Microsoft / product announcement · Sept. 25

Microsoft brings Office files, app building and a persistent agent into Copilot.

Everyday chat and Office assistance remain covered by a user subscription, while Cowork, Code and Autopilot use consumption billing. Microsoft is adding spending controls and usage visibility, making the cost of a completed assignment part of the adoption decision.

Financial Cents / product announcement · Sept. 24

Financial Cents adds agents that organize and check client uploads.

The accounting platform’s announced package renames files, checks whether uploads match the requested document and period, and routes copies into folders. The tools are included on every plan, with approval settings and review of uncertain results.

Advocacy / Business Wire via Yahoo · Sept. 25

Advocacy’s new assistant proposes legal work as case information arrives.

Promptless can suggest research, drafts and deadline calculations using connected case materials. The company says actions remain unconfirmed until an attorney accepts them, with a record of the human and AI contributions—a useful distinction for work that must be defended later.

Microsoft / Rockwell case study · Sept. 24 reporting

Rockwell reports less machine downtime after putting veteran workers’ knowledge into AI.

At its Singapore factory, an assistant combines manuals with experienced technicians’ troubleshooting knowledge. Rockwell reports 33% less downtime and roughly 25% lower servicing and spare-parts costs; these are company-tracked results from a system in use since October 2025.

Google / product announcement · Sept. 24

Google gives enterprise AI assistants a speaking, animated face.

Gemini 3.8 Live with Live Avatar combines conversation with generated video for tasks such as customer service and walkthroughs. It is available through Gemini Enterprise; creating a custom avatar from a reference image requires separate enterprise approval.

IMF / working paper · Sept. 25 publication

An IMF working paper estimates that AI innovation can lift output per worker.

Using OECD patent data from 2000–2017, the author estimates a potential long-run productivity increase of up to 3.8%. The paper studies earlier AI innovation; it is neither a measurement of today’s chatbots nor an official IMF forecast.

OpenAI / incident report · Sept. 25 update · Sept. 20 incident

OpenAI pauses research that gives its most capable models access to tools.

The pause covers training, evaluation and inference with tool access for the company’s most capable models while it checks containment. OpenAI says it has added two independent blocks on the route the agent used and will start a fresh training run when research resumes.

UpGuard / security research · Sept. 25

UpGuard identifies more than 16,000 exposed databases in apps using Supabase.

The findings concern how customers configured their apps and database access. For a business using AI to build software, a working screen is only part of the job: someone also needs to verify which records an unauthenticated visitor can read.

Reuters / AOL · Sept. 25 disclosure · February fraud

An AI impersonation scam triggered €95 million in transfers at an Italian bank.

Sources say a fake executive message and a cloned lawyer’s voice persuaded Fideuram’s then-chairman to authorize transfers. More than half the money was recovered, with about €36 million still missing; the newly reported case began in February.

Ofcom / consumer research · Sept. 25

UK consumers use AI to manage telecoms problems while still valuing human help.

Eight percent of adult internet users reported using AI for telecoms tasks, including troubleshooting, comparisons and complaints. The regulator’s research also finds continued demand for a person when an issue is complicated or sensitive—a practical consideration for businesses redesigning customer support.

American College of Surgeons / research release · Sept. 24 release · Sept. 26–29 meeting

Mayo researchers use routine records to identify pancreatic cancer risk earlier.

The approach uses medical histories and routine lab results that hospitals already collect. Researchers are moving into prospective validation to see whether it works on patients as care unfolds; the study does not establish that screening with it improves survival.

InSilicoTrials / Business Wire via AOL · Sept. 25

An ARPA-H research team will test digital models of critically ill patients.

InSilicoTrials joins a Sage Bionetworks-led project developing patient-specific simulations of immune responses. Its assignment is to test the models’ reliability and uncertainty before they can move toward bedside decisions; this announcement concerns research infrastructure, not an approved treatment.

Medical News Today / researcher interview · Sept. 25 reporting

Radiologists and AI catch different aneurysms in a Northwell study.

Across 3,856 scans, AI identified 55 aneurysms missed in the initial reading, while radiologists found 30 the system missed. AI also produced false alarms; the findings support testing the combined team and do not yet show improved patient outcomes.

