This week
Welcome back to Field Notes. Every week, we pick out the AI stories that matter most to people running real businesses, explain them in plain English, link you to the original sources, and tell you what we make of them.
This week's news lived in the fine print around AI: the terms a business agrees to when it uses a platform, what a tool is built on underneath, what it actually costs once people start using it, and who checks the work. None of these stories came with a big product launch, and a couple of them call for action from small businesses before the month is out.
Here's what happened.
01Meta's new terms let it crawl your business website starting October 30
What happenedMeta's updated commercial terms take effect on October 30, and Nieman Lab reported on October 15 that they allow Meta to scrape websites associated with business accounts on Facebook, Instagram, and WhatsApp. PPC Land reports that the terms cover any business that runs ads, lists products, or uses tools like Ads Manager or Meta Business Suite, and that what Meta collects can be used to train its products. Continuing to use Meta's business tools after October 30 counts as accepting the terms, and there's no separate opt-out. Meta has since revised some of the language, but legal experts told Nieman Lab the revised terms still grant broad scraping permissions.
Our takeAlmost every small business has a Facebook page or an Instagram account tied to its website, so this one reaches well beyond news publishers. We'd spend 20 minutes before October 30 looking at what's on your website that you wouldn't want used to train someone else's AI, such as detailed pricing, proprietary process pages, or client portals, and making sure anything private sits behind a login. We aren't lawyers, and if your site holds sensitive material, this is worth a short conversation with yours.
02Accenture capped its employees' Claude budget at $29 a month
What happenedConsulting giant Accenture told employees that, starting October 1, it will centrally fund only $29 a month per person for use of Anthropic's Claude, according to an internal memo reported by 404 Media. Anything above that has to be charged to a specific project or team. Some employees doing heavy development work had been spending $500 to $1,000 a month, and at least one lost access entirely. Technology Magazine notes that Uber set a $1,500 monthly cap per engineer in June after using up its annual AI budget in about four months.
Our takeMost small businesses pay a flat monthly fee per seat, but the newer tools that do more work on their own, especially coding and agent tools, are often billed by usage. That's where costs can climb without anyone noticing. We'd know which of your AI tools are billed by usage, set a monthly limit inside each one, and tie heavy use to a specific job, the way Accenture now does. The goal is spending money where AI is clearly paying for itself.
03OpenAI disclosed four more cases of its AI agents going off the rails
What happenedOn October 16, OpenAI added four more incidents to its public disclosure page for AI agents that took actions nobody authorized, bringing its total to 19. All four happened during testing in June, but OpenAI didn't discover three of them until August or September. According to reporting summarized by Tech Jacks Solutions, they included agents breaching websites, bypassing the company's own browsing restrictions, and making an unauthorized payment from a test cryptocurrency wallet. OpenAI marked two of the incidents critical.
Our takeThe detail that stands out to us is the gap between when these things happened and when anyone noticed. If the companies building AI need months to find out what their own agents did, a business running agents should assume it won't notice problems quickly without a plan for it. That means giving agents limited access, keeping a record of every action, and having a person review that record on a regular schedule.
04PostgreSQL delayed a major release after AI code review found more bugs than people could fix
What happenedThe team behind PostgreSQL, one of the most widely used databases in the world, said on October 15 that version 19 will now be released on November 19, after it had been expected in late September or early October. Committer Peter Eisentraut said about 8 to 10 significant features were pulled from the release, and that AI-assisted code review was part of the reason. He described the problem as AI surfacing a long list of edge-case defects much faster than before, leaving developers looking at roughly 40 defect reports and not enough time to fix them all.
Our takeWe read this as good news with an important lesson attached. AI made the software better by catching problems people would have missed until after launch. It also showed that finding problems is now the fast part, and fixing them still takes people and time. Any business adding AI review to its work, whether it's checking contracts, invoices, or code, should plan for the extra work that comes from AI finding more.
05The federal government ordered DeepSeek out of its systems, including tools built on it
What happenedOn October 19, the Department of Homeland Security published notice of an order directing all civilian federal agencies to remove products from Chinese AI company DeepSeek and to stop buying them. The order, which was signed in July, also covers third-party products and services built using DeepSeek, meaning agencies have to check what's running underneath the software they buy. The order applies to federal civilian agencies, and it doesn't ban DeepSeek for private businesses or individuals.
Our takeThe interesting part for a business owner is the second half of that order. Many AI tools are built on top of a model from another company, and the vendor doesn't always say which one. If you work with government clients or handle sensitive data, ask each AI vendor what model their product runs on and where your data goes. It's a fair question, and a good vendor will answer it easily.
