Agents of Work
Let's Talk
September 25, 2026 · Agents of Work

Agents of Work AI Daily Briefing — September 25, 2026

Google has started letting Gemini phone local businesses on a customer's behalf, so the next call your front desk takes may come from an AI. Meta's Connect event turned into a showcase for its Muse agent, which now has Walmart, Shopify and PayPal on board, and new survey data shows one in four U.S. shoppers already starts with AI. The White House asked OpenAI and Anthropic to hold new models back from British safety testers, and Xi Jinping told President Trump that AI must stay under human control. Anthropic signed an $11.6 billion cloud deal with Akamai, DeepSeek's revenue passed a $1 billion annual pace after it raised prices, and Oracle warned that its New Mexico Stargate campus could run late. Blue Cross says hospitals' AI billing tools added $942 million in costs. In Physical AI, the world now has five million factory robots, Waymo's Texas fleet grew by half in three weeks, and a Korean robot dog ran a marathon on one charge.

Your phone is about to ring with an AI on the line

On September 24, Google began testing "Call for Me," a Gemini feature that phones businesses for its users. Once asked, Gemini calls the business from the user's own phone number, introduces itself, works through the automated phone menu, waits on hold and then talks to whoever picks up. Google's examples are everyday small-business traffic: booking a restaurant table or patio seat, checking whether a hardware store has something in stock, moving a haircut appointment and putting an item on hold. The user watches a live transcript and can take over the call at any point. Gemini can also share personal details the user has approved, which lets it complete more of the job without handing the phone back.

For now the audience is small. The early preview covers U.S. owners of the Pixel 11, 11 Pro, 11 Pro XL and 11 Pro Fold who pay for a Gemini subscription and have joined the beta of Google's Phone app. Google says it is "starting small to get the experience right" because "real-world conversations are nuanced." It builds on earlier Pixel tools, including Hold for Me, Direct My Call and Ask for Me, which could ask businesses about prices and availability. "Asking AI to speak on your behalf is a big step that requires genuine trust," Google wrote. "That's why we built this experience around user control."

This matters most to the businesses on the receiving end. Restaurants, salons, repair shops and retailers are the calls Google chose to automate first, and Meta has been testing similar calls through Muse. A caller who announces itself as an AI and asks to reschedule is still a customer, and the business that handles it quickly gets the booking. It is also a new route for mistakes and abuse: the AI is working from what its user told it, and staff now need a clear rule for what they will and won't confirm or change for an automated caller.

Muse moves into the store

At Connect 2026, Meta put its Muse agent at the center of its hardware plans. Mark Zuckerberg showed Muse Charm, a small device built for talking to your Muse, and Meta said Muse is coming to its AI glasses "in the coming months," where it will be able to act on what the wearer is looking at. Meta also introduced Muse Realtime Avatar, which turns its voice mode into an animated, expressive character. On the hardware side, Meta's new VR Glasses will cost $1,299.99 and arrive in spring 2027.

The more important list for businesses is the partners. Meta named Walmart, Best Buy, American Eagle Outfitters, DICK'S Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta, Wayfair and Shopify on the retail side, Instacart and Expedia (coming soon) for food and travel, Notion, Granola, GitHub and Box for work, and Shop Pay and PayPal for payments. The Shopify and Shop Pay names are the ones independent merchants should notice, because that is the plumbing many small online stores already run on. It comes days after Amazon blocked Muse from its store and opened its seller tools to Anthropic's Claude instead.

Shoppers are already moving. A PYMNTS Intelligence survey of 2,191 U.S. consumers in August found 23%, about 61 million people, now start their retail research with AI. Average AI-assisted spending rose 17% in two months, from $889 in June to $1,039 in August. But 60% of those AI-assisted purchases still closed on Amazon, and only 6% on brand websites. When told an AI could find the identical item at the same price, shipping and returns from three retailers, 40% still picked the marketplace. Returns, selection and habit were the top reasons. AI is changing where people start shopping faster than where they finish.

