Agents of Work
Let's Talk
October 5, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 5, 2026

Summary

Washington picked its referee for AI this weekend, and the instructions are to keep the industry moving. At the same time, AI moved closer to the money: ChatGPT will show ads while it makes pictures, a top model outscored licensed accountants on the monthly close, and brokerages and business banks are giving agents access to accounts. Federal rules are months away, so the limits are yours to set.

Highlights

  • Federal AI rules are now on a 120-day clock, written by agencies told to avoid overregulation.

  • The FTC chair investigating AI labs is also a vice chair of the new federal task force.

  • ChatGPT will show ads during image generation to free users; Pro and Enterprise accounts will not see them.

  • A top AI model scored 100% on month-end close tasks where licensed accountants averaged far lower.

  • Robinhood's trading agents can touch only a separate account, a limit worth copying for any agent.

  • Only 37% of shoppers would let an AI agent pay for them, so expect agent orders to stay rare for now.

  • Cloud spending limits from Amazon and Google pause a project but are not true hard caps.

  • IKEA will be the first shipper on Kodiak's driverless trucks, starting with a 219-mile Texas highway run.

  • A new safety standard puts duties on companies that operate mobile robots, not only on the makers.

Quick Takes

  • Stripe buys a small-business lender. Stripe agreed to acquire Parafin, which supplies the financing offers small businesses see inside DoorDash, Gusto, Jobber and Mindbody and has helped more than 60,000 businesses get capital since 2020. The price was not disclosed. Stripe's pitch is the gap banks have left: 41% of US small business loan applications were approved last year, down from 59% in 2015. Expect more loan offers inside the software you already use, and compare their cost with a bank line before accepting one. Read more

  • Cloud spending limits arrive, with a catch. Coding agents can start servers and call paid services faster than anyone watches the bill. Amazon Web Services added a monthly spend limit on September 16 that pauses a project when it reaches the threshold, still in limited release, and Google Cloud launched Spend Caps for specific services in July. Neither is a true hard cap that returns errors and stops charges, so a paused project can cost money again once someone resumes it. Read more

  • OpenAI's legal exposure widens. OpenAI has alerted more than 100 organizations to unauthorized activity tied to its AI agents, Semafor reported, and California Attorney General Rob Bonta has subpoenaed the company over cybersecurity incidents involving its models. After criticism from staff, OpenAI's president also withdrew the second half of a $50 million commitment to an anti-regulation political committee. If a vendor's agent can reach your systems, ask how you would be told about an incident. Read more

  • Patch self-hosted GitLab AI Gateways. GitLab fixed a flaw rated 9.9 out of 10, tracked as CVE-2026-90970, that let a logged-in user with Duo Agent Platform access break out of a prompt template and run commands on the gateway. It affects companies running their own AI Gateway, versions 18.1.6 through 19.4, and the fixes are 19.2.4, 19.3.2 and 19.4.1. GitLab.com is not affected, and there is no evidence of attacks so far. Read more

What's Covered

  • The Super Intelligence Force: who sits on the new federal AI task force, what it must deliver, and what its chair has already said.

  • Ads in ChatGPT images: how the new ad format works, who sees it, and what Europe will force into the open.

  • AI and the monthly close: a study pitting licensed accountants against AI models, and what it did not test.

  • Agents with money: Robinhood's trading agents, Airwallex's business accounts, and a card-industry plan to record what an agent was allowed to buy.

  • Physical AI: Kodiak and IKEA's driverless trucks, Lowe's drone delivery pilot, a new safety standard for mobile robot users, Hello Robot's elder care grant and RobCo's valuation.

Featured News

Washington names an AI czar and gives him 120 days

President Trump created a federal task force called the Super Intelligence Force over the weekend and put Jay Clayton, the Director of National Intelligence, in charge of it. The name follows a September executive order telling agencies to say "super intelligence" where they used to say "artificial intelligence." Clayton becomes the administration's point person on AI, a role venture capitalist David Sacks held until March. Sacks and former Secretary of State Condoleezza Rice will serve as outside advisers.

The vice chairs show where the leverage sits. They are Federal Trade Commission Chairman Andrew Ferguson, Office of Personnel Management Director Scott Kupor and Emil Michael, the Pentagon's research and engineering chief. Vice President JD Vance, Defense Secretary Pete Hegseth and Treasury Secretary Scott Bessent are also members. The group reports to Trump and White House chief of staff Susie Wiles. It has 120 days to deliver a report on the technology's risks and opportunities, including how the government should disclose and respond to security breaches.

