook first. A shared site opens only for people who can open the notebook behind it. It works the way a spreadsheet with hidden tabs does: the chart is what people see first, but every file in the notebook comes with it. If the payroll export behind the staff costs is in there, every department head can read their colleagues' salaries. So build the site in a notebook that holds only the files you would send anyway.

Share this story

The system raises an alert when a company's software breaks, a service that most companies rent from a vendor. Replit's engineers directed the agents, and Replit now sells the same setup as Missions: a customer states the goal and sets a spending limit.

That limit puts a price on leaving a vendor. Rebuilding a tool used to be a project that nobody could cost, so buyers usually accepted the increase at renewal. Now IT can run a capped attempt to rebuild the features a team uses, and every price increase is measured against what that attempt costs.

Share this story

More than a decade ago, a mobile operator ran a campaign to keep customers, and the share who left rose from 9% to 10%. A call near the end of a contract reminds a contented customer that they are free to go. The operator later found a group of customers whom the contact did help.

An AI agent can write to every renewing customer for almost nothing, so nobody asks who should hear from it. Only a random group that the agent never contacts can answer that: if more of that group renew, the messages are driving customers away.

Share this story

Under Poppy, a company publishes a file on its website that tells a visiting AI agent what it may do there for a customer. Meta's Muse agent looks for that file first on every site it visits, and every Shopify store now publishes one automatically.

Banking apps made the same change before. Budgeting apps once signed in with the customer's bank password; open banking replaced that with access that the customer grants for certain actions only. With Poppy, the customer signs in through a normal login and chooses what the agent may reach. The business bank Mercury lets agents check balances and start payments, and each payment still goes through the account's own approval rules.

Other companies will copy that design: the agent looks things up and starts requests, and the approval rules the company already has decide what goes through.

Share this story

Neon finished the tasks on a set of hard tests for $3.74 each. Gemini 4 Carbon, Google's top model, cost $7.76 for the same score at its standard price, which returns on 1 December after a discount. Neon's price per token is 25% lower, and the rest of the saving comes from writing about a third fewer tokens.

A bill is the tokens a model writes multiplied by the price of each one, and a price list shows only the price. So what a job costs can only be learned by running your own job on each model.

Share this story

THE DAILY BRIEF

Get the next edition in your inbox.

A five-minute read, every weekday morning.

Free forever · Unsubscribe anytime

Share this

Other Editions

Newest first


THE DAILY BRIEF

Read the next one first.

A five-minute read, every weekday morning.

Free forever · Unsubscribe anytime