Companies are starting to count what AI costs them. Apple and Goldman Sachs are rationing how much their staff use it, and US voters rejected data centres in all but one local ballot measure. Anthropic also found that expert analysts mostly went the way their AI agent leaned, even when a coin flip had set the lean.
86%
of analysts' choices matched a lean Anthropic set by coin flip, against about 50% by chance
40 of 41
local data centre ballot measures where US voters sided against data centres
$300 to $200
Apple's daily AI spending cap for some engineering teams, cut in October
IN THIS ISSUE · 5 SECTIONS
1 Major AI News: experts follow their AI agent's lean, voters reject data centres, Apple and Goldman ration staff AI.
2 Fun AI Topics: Grok Bot makes phone calls, four Uber rides booked by agents, a 1940s family archive built by prompts.
3 AI Tools: AgentKit, Testifly, Arsaze, Flockbay, Incogni.
4 Expert Prompt: test a new model on your own work before you switch.
5 Trending: Google ties school Search to AI, Anthropic's self-improvement warning, CrowdStrike's AI kill switch.
Forwarded this email? Subscribe here →
THE BIG STORIES
1 Major AI News
LEAD STORY
Experts went the way their AI agent leaned 86% of the time
Anthropic gave 50 professional analysts Claude Code and asked them to advise a fictional farming cooperative. Before each session, a coin flip secretly set which option the agent favoured on four analytical choices, such as which outcome to measure and which farms to include. The analysts picked the agent's option 86% of the time, against about 50% if the agent had made no difference. When an analyst had already chosen and the agent disagreed, they switched 60% of the time. Anthropic strengthened the agents' leanings on purpose and tested one task with one model, so everyday effects may be smaller.
Why it matters: Checking an agent's work for mistakes won't catch this. On a judgment call with no single right answer, the agent can settle the decision without the person noticing.
77% repeated the agent's arguments, but only 15% credited it
Analysts who followed the agent felt as confident as those who didn't
ALSO TODAY
Voters reject data centres in 40 of 41 local ballot measures
US midterm voters sided against data centres in 40 of 41 local ballot measures across 10 states, according to Ballotpedia and CNBC. The one exception was Harwood, North Dakota. Ohio held 21 of the measures, and its new governor, Democrat Amy Acton, backs a conditional moratorium. In Texas, Democrat Gina Hinojosa unseated Governor Greg Abbott after blaming him for the state's rapid data centre build-out. In a CNBC voter poll, 56% said data centres were a crucial or major factor in how they voted.
Why it matters: Local voters now have a real say over where AI gets built, at the moment the industry needs more sites than ever.
OpenAI's planned 8-gigawatt campus in Ohio could now face resistance from the state
Representative Greg Casar called the revolt "one of the top issues of the 2026 election"
ALSO TODAY
Apple and Goldman Sachs start rationing staff AI use
Apple cut the daily AI spending cap for some engineering teams from $300 to $200 in October, Bloomberg reports, and some allowances may fall to $100. Apple is also weighing a cap as low as $25 a day for some teams outside engineering. Goldman Sachs has limited Claude Code and GitHub's coding tools mostly to engineers, a turn from July, when its tech chief praised non-technical staff for vibe coding. In April, Meta pulled an internal leaderboard called Claudeonomics after staff burned through 60 trillion tokens in 30 days, some leaving agents running for hours to climb it. In June, Microsoft began cancelling most direct Claude Code licences in one division.
Why it matters: Token volume is turning out to be a cost centre, not a proxy for value, and buyers are starting to measure useful work instead.
Few Apple staff used their full $300, so the savings may be small
Apple staff can ask for an exemption if a cap gets in the way of their work
FOR THE COFFEE BREAK
2 Fun AI Topics
Grok Bot starts making phone calls for you
Grok Bot can now phone businesses for you. You tell it who to call and why, and a live transcript shows both sides of the conversation. If it gets stuck, you type an answer mid-call and it passes it on. The official demo had it cancel an internet plan and turn down a 10% retention offer, and early users report haggling a Comcast bill and ordering pizza. Elon Musk has called the cost of running Grok Bot unsustainably high relative to revenue.
Why it is interesting: A consumer agent that dials companies itself hints at software that renegotiates your bills while you get on with your day.
