Reading the money behind the tools you buy

The common thread today is money. Several publishers reported the scale of what artificial intelligence is now earning and costing, while others shipped features that decide whether you can trust what AI produces inside your daily tools. Both halves matter when you choose software, because the vendor's finances and the product's reliability are two sides of the same decision.

What a record quarter tells a buyer

When you adopt a tool, you are making a bet that the company behind it will still be there in three years, still investing, still answering support tickets. The cheapest way to read that bet is the vendor's own earnings. Microsoft reported a record quarter, with revenue of $92.3 billion, up 19% on the year before [1]. For a buyer, the useful signal is not the headline number but the direction: a vendor growing at that pace is unlikely to abandon the products you depend on.

Two further disclosures sharpen the picture. Microsoft said for the first time what Azure earns in dollars, which it had never previously broken out [2]. Visibility like that helps you judge how committed a vendor is to the part of its business you actually buy from. The company also recorded $10 billion in revenue from its commercial arrangements with OpenAI in the quarter ending September 30 [3]. That figure cuts both ways. It shows how much money AI partnerships now move, and it shows how tightly the two firms are bound together. If a core feature in your stack depends on that relationship, the relationship itself is now part of your risk assessment. We covered the same question of vendor stability in yesterday's briefing.

The debt that has not come due yet

Strong earnings at the top of the market can hide stress lower down. SaaStr highlighted a report that the large defaults on software debt are still ahead, not behind. It cited Solera, a maker of software for vehicle maintenance, as having about $8 billion in loans coming due in 2028 and 2029 [4]. The teaching point here is simple and often ignored. Many of the tools a small business runs on are owned by private-equity holders who financed the purchase with debt. When that debt matures, the owner must refinance or sell, and either event can change pricing, support, or the roadmap you were counting on.

You cannot audit every vendor's balance sheet. You can do two practical things. First, keep your most important data in formats you can export and move. Second, avoid building a critical process on a single tool that would be expensive to leave. Neither step is dramatic. Both are cheaper than migrating in a hurry because an owner changed the terms.

Checking what AI writes into your spreadsheet

The second half of the day was about trust in the output, not the vendor. Google added cell-level citations to the AI fill feature in Sheets. When the AI function generates a value, it now shows the exact web source it used, and the feature supports larger requests [5]. This is a small change with a large principle behind it.

The principle is that an answer you cannot trace is an answer you cannot use for anything that matters. If AI fills a column with company sizes, revenue figures, or contact details, the number is only worth as much as its source. A citation attached to each cell turns an opaque guess into something a person can check in seconds. That is the difference between data you can put in front of a client and data you quietly hope is right. The lesson applies well beyond one spreadsheet product: whenever a tool offers to generate facts for you, ask where it lets you see the source. We wrote about drawing this line between useful and unsafe automation in what AI should and should not do in your business.

Agents moving into the developer's workflow

If your business builds any of its own software, the shape of that work is changing. GitHub used its Universe conference to describe itself as an end-to-end software delivery system, laying out a vision for every layer of the stack [6]. Alongside it, the company's annual Octoverse report said commits and pushes reached new highs in 2026 as developers build with agents [7].

The concept worth understanding is that "agent" here means software that carries out multi-step work on its own, not a chat window you type into. For a business choosing tools, the question is not whether this is impressive. It is whether the extra output comes with the same review you would expect from a person. More commits is only good news if each one is checked, tested, and understood. The trade-off to watch is speed against oversight: a team that ships faster because agents do more of the typing still needs someone accountable for what ends up in production.

Measuring which AI is good at your actual work

The last item is about how you compare models at all. Zapier announced that its AutomationBench has joined a collection of practitioner-built benchmarks that compare AI models on real business workflow automation [8]. This matters because most published AI scores measure exam-style tasks, not the dull, specific work most companies need.

The idea to take away is that a benchmark is only useful if it tests the job you have. A model that writes elegant essays may be poor at reliably moving a record from one system to another. When a vendor or a model claims to be good at automation, ask what it was tested on. A score drawn from real workflow tasks is worth more than a general ranking, because it predicts behaviour on the work you would actually hand it.

What to take from the day

Read these items together and a usable rule emerges. The money news tells you which vendors are investing and which may be under pressure, so you can weigh how much to depend on each. The product news tells you to demand traceability: a citation on a generated value, a review on an agent's commit, a benchmark drawn from real work. Neither half is enough alone. A financially sound vendor can still ship an AI feature you cannot check, and a well-designed feature can still sit inside a company whose debt comes due. Choosing tools well means holding both questions at once, every time.