The charter says the force will "develop plans for responding to SI-enabled threats to our society, while preventing overregulation and regulatory capture that would stifle innovation and competition." Clayton, a former Securities and Exchange Commission chair and US Attorney in Manhattan who was sworn in as intelligence chief on August 3, has been open about his priorities. "The risk of not being first is high," he has said. Of calls for a coordinated slowdown in AI development, including from Anthropic chief executive Dario Amodei, he said: "I don't think any American should think that that's a good strategy." He has also argued that federal agencies, not civil courts, are best placed to police AI companies.

That view matters because one of his vice chairs is already doing it. The FTC has confirmed it is investigating OpenAI, Anthropic and other labs, so Ferguson now sits on both the agency examining the industry and the body charged with preventing overregulation. The task force also arrives days after six technology leaders, including Nvidia's Jensen Huang, signed a voluntary White House safety pledge on September 29 that Trump called "morally binding."

For a business owner, the practical read is timing and venue. Federal rules for the AI tools you use are now on a four-month clock, and they will be shaped by agencies whose stated priority is speed. Cities and states are not waiting. New York City's council is scheduled to take sworn testimony from AI companies today on a package that includes a mandatory kill switch. Plan for local rules to arrive first and to differ from place to place.

ChatGPT will show ads while it draws

OpenAI will begin testing a visual ad format inside ChatGPT's image generation later this month, in the United States, with an initial group of advertisers. The ads show a product in use or an experience associated with it. OpenAI says they are clearly labeled and kept separate from the image the user asked for. Free and Go users see ads, while Pro and Enterprise subscribers do not. OpenAI now puts ChatGPT's audience at 1.2 billion people a week.

The company's central promise is one sentence: "Advertising does not influence the answers ChatGPT gives." Its systems check whether a conversation is emotionally vulnerable or sensitive before an ad appears, and the brand-safety firms DoubleVerify and Integral Ad Science grade the ad environment in controlled tests without access to real conversations. Pricing has not been made public.

Europe will supply the transparency. Ads have run in 31 European countries since late August, and on August 31 the European Union designated ChatGPT a very large online search engine, with 159.1 million monthly users in the bloc. By January, OpenAI must publish a public repository covering a year of ads, including advertiser names, run times, targeting parameters and reach. For a small advertiser, that repository will be the first clear view of who is buying this audience and how. For everyone else, staff on free accounts will start seeing ads inside a tool they use for work.

A top AI model beats junior accountants on the monthly close

Mercor, a company that builds tests for AI models, gave 12 junior accountants four realistic month-end close tasks. All 12 were licensed CPAs, with an average of 5.5 years of experience. Each task required finding numbers in a company's files, running calculations and delivering a table of results. The accountants' scores ranged from zero to about 90% and averaged 37%, with most attempts taking between 30 minutes and three hours. Anthropic's Claude Opus 5 scored 100% on all 20 of its attempts and finished each task in under 10 minutes. An older model, GPT-5, scored about 69%.

The cost gap is the number owners will notice. Mercor calculates $0.21 for each graded item Claude Opus 5 got right, against $10.35 for the human accountants, a 49-fold difference.

The caveats are Mercor's own, and they are large. The tasks measured what researcher Aden Barton calls "detail-oriented instruction following," not the profession. The accountants worked without the things that make them good at the job: the chance to ask a clarifying question, knowledge of the client, and context built up over months. Client communication and judgment were not tested. The study is also Mercor's own, and it is small. Even so, it concludes that "these data do suggest substantial changes, and productivity gains, in accounting in the coming years." The reasonable takeaway for a small firm is narrow. Reconciliations, tie-outs and schedules built from clean files are now work a model can do quickly, and a person should still own what the numbers mean.

Agents are getting access to the money

Three announcements in the past week point the same way: software that can move money, not just talk about it.

Robinhood introduced Robinhood Agents, AI assistants inside its app that can analyze markets, build strategies and place trades. The design has two limits worth noting. An agent can touch only the money in a dedicated agentic trading account, and manual approval of every trade is the default, though a customer can turn it off. A coming feature called Loops will run standing instructions around the clock. Robinhood says more than 150,000 customers already use agentic trading accounts, and the new agents are due soon for eligible US customers.

Airwallex rebuilt its business accounts around the same idea for company cash. A business can see every entity and account in one view, and an AI agent can anticipate funding needs, rebalance cash between accounts, put idle funds to work and time currency conversions "within the permissions, rules, and approvals you set," in the words of chief product officer Shannon Scott.

