Do this: Your AI bill is not one number. Some jobs quietly run on the newest top-tier model while an older one would do. Two minutes gives you the downgrade list.
PASTE THIS
Act as a cost auditor for a small business that is not technical. Here is every AI job we run and its tool. For each, reply in one line: does this job need the newest top-tier model, or would a cheaper or open model do the same work? Then name one free model to test and the single test to run. Rank the list by what we would save in a month. If unsure, write 'test it' rather than guessing.
You'll get: A ranked downgrade list, one test per job: move the repetitive half of your AI work to a cheaper model, keep the top tier for the work that earns money, and run a test this week.
Two files that tell the crawlers what is not theirs
Put this in the robots.txt file at the root of your website. It is a request, not a lock: crawlers that behave will stop, and the ones that ignore it were never going to ask. User-agent: GPTBot Disallow: / User-agent: ClaudeBot Disallow: / User-agent: Google-Extended Disallow: / User-agent: CCBot Disallow: / User-agent: PerplexityBot Disallow: / User-agent: Bytespider Disallow: / Then add one line to your terms page and your invoice footer: Content on this site is not licensed for AI training. Licensing enquiries: [your email]. No developer, no plugin, under an hour. Refresh the list each quarter, because crawlers rename faster than terms change hands. The maintained list sits at github.com/ai-robots-txt/ai.robots.txt.
Vendors rarely raise the price now. They rewrite the terms and let the default do the work. Meta's new commercial terms arrive on 30 October, and every business page falls under them. The clause nobody reads is the one that costs money. Read one default this week and change it.
Meta's commercial terms take effect on 30 October 2026. They let it run crawlers, bots and spiders across the sites tied to your business accounts, and use what they collect to train its AI models, The Register reports. Ads, catalogues and Business Suite pages are in scope.
So what: before 30 October, open the terms on every Meta account that carries your brand, note what you would not want used for training, and move the master files off the platform. Do it this week, while the decision is still yours.
The cheap model caught up with your bill
Open-weight models handle 71.5% of the tokens through Vercel's AI gateway, for close to a quarter of the spending, The Register reports. Nvidia's new RTX Spark laptops, on sale today, run models on your desk instead of renting them by the token. So what: this week, list your five most repetitive AI jobs, ask which model each needs, and move one to a cheaper option before your next invoice.
Your AI vendor can drop you overnight
Replit's chief executive said this week that his company no longer works with Anthropic, and that cost was part of it, TechCrunch reports, while Anthropic woos investors ahead of a stock listing, the BBC reports. Neither needed your consent. So what: this week, write down which tools depend on a single model supplier, and keep one free alternative for the two jobs you cannot pause.
AI access is being rationed on both sides
OpenAI, Meta and Google want agents to browse and buy for you, but popular services are refusing to let them in, The Verge reports. Elsewhere, Paysend limits general-purpose AI tools to high performers who can show what the tools add, as other UK employers push them out to everyone, CNBC reports. So what: this week, decide who gets the tools, and put a human check on anything an agent buys or sends.
This week's question: we issued AI licences to the whole team and half of them sit unused. What now? Hand them to the people who actually use them. Pull the usage list for any seat you pay for, see who opened the tool in the last 30 days, and cut the rest at renewal. A licence nobody opens is shelfware, and the two people whose output you would notice will get more from it than a whole floor of casual users. Then write one rule for the tool everybody keeps: what may be pasted in, and who approves what leaves the building.
Do you know which model your busiest AI tool runs on?
A. Yes, and we know the price
B. No, whatever the tool picked
C. We know the tool, not the model
D. We test a cheaper one most months
Reply with one letter — results next issue.
DeepSeek pushed out an updated model this weekend that it says is faster, leaner and gives shorter answers. What is the version called?
Answer tomorrow — and no, we are not printing it here.
YESTERDAY'S ANSWER
China is weighing an export-control list for AI. How many technologies would it add, and which two kinds does the report name?
23 technologies, including large language models and chip designs.
About 5 minutes to read. One thing you can use today every issue.
Every Friday the desk answers one reader's AI question in depth. Reply with yours — the best one gets the full writeup.
Know an owner drowning in their inbox? Forward this — the 2-Minute Win is worth 2 minutes.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.
8735 Dunwoody Place STE R, Atlanta, GA 30350