Customers pick businesses inside ChatGPT, and Binance's AI tool helped thieves reach its wallets

When the question is local, such as a plumber, a dentist or a van rental, the customer often chooses without leaving the chat. Ask ChatGPT for the best one in a US city and it nearly always answers with a list of business cards, about 12 names with star ratings. The customer picks one and phones. Few people open a website: ChatGPT sends local businesses less than 1% of the visitors that Google sends them.

So the website is no longer where a business gets judged. The shortlist is built from listings and review pages, and the ratings on those cards match the business's Google profile. An old phone number or wrong Saturday hours on a listing goes straight into the answer. The lost customer leaves no trace, because there was never a visit to record.

For example, a tool rental company with 30 depots could have one person ask ChatGPT, for each town, the three questions that its customers ask most: who rents small excavators nearby, who is open on Saturday, and who delivers today. Those 90 questions take about half a day. The result is a table that shows, for each depot, whether it is named, whether its phone number and hours are right, and which competitor appears instead. Each wrong line points to one listing to fix.

From now on, the useful figure is how often ChatGPT names the business, along with how many new callers say they found it there.

In September, a fake recruiter on a job site took a Binance software engineer through two interviews and then asked him to download a coding assignment onto his work laptop. The assignment carried malware. On October 20, the attackers used his laptop and an internal Binance AI tool to take over the computers of two engineers who worked on the company's crypto wallets. Nine days later, about $6.34 billion was gone.

An internal AI assistant is useful because it can reach many systems and the work of many colleagues. Whoever controls a laptop where it is signed in can use that reach too, and that is how one infected machine became three. So an assistant needs the limits a new employee in the same job would get: access to what the work needs, and nothing more.

The tool reads an application for a kitchen or loft extension, checks it against the council's planning rules and its list of required documents, and drafts the officer's report. In pilots at three English councils, officers approved more than 70% of those reports without changes.

The time saved shortens a family's wait only if the council puts it into the queue, instead of letting it spread into email and meetings. Any team whose AI halves a checking step, such as matching invoices, has the same choice to make before the hours disappear.

In June, Uber capped what each engineer could spend on coding tools.

A cap makes people ration the valuable requests along with the routine ones. Most requests are routine, such as summarizing an email or pulling the fields from an invoice, and they go to the strongest model because it is the default. Smaller models do that work well at a fraction of the price.

If 80% of a finance team's requests move to a model that costs a fifth as much, the bill falls by almost two thirds, and nobody has to ration the requests that matter.