Half of American shoppers now use AI to shop. Not "might someday." Did it, last month. A new survey from NielsenIQ puts the number at 51%, the first time it has crossed the halfway line, and the rest of this week's news reads like the follow-up: AI that phones shops for the customer, AI with a face that talks in 97 languages, and a US regulator reminding everyone who pays when an AI agent messes up.
I went through the week so you don't have to. Five stories, a few quick takes, and one thread running through all of them: the customer's side of the conversation is getting its own AI. Your side had better keep up.
Half of US shoppers now use AI to shop
On September 24, NielsenIQ (NIQ) published the latest reading of its monthly Agentic Commerce Tracker. 51% of US consumers said they used at least one AI tool to help them shop in the past month. The most common use was product recommendations (20%), followed by AI personal shopping assistants (16%). NIQ's North America president called it the moment AI shopping went "mainstream."
Two caveats, because I'd rather you trust me than be impressed. It's a monthly quick survey of about 500 people. And it's the US, not Iran. But the direction is the same everywhere, and NIQ's most useful finding isn't the headline number. It's how people use it: AI doesn't replace shopping, it compresses the decision. People ask it to compare, judge value and cut the list down.
Why it matters to you: a shopper who has already narrowed things down arrives in your DM with a sharper question and less patience. "Is this one better than the other one for oily skin?" "Why is yours more expensive?" If your answer is "price sent in DM 🙏" three hours later, the comparison has already been won by someone else. The seller who explains, compares and replies fast is the one who stays on the shortlist. We've written about what to say when a customer says "too expensive", and that question is only getting more common.
Google's Gemini will now phone the shop for you
Also on September 24, Google started testing "Call for Me" on Pixel 11 phones in the US. You tell Gemini what you need, and it calls a local business for you: it gets through the phone menu, waits on hold, then talks to whoever picks up. It can ask whether something is in stock, book a table or move an appointment. You watch a live transcript and can take over at any moment. Google says it won't handle payments, passwords or card numbers, and it can't call emergency services.
It's an early preview, US-only, and you need a paid Google AI plan plus the beta version of the phone app. So no, an AI isn't going to call your shop in Tehran tomorrow.
Why it matters to you: look at the list of what it asks. Stock, availability, bookings. Those are exactly the boring, repetitive questions that already fill your inbox. When the customer's assistant starts asking them, it won't wait politely until morning; it'll move on to the next shop on the list. The businesses that answer instantly, with correct stock and prices, win both the human shopper and the machine one. That's the case for answering customers around the clock, and it just got stronger.
Google gives AI agents a face that speaks 97 languages
Google made Gemini 3.8 Live with Live Avatar generally available for its business customers on September 24. It's a live AI agent with an animated face: it talks, listens, sees your camera or screen if you share it, lip-syncs and keeps its expressions steady across 97 languages, and can switch languages mid-conversation. It can also look things up in the background without pausing, which is the difference between a demo and a useful assistant.
Businesses can pick preset faces; making a custom avatar from a reference photo is still limited to approved customers. Every piece of audio and video carries Google's invisible SynthID watermark, so it can be identified as AI-made.
Why it matters to you: this is enterprise tech for now, priced and packaged for big companies. But watch the pattern. Last month it was cheap AI voices; this month it's faces. A video "shop assistant" that greets visitors in their own language is coming down the price curve toward small sellers. My advice: don't rush to put a fake face on your brand. Customers forgive a text bot that says it's a bot. They're much less forgiving when they feel tricked.
The FTC's message: your AI's mistakes are your mistakes
At a Reuters conference in Austin on September 25, US Federal Trade Commission chair Andrew Ferguson pushed back on the idea that AI agents act "on their own." He said that when companies claimed a system had slipped out of control, the logs often showed it had been told to do exactly what it did. He said he'd keep resisting what he called the "anthropomorphizing" of these tools, and suggested the people who build and instruct agents are the ones responsible for harm.
This isn't a new law. It's the country's top consumer-protection regulator telling the market how he thinks about blame.
Why it matters to you: Iran has no FTC, but your customers have the same instinct. If your assistant promises free delivery you don't offer, or quotes a price that doesn't exist, "the AI said it" is not an excuse anyone accepts. So:
- Give your assistant real, current facts: prices, stock, delivery rules. Guessing is where trouble starts.
- Write down what it must never promise (discounts, refunds, delivery dates you can't keep).
- Decide in advance when the AI should hand the chat to a human, and make sure a human actually picks it up.
Microsoft puts AI agents on the meter
On September 25, Microsoft rebuilt its Microsoft 365 Copilot app around three parts. Home puts chat, email, calendar and tasks in one place. Code lets non-programmers build small apps and dashboards by describing them. Autopilot is an always-on agent with its own cloud computer that can handle recurring work without someone pressing a button. Home and Code roll out to early-access customers in the coming weeks; Autopilot goes into private preview.
The part to notice is the bill. Everyday Copilot chat stays on a normal per-user license, but the agent features move to usage-based billing. The more work the agent does, the more you pay.
Why it matters to you: the industry is quietly splitting AI into two prices: chat is a flat fee, "doing things" is metered. Expect every tool you use to go the same way. And prices don't only go down. The Information reported this week that DeepSeek, the famously cheap Chinese AI lab, raised its API prices by 2.3 to 4.5 times last month and says customers stayed anyway. Before you sign up for any AI tool, ask how it charges when you get busy, not only on a quiet day.
Quick takes
- Gemini 4 is coming "much earlier" than year-end. Google DeepMind's head, Koray Kavukcuoglu, said on September 23 that the model is in early post-training. No date, no specs, no prices yet.
- OpenAI DevDay is on September 29 in San Francisco. Expect developer tools and agent news; we'll report what actually ships, not the hype.
- DeepSeek's revenue run rate passed $1 billion, more than double a few months ago, according to The Information, as it finalizes a large funding round ahead of a planned Shanghai listing.
- A bill to ban "superintelligence." Sen. Bernie Sanders and Rep. Greg Casar introduced a US bill on September 23 to prohibit AI systems that exceed human ability across most areas. It has little chance of passing soon, but it shows how loud the safety debate has become in Washington.
If you want the backstory on how fast shopping agents moved this month, our digest on Amazon's new seller agent is a good place to start.
Here's the thread: customers are arriving with AI already in their hands, sharper questions and less patience. The shops that win are the ones that answer every message fast, with the right facts, on every channel. That's the job we built Vardast for: an AI assistant that answers your customers on Instagram, WhatsApp, Telegram and your website, knows your products and hands tricky chats to you when it should.