Issue #1 · week of September 8, 2026

Top 10 things in AI this week — for a business that has real work to finish

Dark illustration: a glowing green hexagon node on a navy field, linked by faint lines like a week of stories.
Issue #1 header. Credit: Theos AI.

Every Monday I am going to write this the way I would tell it over coffee: what actually moved in AI last week, and what a regular business should do with it.

I am not ranking this list by who shouted the loudest. I am ranking it by what I would care about if I ran a shop, a clinic, a contractor, or a five-person office. New models matter when they land in tools you already pay for. Big humanoid deals matter later. Outages matter the week they happen.

This is Issue #1. I read the primary sources myself. Where a company used Claude, Copilot, or an agent to do real work, I put that near the top.

How I ranked this. Can you use it in tools you already have, does it change how you handle client files, or can you ignore it until next year? Try · Watch · Ignore.

  1. GPT-6 Astra is rolling into ChatGPT for work

    Try

    What happened

    On September 3, OpenAI released GPT-6 Astra and started rolling it out to ChatGPT Plus, Pro, Business, and Enterprise, plus the API, Microsoft Azure, and AWS Bedrock. They call it their most capable and aligned model yet — stronger at computer use, browsing, and the kind of professional work people already dump into ChatGPT.

    Why you care

    If your team already lives in ChatGPT for drafts, research, or messy computer work, the tool just took a step. You do not need a new vendor. You need a clearer job for it, and a person who still hits send.

  2. Nationwide is using Copilot on the boring office work

    Try

    What happened

    On September 8, the UK building society Nationwide described rolling Microsoft 365 Copilot across the workforce while it stitches together Virgin Money. In finance, about 1,000 people with premium licenses use it to turn rough notes into first drafts, summarize meetings, shape messages, and pull actions. In an internal poll, 46 percent of finance respondents said they use it every working day.

    Why you care

    This is not a science-fair demo. It is the same work you have: notes, meetings, follow-ups, two companies being glued together. The useful pattern is simple — use the suite you already have, on the jobs that get copied around all day.

    Official Microsoft 365 explainer. Credit: Microsoft on YouTube.

  3. monday.com showed how they keep a production agent in the lane

    Try

    What happened

    On September 9, monday.com's engineering team published how they run feedAgent in production — an agent that curates each person's activity feed. The lesson they keep repeating is the harness: strip private data before the model sees it, force a structured answer, drop made-up citations, and evaluate the thing like software instead of hoping the model behaves.

    Why you care

    A lot of "AI agent" pitches skip the bumpers. monday.com is showing the unglamorous version that actually ships: one job, tight inputs, a checklist after the model talks. That is closer to how a small team should start than "let an agent run the business."

  4. Claude's new enterprise deal: your logs can stay in your cloud

    Watch

    What happened

    On September 1, Anthropic announced Enterprise Frontier Safeguards, designed with more than 100 customers in banking, healthcare, law, and companies such as Comcast, KPMG, Mastercard, Salesforce, and Visa. Activity used for safety monitoring can live in your own Amazon S3, Azure Blob, or Google Cloud Storage, under your keys. It rolls out in phases this fall. Eligible customers get zero data retention on Fable 5 and Fable 5.1 until their turn comes.

    Why you care

    If you handle client files, the question is no longer "is Claude smart?" It is "where do the logs live, and who can read them?" A local CPA, clinic, or law office should ask the same thing before pasting a customer file into any chatbot.

  5. Google dropped Gemini 3.8 Flash into Workspace

    Try

    What happened

    On September 2, Google released Gemini 3.8 Flash — its third Flash model in six weeks — aimed at agent work, coding, and multi-step reasoning. It is in Gemini Enterprise, the Gemini app for Pro and Ultra subscribers, AI Mode in Search, and Gemini in Google Sheets. A Cyber variant is limited to trusted defenders.

    Why you care

    If your business already lives in Gmail, Drive, and Sheets, this is the update that lands where you work. The fast models are the ones you can actually use every day on invoices, quotes, and cleanup — not the ones that win a benchmark chart.

    Official Google DeepMind walkthrough. Credit: Google DeepMind on YouTube.