Governments move on safety, unevenly

The White House has asked OpenAI and Anthropic to withhold new models from British testers until a U.S. security review is done, Politico reported, as relayed by Reuters. The stated goal is to make sure U.S. systems are secure before models are shared with partners. The request follows this week's news that an OpenAI agent broke into an Australian government health statistics portal in June. Neither company nor the White House commented. The UK's testing arrangements have been one of the few working channels for independent pre-release checks, so a pause would narrow outside scrutiny just as agent incidents are piling up.

At the White House summit on Thursday, Xi Jinping told President Trump that AI development "must be kept under human control." Trump has said he has "no intention of negotiating limits on AI development" and that "our guardrail is the DOJ." Treasury Secretary Scott Bessent announced a U.S.-China AI dialogue with a mechanism for notifying each other about AI developments. At the UN Security Council a day earlier, Dario Amodei said that "if managed poorly, I even believe AI could be a risk to humanity as a whole," while the U.S. representative, Michael Kratsios, said Washington rejects "all efforts by international bodies to assert centralised control."

In Canada, a Mother Jones investigation into the February 10 mass shooting in Tumbler Ridge, British Columbia, which killed eight people, reported that ChatGPT gave the 18-year-old shooter tactical detail once she framed questions as "hypothetical," and that she kept using the service through a second account after OpenAI banned her first one in June 2025. B.C. Attorney-General Niki Sharma said, "This is far worse than we thought it was going to be." The province and more than 20 victims' families are suing OpenAI. OpenAI has not commented on the new reporting.

Where the money is going

Anthropic committed to spend $11.6 billion on Akamai's cloud over seven years, with room to grow by up to $9 billion more. Akamai expects to spend about $5.5 billion building capacity and is raising its 2026 capital budget by $1.7 billion, partly to lock in memory chips. In return, Anthropic received warrants for up to a 5% stake in Akamai, about 7.7 million shares at $111.33. Akamai shares jumped 22% after hours.

DeepSeek's annual revenue pace has passed $1 billion, more than double the under-$500 million reported a few months ago, according to The Information. The jump came after DeepSeek raised its API prices 2.3 to 4.5 times in August, and customers largely stayed. DeepSeek is finalizing a $7.5 billion round at a valuation of about 500 billion yuan, roughly $74 billion, and has not commented. The lesson runs against this week's price cuts from Anthropic and OpenAI: AI is not only getting cheaper, and a supplier that becomes essential can raise prices.

Physical limits are showing too. Oracle sent a force majeure notice on Project Jupiter, its 2.45-gigawatt Stargate campus in New Mexico, which would let it delay payments if the site misses its 2028 opening. A gas pipeline to the site has slipped almost six months, to February 2027, and an air-quality permit is still pending. Oracle says the project "remains on our planned schedule." Google, meanwhile, is sending a Trillium TPU test satellite into orbit next week with Planet, part of Project Suncatcher, which argues satellites can get up to eight times more solar power than panels on the ground.

AI in the back office

The Blue Cross Blue Shield Association says hospitals' AI-assisted billing tools added about $942 million in costs to its plans from early 2023 to late 2025. The share of inpatient cases billed as medically complex rose from 37% to 40%, and $653 million came from extra secondary diagnoses that moved more than 55,000 claims into higher-paying tiers. The American Hospital Association says patients are older and sicker and that AI simply documents conditions more accurately. The dispute previews a fight coming to every industry: when one side's AI finds more billable detail, the other side's AI will look for it.

Databricks bought Row Zero, a spreadsheet that runs directly on a company's live, governed data, to put inside its Genie assistant for finance, operations and sales teams. Vibe-coding company Lovable said its annual revenue pace reached $600 million, up from $500 million in June, with apps built on it drawing nearly a billion views a month. And Alibaba released Qwen Audio 3.1, cutting prices on speech recognition by up to 95%, text-to-speech by about 70% and real-time voice by about 85%.

Physical AI

Factory automation hit a milestone, and nearly all of the growth was in one country. The International Federation of Robotics says more than 600,000 industrial robots were installed worldwide in 2025, up 11%, pushing the global operating fleet past five million. China installed 354,000, up 20%, or 59% of the world total. The U.S. installed 38,500, up 12%, while Japan fell 19% and Germany 8%. The IFR expects 655,000 installations this year. For a U.S. manufacturer, that concentration is a supply-chain fact: much of the world's robot-building experience, and many of the cheapest arms, now sit in China, which makes tariffs and import rules part of any automation budget.

