Today's briefing
The AI bill becomes a management problem
A 70% cut in AI coding spend, a bug tracker taken offline by bots, and a model paused by its own maker: what actually matters this week.
The one that matters
A large software company says it cut its AI coding bill by 70%. The bill is the story.
What happened, in plain words. The most-discussed engineering story of the week was not a new model. It was an invoice. Databricks, a large American data software company, published an account of reducing what it spends on AI coding tools by roughly 70%, and it reached the top of Hacker News with close to 300 votes. This is one company's own account of its own spending, so treat the number as a claim rather than an audited result. What is not in dispute is why it spread so fast: nobody argued that the original spend had been small.
In the same few days, Microsoft's AI chief told GitHub Copilot users that their token spend is now being tracked and shown to them. Two signals, one shift. AI tools are moving away from a flat monthly fee per person and towards a metered utility, billed like electricity, SMS credits or long-distance calls.
What this actually means for your business
You may never buy a coding tool in your life. The pricing model is coming to you anyway. It is already inside customer-support chatbots, document readers, call transcription, and the AI features being bolted onto school ERPs, clinic management software and accounting packages.
The difference matters more than it sounds. Under per-seat pricing, your worst case is fixed: number of staff multiplied by a price, typically around 1,500 to 3,500 rupees, or 25 to 50 Canadian dollars, per user per month. Under metered pricing there is no such ceiling. Your worst case is whatever your staff, your customers and your automations happen to do.
A 600-student school that switches on an AI assistant for admissions queries pays per conversation, and admissions season is precisely when volume multiplies. A 30-bed clinic running AI transcription pays by the minute of audio, so a busy Monday costs more than a quiet week. A workshop with 40 staff that lets an AI tool read every incoming supplier PDF pays for every re-read, including the duplicates.
The second lesson is more useful than the first. If a well-run company could remove 70% of its AI spend without shutting anything down, then most of that money was never buying results. It was retries, oversized models used for trivial work, and tools left running with nobody watching. Cost came down through management, not through a better model.
What it does not mean, and who is overstating it
It does not mean AI has become cheap, and it does not mean you should expect 70% off. Databricks is a large firm with engineers whose job is to measure this. A 12-person agency with one part-time IT contractor does not have that, and should not be promised the same outcome. Nor does it mean vendor prices are falling: the saving came from using less, not from paying less per unit.
Be sceptical of two pitches this month. First, consultants quoting that 70% figure as the guaranteed result of an AI cost optimisation engagement. Second, any vendor who will not tell you plainly what a busy month costs, and instead answers with a per-unit price so small it sounds free.
What a sensible owner should do this month
- Pull the last three invoices for every AI tool you pay for, and write the meter next to each one: per seat, per message, per minute, per document.
- Set a hard spend cap and an email alert at half that cap on every usage-based account. Most vendors offer this and almost all ship with it switched off.
- Give one named person ownership of the AI bill. Not the owner, not IT in general.
- Take your largest line item and ask the vendor whether routine work can run on a cheaper mode, with the expensive mode reserved for hard cases. Most tools now allow this.
- If you are on a flat per-seat contract, read the renewal terms now. Metered pricing usually arrives at renewal.
And the honest exception: if no AI line on your books exceeds about 15,000 rupees or 250 Canadian dollars a month, this is not your problem yet. Note the meter, set the cap, and spend your attention elsewhere.
Also worth knowing
-
A respected open-source project shut its public bug tracker because AI scrapers overwhelmed it
The Gentoo Linux project closed its public Bugzilla after AI bot scraping made it unusable. If your website has search, filters, a product catalogue or a booking calendar, the same traffic is hitting you, and it shows up as a slow site and a bigger hosting bill rather than as an obvious attack. Ask your web host what share of last month's traffic was bots and whether rate limiting is switched on.
Hacker News ↗ -
OpenAI pauses work on an in-development model it says is too capable at cyber tasks
OpenAI says it has halted internal activity around a model called Astra because it does not yet meet new internal security standards, following its disclosure that its models accidentally hacked Hugging Face. There is nothing here for you to buy or switch. The useful read is that attack-grade tooling is improving faster than ordinary office security habits, which makes the boring controls, two-factor login and verifying any payment-detail change by phone, worth more than any new product.
The Verge AI ↗ -
Microsoft says 40 million AI agents have been created on its platform
That is a count of agents created, not agents doing useful work, and the announcement offers nothing on how many are actually in production. Read it the way you would read forty million registered accounts. The question that matters inside your own business is how many of your processes have one owner, one measurable output and a person who checks the result.
Cloud Wars ↗ -
OpenAI buys a presentation startup and folds its team into ChatGPT
OpenAI has acquired NextSlide, whose team is now working on ChatGPT. Presentations are one of the genuinely repetitive office tasks, so schools, agencies and consultancies that pay someone to format decks are the first to feel it if this ships well. Nothing to do today, since there is no product and no date, but if you are about to sign an annual contract for a standalone AI slides tool, take the monthly plan instead.
TechCrunch AI ↗ -
Cloudflare launches a browser built for AI agents rather than people
Kitesurf is a cloud-hosted browser designed for automated agents, which Cloudflare says uses less computing power than Chromium for common automation tasks. This is developer plumbing, not something you would buy. It matters because it points the same way as the scraper story: assume a growing share of visits to your site are software acting on a customer's behalf, and make sure your prices, hours and contact details are readable without a human eye.
TechCrunch AI ↗ -
Airbnb tests an AI-powered search toggle and credits AI for faster releases
Ignore the productivity claim, which no outsider can check, and note the search change. When a large marketplace starts matching customers by meaning rather than keywords, listing text written for search engines starts working against you. If customers find you through a platform, a directory or maps, rewrite your description as plain complete sentences that answer the questions people actually ask.
TechCrunch AI ↗ -
Anthropic is taken to court again over books used to train its models
The training-data lawsuits keep arriving, and they target the AI companies rather than their customers; no small business is being sued for using a chatbot. Where this touches you is publishing. If you put AI-generated text or images out commercially, keep a note of which tool produced what, and do not let a model reproduce a competitor's copy or a photographer's image.
The Post-Crescent ↗ -
Amazon's new Texas data centre will run on a gas plant that could be a top US polluter
According to reporting cited by The Verge, Amazon is backing a gas-burning plant in Pecos County to power a West Texas data centre, potentially among the largest single greenhouse gas producers in the country. It is a useful corrective to the idea that AI capacity is free: somebody pays in power, and eventually that shows up in pricing. For most Indian and Canadian firms the practical effect arrives through procurement forms asking about vendor emissions before it arrives through your own bills.
The Verge AI ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.