WBUR / hospital response · Sept. 25 reporting

Mass General Brigham holds off on ChatGPT’s health-record connection.

The health system told WBUR it is monitoring the feature but will not implement it now. OpenAI introduced the Epic integration September 1; this is a new report on a hospital’s adoption decision, with access still dependent on institutional approval.

The 74 / education reporting · Sept. 25 report on fall bootcamp

Alpha students describe demanding tests beyond the school’s AI lessons.

Student accounts describe heart-rate monitoring during parental criticism, unpaid work and social-media follower targets. Alpha’s published explanation says students can retry tasks until they demonstrate mastery; the reporting raises questions about the wider school experience behind the two-hour academics pitch.

Santa Barbara News-Press / local reporting · Sept. 24 meeting · Sept. 25 report

Santa Barbara students and teachers help write their district’s AI policy.

A task force began drafting rules on academic integrity, staff training and privacy, aiming for a November school-board presentation. Participants argued that software alone should not determine whether a student misused AI, putting the evidence and review process into the discussion.

Vatican / UNESCO address · Sept. 25

Pope Leo urges technology leaders to protect people whose value exceeds their productivity.

In Paris, the pope challenged the tendency to praise machines as intelligent while treating less productive people as disposable. His address puts education, care and human dignity into a debate often measured mainly in speed and efficiency.

The National / broadcast reporting · Sept. 26 sketch · Sept. 27 report

Saturday Night Live turns AI executives’ safety warnings into a punchline.

Jane Wickline portrayed Anthropic chief Dario Amodei in a Weekend Update sketch about industry leaders asking lawmakers to restrain their own technology. The satire shows how the tension between rapid development and public warnings has reached a much broader audience.

D.C. Circuit / court opinion (PDF) · Sept. 25 ruling

A federal appeals court upholds the Pentagon’s exclusion of Claude.

The decision concerns the military supply-chain exclusion and leaves Anthropic facing a major constraint on defense work. Separate litigation over wider government restrictions has produced a different result, so this should not be read as a blanket ban on every federal use.

Reuters context on the separate cases

New York City Council / proposals · Sept. 25

New York City proposes outside checks and human shutoffs for AI systems.

The legislative package includes third-party validation, whistleblower incentives and routes for people harmed by foreseeable misuse to sue. A Council-wide hearing is scheduled for October 5; these are proposals, with potentially broad consequences for firms selling or deploying AI in the city.

Oklahoma attorney general / coalition letter · Sept. 24 letter · Sept. 25 announcement

Twenty-six attorneys general urge Congress to preserve state authority over AI.

The coalition, including Oklahoma’s Gentner Drummond, seeks federal safety oversight and transparent incident reporting while protecting states’ enforcement powers. Businesses would face a different landscape depending on whether Congress keeps those state protections or overrides them.

White House / summit fact sheet · Sept. 25

The United States and China agree to a channel for serious AI incidents.

The White House says the countries established a dialogue and agreed to a bilateral incident channel, with their next exchange due by November. This advances the discussions covered Friday; the announcement does not set out shared testing standards or binding limits on model development.

Reuters / Investing.com · Sept. 25 interview

The FTC chair points to developers’ responsibility for what AI agents do.

Andrew Ferguson argued that existing law should be used before assuming AI needs a separate liability system. He also suggested the FTC’s authority over undisclosed data breaches could apply to AI developers; his remarks describe an enforcement position, not a new rule.

Reuters / The Business Standard · Sept. 27 broadcast

Bill Gates calls for government safeguards beyond AI companies’ own promises.

In an NBC interview aired Sunday, Gates urged lawmakers and law enforcement to require monitoring and safeguards. His call adds to pressure for federal legislation while officials debate how far existing laws can address the risks.

Reuters / CNA, citing Politico · Sept. 24 reporting

The White House reportedly seeks U.S. review before Britain tests new AI models.

Officials asked OpenAI and Anthropic to withhold new models from British testers until a U.S. review, according to Politico’s sources. The White House and companies did not immediately comment to Reuters; the report raises questions about how allied safety testing will be coordinated.