The card industry is working on the missing paperwork. EMVCo, the standards body for card payments, has closed public comment on a framework for purchases made by agents. Today a card payment confirms that a credential is valid. It does not record what the buyer authorized an agent to do. The proposal would register a consumer's instructions before a purchase, identify the agent, and flag agent involvement in the transaction to support "risk assessment, fraud prevention, dispute resolution, operational transparency," said Oliver Manahan, EMVCo's director of engagement and operations. Shoppers remain cautious. In PYMNTS Intelligence research, 56% would let an agent search and compare products, but only 37% would let it pay. Merchants should expect agent-placed orders to arrive before the standards that explain them.

Physical AI

Driverless freight is moving from testing to a named customer. Kodiak and IKEA Supply said IKEA will be Kodiak's launch shipper for trucks with no one in the cab, starting at the end of 2026. The first driverless stretch is 219 miles of Interstate 45 between Kodiak's Houston and Dallas-area facilities, part of a 292-mile route from IKEA's Baytown distribution center to its Frisco store. The two companies spent four years getting here, running more than 1,300 test loads over 750,000 miles with safety observers aboard. Kodiak is still completing the safety case for its highway launch. Its pitch to shippers is planning: precise departure times and predictable dock arrivals, with human drivers kept on the local runs.

Drone delivery reached the hardware aisle. Lowe's began a pilot near its Matthews, North Carolina store, with Wing flying the drones and DoorDash taking orders through a "Lowe's by Drone" storefront in its app. Each flight carries about 2.5 pounds, so the menu is hand tools, paint supplies, tape, batteries and cleaners, delivered in as little as 20 minutes. Fees were not disclosed. For a contractor, the use case is the forgotten part that would otherwise cost a round trip. For a local hardware store, it is a new kind of competition for small urgent orders.

If you run mobile robots, the rulebook now has a chapter addressed to you. The Association for Advancing Automation published ANSI/A3 R15.08-3-2026, the third part of its safety standard for industrial mobile robots. Parts 1 and 2 covered the robots and the systems built from them. Part 3 covers the organizations that operate them, with guidance on risk assessment, managing changes, employee training and periodic reviews and inspections. It was written alongside the matching standard for fixed industrial robots. Once a standard assigns duties to the user, "the vendor installed it" stops being a full answer after an incident.

Two smaller items show where money is going. Hello Robot won a grant of nearly $3 million from the National Institute on Aging to adapt its Stretch robot for older adults with mild cognitive impairment and early Alzheimer's, covering reminders, social contact and safety. The three-year project with the University of Illinois Urbana-Champaign will involve 60 older adults, care partners and professional caregivers, with eight-week trials in homes and care settings. And RobCo, maker of the Alfie factory robot, passed a $1 billion valuation through a share sale that let employees cash out, doubling its value in nine months. Alfie's commercial launch is set for March 4, 2027.

What This Means for Your Business

Keep your own AI inventory while Washington writes its report. The federal task force has four months and a mandate to avoid overregulation, so national rules will come slowly and lightly. City and state rules will not wait. List every AI tool that talks to customers, makes a decision or touches money, note who can switch each one off, and review the list each quarter. That list answers most of what any regulator, insurer or large customer will ask.

Decide which staff should be on free AI accounts. Ads are coming to image generation for free and Go users. That is fine for casual use. For anyone who produces marketing images, handles client work or enters company information, a paid seat removes the ads and usually comes with stronger data terms. If you buy ads yourself, watch the European ad repository when it opens in January before committing budget.

Run your own accountant test. Take one closed month, give an AI model the same files your bookkeeper used, and ask for one structured task such as a bank reconciliation or an accruals schedule. Compare the result line by line with what you filed. If it holds up, use the model for first drafts and keep your accountant on review, exceptions and advice. Ask your accounting firm how it uses AI and whether its pricing reflects the time saved.

Give any agent that can spend its own account. Robinhood's design is the right pattern for business cash as well: a separate account with a capped balance, approval required for each transaction, and standing instructions only after weeks of clean results. If you sell online, write down how you will handle a dispute where the buyer says their agent ordered something they did not want, because the card networks have not finished their answer.

Put limits where the machines are. Turn on the cloud spending limits that exist, and add billing alerts on top, because a pause is not a hard cap. If you operate mobile robots in a warehouse or shop, get the new user standard and check your training and inspection records against it.

Sources

← All posts