The priciest AI setup performed worst on invoices, and nine in ten firms report no productivity gain

An MIT test gave the same 44 supplier invoices to competing AI agent setups. The most expensive setup got 91% of them right. A mid-priced build got all 44.

The costly setup was not worse at reading an invoice. It was assembled to be strong at every step, and an agent that resolves a mismatch by re-reading the purchase order and then the invoice again pays for every reading. Teams that cut the price of one component, a cheaper model or less of the document fed in at a time, then paid for more attempts at the same invoice.

Rippling ran 2,100 scored agent runs per model against its own payroll data. In SaaStr's account of the results, the cheapest model tied the most expensive one.

Research teams at the Atlanta Fed, the Bank of England, the Bundesbank and Macquarie University put identical questions to nearly 6,000 senior executives in the US, UK, Germany and Australia, from November 2025 to January 2026. They found 69% of firms actively using AI, and nine in ten executives reporting no effect on productivity or employment at their own company over the past three years. PwC's April study of 1,217 senior executives found that the companies seeing real financial returns were twice as likely to have redesigned how work gets done instead of adding AI to the work as it stood.

More than two-thirds of the executives in that survey use AI themselves, an average of 1.5 hours a week.

A website can now hand ChatGPT a written list of what it is allowed to do: search the catalogue, check a delivery date, take a booking, complete a purchase. That is the new WebMCP standard, and OpenAI has added support for it. Until now, an agent sent to buy something had to click through the pages the way a person does, working out the buttons as it went.

An agent clicking through a checkout built for human eyes stops at the ordinary obstacles, a dropdown that reloads the page or a date field that rejects its format; an action the site publishes for it has one address and one answer.

Agent traffic to websites grew 7,851% in 2025, on sites that publish nothing of the kind, and the common response has been to block them, at three times the earlier rate. In a server log, a blocked agent and a blocked scraper are the same line; one of them was shopping.

Of that agent traffic, 79% arrived on product pages and search results, the two pages a buyer reads before buying.

Anthropic has put a browser inside the Claude desktop app. Its Cowork assistant opens a window of its own, signs into a website, clicks through the forms and pulls data out of pages that have no software connection to anything else.

That covers a large part of an operations team's week: the supplier dashboard where prices are updated by hand, the insurer's claim portal, the government filing site, the ageing inventory tool that only opens in a browser. Ready-made connectors exist for Gmail and Salesforce. None exists for a regional carrier's rate portal, and someone logs into that one every Tuesday.

Claude's desktop app is an ordinary subscription, and it can now read a screen, type into a field and download a report with no connector at all. What it cannot supply is the judgment about whether the figure it copied belongs in the box it typed it into, and a page it reads can also carry a sentence written to be taken as an instruction, which Claude has no reliable way to separate from its user's own words.

Whatever Claude signs into in that window, it never sees the tabs and saved passwords in the person's own browser.

Plaud One is a pair of earbuds at $249 whose charging case carries its own cellular connection, so it records and transcribes a meeting or a phone call with no phone nearby.

California, Illinois and Pennsylvania are among the US states where every person on a call has to agree before it is recorded, which makes a recorded client call or job interview a legal problem rather than an awkward one. The audio and its transcript go to the vendor's servers, and so does whatever pricing or salary detail was said out loud.

Relay, a London company, released dictation software this week that also acts inside other apps, sending a message or putting a meeting on the calendar, which means it reads what is on the screen. Its own microphone arrives in early 2027 at $150, plus $15 a month.

The class action against Workday covers more than a billion job applications its screening software processed. The employers who used it carry a duty of their own: Illinois requires them to tell applicants when AI is involved in a hiring decision, and Colorado's law takes effect in 2026.

School district HR teams are using the tools on the paperwork anyway. AI makes the first pass through applications against the same written rubric a human committee then applies, and the committee keeps the final decision. The rubric is what an employer can produce later when an applicant asks why.

One district's HR team also turned the tools on its own records, cleaned up 40,000 employee files and counts more than $300,000 saved this year.