AI shopping assistants choose sellers on delivery and returns, and California limits AI firings

For the first time, more than half of American consumers used an AI tool to help them shop in the past month. An assistant answers a request such as "a garden bench under $300, delivered before Saturday, with free returns" by reading across sellers' pages, and it can only recommend what it can confirm there. The advert never reaches it; the delivery promise, the returns policy and the reviews do.

That moves part of the sales job to the people who run the warehouse, the deliveries and customer service. Brand websites get only 6% of AI-assisted purchases, down from 10%, and half of the shoppers who buy on marketplaces would buy direct if delivery were as fast. A seller whose terms an assistant cannot read is easier to leave out, before any person sees its page.

For example, a furniture retailer's returns page says 30 days, a PDF says 14 days for large items, and the delivery date appears only after the shopper types a postcode at checkout. An assistant asked for that bench cannot confirm either term, so it sends the shopper to a marketplace. One plain page with delivery times by region and the real returns rules is about a day's work. Late deliveries also become one-star reviews that the assistant reads, so the on-time rate starts to work like advertising.

Within two years, operations teams will be judged on how their promises read to a machine, the way marketing teams have been judged on search rankings.

ElevenLabs could already clone voices. The new model makes a faithful copy from 10 seconds of audio, which is one voicemail greeting or the first answer in a recorded webinar. The copy stays recognizably the same voice in more than 90 languages, so a criminal can make an English-speaking finance director call a Brazilian supplier in Portuguese.

Voice-cloning fraud was already rising sharply, and a familiar voice on the phone now proves nothing. A request to change bank details is confirmed by calling back on the number already in the supplier file, never the one the caller gives.

A Gem was a custom assistant that worked only inside Gemini. A Skill is a saved set of instructions that Gemini uses when someone types "/" and its name, or runs by itself when a request matches. Anthropic created the format as an open standard, and OpenAI and Microsoft already use it. Gems convert automatically, in November for personal accounts and March 2027 for business accounts, but Skills do not yet support every Gem feature, so each one deserves a test.

A shared format also means shared marketplaces. Skills.sh, one of them, grew to a million skills in seven months, and about a third carry security flaws. A skill can include scripts that the assistant runs with the access of the person using it, so an unknown skill deserves the same caution as an unknown app on a work laptop.

Old systems do get in the way: more than four in five technology chiefs have cancelled an AI pilot because of them. Replacing them is still the riskier cure. An old ERP holds years of rules that nobody wrote down, such as the customer on 90-day terms or the part number that changed in 2019. A replacement must rediscover every one before its AI can do the job. AI added on top reads what the old system produces and leaves those rules running.

The system that looks outdated is often the only complete record of how the company works.

That spending is up from under $2 billion in 2025, and most of it will arrive inside the planning software that companies already rent. When a port closes, a planner can spend a morning emailing forwarders for quotes; an agent can collect them and rebook a stuck container in minutes.

The agent will use whatever lead times and carrier rates are in the system. A planner knows which supplier always ships a week late, and the agent knows only what is written down. The data clean-up that has waited for years now decides how fast a company recovers.

From July 2027, when software flags a California worker for discipline or dismissal, a person must check the finding against other evidence, such as personnel files and manager evaluations, and the worker must be told what data the system used. A warehouse tool that scores how fast staff pick orders falls under this as soon as its scores lead to a warning. A separate law requires notice when AI causes a mass layoff. A manager who simply approves the system's flag is not the review the law requires.