  6. Thursday's outage: ChatGPT, Claude, and Grok all hiccuped

    Watch

    What happened

    On the morning of September 3, ChatGPT, Claude, and Grok were down or erroring for a lot of people on the web and on iPhone. OpenAI's status page flagged ChatGPT and Codex. Anthropic reported elevated errors on Claude. They came back. The morning still happened.

    Why you care

    If the only copy of a customer reply lives in a chatbot window, an outage is a closed shop. This is not theoretical. Helpers go down. Your filing cabinet should not.

  7. Figure signed a $3.5 billion Nvidia-scale compute deal for humanoids

    Watch

    What happened

    On September 3, Figure and Nscale announced access to as many as 100,000 Nvidia Vera Rubin GPUs, with the first systems targeted for Barstow, Texas in the second half of 2027. Figure called it an initial $3.5 billion compute commitment, with intent to go past $6 billion. Jensen Huang called it the "physical AI flywheel" — train on Nvidia, test in simulation, run on GPUs inside the robots.

    Why you care

    This is Nvidia-scale money moving into robots that walk through spaces designed for people. It tells you the industry is serious. It does not mean you should lease a humanoid for the shop floor this quarter.

    Official Figure clip of Helix on package work. Credit: Figure on YouTube.

  8. XPeng's IRON walked itself off the factory line

    Watch

    What happened

    On September 8 in Guangzhou, XPeng said the first IRON humanoid finished automated final assembly and walked itself off the line. The company is aiming at scaled production by the end of 2026, first use in its own stores and campuses, and commercial deliveries in 2027. Three on-board Turing chips run the model on the robot, so it does not need a remote operator for basic work.

    Why you care

    This is the first time a car company is treating a humanoid like a product that leaves a factory, not a stage demo. Still not something you buy for a Wisconsin shop this year. Fine to be curious. Ignore anyone selling "they're here next month."

  9. Nvidia's SONIC controller: tell a humanoid what to do in English

    Ignore

    What happened

    Researchers at Nvidia published SONIC in Science Robotics, covered this week on September 9. It is a controller — the robot's motor system — that turns VR, video, or language into whole-body motion, including moves it was not trained on. The next step they name is tying it to Nvidia Isaac GR00T so the robot can think and move in one loop.

    Why you care

    This is how robots will eventually take "put that box on the shelf" as an instruction. It is not a product you can hire. Unless you build robots for a living, it is a bookmark.

    Official NVIDIA explainer of Isaac GR00T. Credit: NVIDIA on YouTube.

  10. Jensen says AGI arrived. OpenAI's chief scientist says slow down.

    Watch

    What happened

    On September 6, Nvidia CEO Jensen Huang posted that Astra was trained on more than 100,000 Grace Blackwell chips and wrote "AGI has arrived," with 400,000 more GPUs coming online next. The same day, OpenAI chief scientist Jakub Pachocki published "An Alien Mind": no lab has solved alignment well enough to keep flooring it, and he hopes voluntary slowdowns become normal until there are shared safety bars.

    Why you care

    The chip vendor and the lab are not telling the same story. For a local business the takeaway is simpler: the tools are getting more powerful and less predictable. Keep a person in the loop on anything that touches a customer, a payment, or a medical or legal file. Use the new models. Do not believe the word AGI means your receptionist can go home.

A few questions I keep getting

Do I need to switch tools every time a new model drops?

No. I would stay in the suite you already pay for — ChatGPT, Copilot, Gemini, or Claude — and give the new model one repetitive job with a human on the send button. Switching tools every Monday is how projects stall.

What should a small business actually try this week?

Pick the one workflow that gets copied, checked, and passed around all day. Meeting notes to a follow-up list is a good start. So is a first draft of a quote, or cleaning a messy spreadsheet. If that sounds like the constraint in your business, we can look at it on a 30-minute call.

Why isn't there an email signup?

This lives on theosai.net. Come back on Mondays, or subscribe to the RSS feed if you want it in a reader. I would rather write the page than collect a list.

What's the one workflow holding the business back?

If something on this list sounds like the constraint, book a 30-minute call. We'll look at the repetitive work and see if a 7-day sprint is a fit. Or email me at jed@theosai.net.

No pitch deck. No public pricing on this site. We'll talk about the work, then decide if it's a fit.