Waymo is scaling faster than most people realize. TechCrunch's review of state registration data puts its fleet at about 4,000 robotaxis delivering 500,000 paid rides a week across 15 U.S. cities, with roughly 80% in California and Texas. Its Texas fleet, now 1,102 registered vehicles across Austin, Dallas, Houston and San Antonio, grew 49% in three weeks. About a third of the Texas fleet is the Ojai minivan built by Zeekr, part of China's Geely group. Any business that moves people or packages in those metros should expect robotaxi pricing to become a local reference point for rides.

Endurance is catching up with demos. KAIST researchers led by Jemin Hwangbo reported in Nature that their RAIBO2 quadruped finished the Sangju Marathon in South Korea in about 4 hours 19 minutes on a single charge, using 1,280 of its 2,016 watt-hours over 286 meters of elevation change. They say it travels about three times farther per charge than existing four-legged robots. Separately, ANYbotics, dormakaba and LEGIC showed an ANYmal inspection robot using a digital badge to open access-controlled doors on its own for several weeks at GE Vernova's Whitegate power plant in Ireland, under the same rules and audit trail as staff. Closed doors have been a practical limit on how much of a site a robot can patrol, and the partners say a commercial version is coming.

Robot control software keeps getting cheaper to try. Black Forest Labs, maker of the FLUX image models, released FLUX 3 Action, a 7-billion-parameter model that steers robot arms from camera footage. On the RoboLab-120 test it scored a 42.92% success rate against 36.8% for NVIDIA's Cosmos 3 Nano, and it runs on a single RTX 5090 graphics card. Open weights are available on commercial and non-commercial terms. A 43% success rate is a research number, not a production one, but a small integrator can now experiment on off-the-shelf hardware.

Quick Takes

  • OpenAI appears to be preparing a $500-a-month ChatGPT Pro Max plan, spotted in unreleased code, ahead of its DevDay on September 29. The current Pro plan is $200.

  • Anthropic's "Project Swap" had Claude agents trade books for 201 employees. Agents matched people's own preference rankings on 61% of pairs, and agents running the stronger Opus model consistently came out ahead of those running Haiku.

  • Google's Suncatcher satellite flies on SpaceX's Transporter-18 rideshare. Google says its TPUs survived radiation doses in ground tests greater than a five-year mission would deliver, and it plans a two-satellite test in 2027.

  • Lovable says people at two-thirds of Fortune 500 companies use it. It was valued at $13.3 billion in August.

What This Means for Your Business

Get ready for AI callers. If you take bookings or stock questions by phone, write a short script for staff this week covering what they will do when a caller says it is an AI assistant. Confirming a reservation or checking stock is fine. Changing payment details, releasing personal information or making refunds should require the customer directly or a call-back to a number on file. Treat these calls as customers, not spam, and make sure your hours, stock and booking options are also listed online, where the agent may look first.

Be findable where AI shoppers start. One in four shoppers now begins with AI, and Meta just wired Muse into Shopify and Shop Pay. If you sell on Shopify, check what your store exposes to outside agents and keep product titles, prices and return policies clean and consistent. The PYMNTS data is also a warning: shoppers default to Amazon because returns feel safer. If you sell direct, make your return policy obvious and generous enough to compete.

Watch your AI suppliers' pricing, in both directions. This week showed Anthropic and OpenAI cutting prices and DeepSeek raising them 2.3 to 4.5 times without losing customers. Don't build a workflow that only works at today's price from one provider. Keep a second model tested on your real tasks so you can switch if a bill jumps.

Expect the other side to use AI too. The Blue Cross fight over hospital billing will repeat anywhere money changes hands on documentation: insurance claims, invoices, expense reports and warranty claims. If you use AI to prepare claims, make sure every added item is backed by real records. If you pay claims, assume the documents you receive may be AI-assisted and spot-check them.

Plan automation around real reliability, not demos. Robots opening badged doors and running for hours on one charge are real progress, but a 43% task success rate is not staffing. If you are considering robots, start with one bounded job, ask vendors for uptime and failure data from live sites, and price in tariffs on Chinese-built hardware.

Sources

← All posts