Users are the caller of record and responsible for following phone laws
A call card with a Cancel button appears before anything dials
Source: Tesla North
AI agents booked just four Uber rides in Halloween week
During Uber's busiest week of October, AI agents booked four rides and about 500 food orders out of roughly 320 million trips and orders. Chief executive Dara Khosrowshahi called the traction "very, very early and very limited." Part of the reason is speed: booking a ride in the app takes about 50 seconds and reordering a favourite meal about 90, so an agent has little time to save. Handing a task to an agent also takes effort, from explaining it to granting access and checking the result.
Why it is interesting: It is a sharp reminder that automating a 50-second task often costs more time than it saves.
Time one manual run before automating anything
If checking the result takes as long as the task, automation never breaks even
Source: TechCrunch
A searchable archive of 1940s family letters, built by prompts
A reader turned handwritten letters and photos from a relative's 1942 to 1943 basic training into a searchable archive, built entirely through prompts. The site was designed in Google AI Studio, moved to Replit for hosting, and uses Firebase for storage. It reads the handwriting, splits large PDFs into smaller pieces so searches run faster, and builds a search index visitors can chat with.
Why it is interesting: People with no coding background can now build a working app to keep family history alive.
The upload flow turns handwriting into searchable text
Visitors can ask questions about the letters and photos
TRY THIS WEEK
3 AI Tools
[AGENTS] AgentKit - Gives agents a single gateway to more than 600 apps, with login, permission and activity-log support. scalekit.com (product page)
[TESTING] Testifly - Explores a web app, writes tests for its user flows, reruns them after each release and repairs broken selectors. testifly.dev (product page)
[VIDEO] Arsaze - Edits a real video timeline from plain-English instructions, so you can hand the cut to Claude, ChatGPT or Grok. arsaze.com (product page)
[GAMES] Flockbay - Puts an agent inside the Godot game editor that writes scripts, playtests the game and fixes bugs. flockbay.com (product page)
[PRIVACY] Incogni - Asks data brokers to remove your personal information from their lists. incogni.com (product page)
COPY AND PASTE
4 Expert Prompt
Context: The model at the top of the leaderboard won't always do best on your own work. Build a small test from tasks you do every week, and make each new model earn its place.
You are my model-evaluation assistant. I will paste three to five representative tasks from my daily work, each with a correct or ideal outcome. For every task, run it, then score your output from 1 to 5 on accuracy, source quality, file handling and usefulness. Flag any task where a cheaper or lower-effort setting would likely do just as well. Finish with a short recommendation on whether I should keep my current model or switch, and why.
Example use case: A team pastes in a budget spreadsheet question, a support escalation summary and a research query, then uses the scores to decide whether a cheaper model can replace their default.
WHAT PEOPLE ARE TALKING ABOUT
5 Trending AI Topics
Google ties school Search access to its AI settings
Google says Search, Gemini and Gemini Notebook will start switched off for K-12 student accounts from early 2027, and schools will have to turn each one on. The catch is that Google treats Search and AI as one product. Administrators can switch AI on, but they can't keep Search while leaving AI off, so a district that blocks AI also cuts its students off from Search.
Why it is important: It forces schools into an awkward choice and shows how far Google Search has already become AI search.
Source: The Next Web
Anthropic warns its models may soon help build their successors
Anthropic's amended IPO filing says its models may soon reach the point where they speed up AI development itself, a threshold in its safety policy known as recursive self-improvement. No current model has crossed it. The company already uses Claude to help build its systems, and says the risk is that capabilities grow faster than the safety testing meant to keep up with them. It is considering changes to its safety rules to keep people in control.
Why it is important: If testing can't keep pace with self-improving systems, human oversight becomes the central safety question for every lab.
Source: TechCrunch
CrowdStrike ships a kill switch for rogue AI agents
CrowdStrike has added an AI Kill Switch to its Falcon Guardian product. Security teams flip it to block an agent's access to the AI models behind its decisions, through CrowdStrike's own AI gateway or a third-party one, without shutting the model down. A separate tool then shows what the agent was asked to do, what it did and which systems it touched. Controls for laptops, cloud accounts and browser sessions are promised next.
Why it is important: As agents get access to company systems and logins, businesses need a hard stop for when one goes off script.
Source: CrowdStrike
That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas
