The AI bill and the strategy you still bring

The companies that sell AI and the companies that run the clouds underneath it spent Thursday describing the same problem from two ends. The bill for AI is very large, and there is no single, tidy way to spend it — which means a business choosing tools has to think about both what its vendors are paying for and what strategy it is expected to bring on its own.

The cost behind the tools you rent

When you pay for a SaaS product, part of what you are paying for is compute you never see. The vendor runs its software on a cloud, and that cloud is built from data centres, chips and power that someone has to fund years before you click buy. That is why the spending plans of the largest cloud providers matter to a small buyer who will never sign a cloud contract directly.

Amazon made the scale of that plain. Its chief executive said the company intends to keep spending heavily to lead in AI cloud computing, with the stated ambition that "Amazon wants to be the king of AI cloud computing and become the first $1 trillion revenue business" [1]. Ambition on that scale is paid for up front. The same results showed the strain, where the report noted that "the biggest ouch of the report was its negative free cash flow of $2.4 billion" [2].

For a buyer, two things follow. First, the cost of building AI has to be recovered somewhere, and part of it arrives as the price of the AI features now appearing in tools you already use. Second, a vendor spending ahead of revenue is making a bet, and you are indirectly along for the ride. None of this is a reason to avoid AI features. It is a reason to read a pricing page carefully and to ask what happens to your costs when the free tier ends. If you are weighing a tool for the long run, it helps to choose software that lasts a year rather than one priced to win a launch.

There is no single AI strategy

The second thread is about how a business should adopt AI at all. The tempting idea is that an organisation picks one model, one assistant, one policy, and rolls it out everywhere. Salesforce argued the opposite on Thursday. Its framing was blunt: "AI is still jagged. Your enterprise strategy should be, too." [3]

Jagged is a useful word here. It means AI is strong at some tasks and weak at others, and the line between the two is not where intuition puts it. A model may draft a serviceable email and then fail at arithmetic a child would get right. If that is true, a single company-wide AI strategy will be wrong in most places. The better approach is to match each use to its own judgement about risk and reward. Automate the task that is repetitive and reversible. Keep a person on the task that is rare or hard to undo.

This is the same discipline that applies to any automation. Deciding what a machine should touch and what it should not is a per-task question, not a per-company one. We cover the reasoning in what AI should and should not do in your business. The point of Thursday's argument is that even the largest vendor of this software is telling buyers not to expect one answer.

Editing files inside the chat

While the strategy debate ran, the tools kept changing shape. Google described a change to where work actually happens: "We're introducing new ways to create and refine files and code in the Gemini app." [4] The stated idea is to let people generate, preview and edit files and code inside the chat window rather than moving between the chat and a separate document.

The concept worth understanding is the cost of switching. Every time a tool sends you somewhere else to finish a job — copy the draft, open another app, paste it, come back — you lose a little time and a little attention. Features that collapse that round trip are not about novelty. They are about keeping the work in one place. The trade-off is that the more a single app does, the more your working day depends on it, and the more you need to know you can get your content out when you want to. That is worth thinking through before any one tool becomes the room you live in. A feature that saves you five minutes a day is a real gain, but only if the switching cost it removes was yours and not invented by the tool in the first place.

When a system of record keeps its customers

The last item is quieter, but it teaches something durable about choosing tools. SaaStr looked at Guidewire, which "sells core systems to property and casualty insurers: policy, billing, and claims" [5]. By its own headline the company reached $1.24 billion in ARR, made a $1.7 billion AI acquisition, and reported its lowest churn ever for a 25-year-old system of record.

A system of record is the software a business treats as the single true place for a kind of data — here, the insurer's policies, bills and claims. The interesting detail is churn. Low churn on a system that old, serving slow-moving buyers, tells you something that a feature list cannot. Once a business has built its daily work on a system of record, it tends to stay, because leaving means moving the data that everything else depends on. That stickiness is a benefit when the system serves you well and a risk when it does not. It is the reason a buyer should care, before signing, how hard it would be to leave. We write about that in what happens to your data when you leave.

The thread across all four is one question in different clothes: who pays for the AI, who decides where it is used, and what you are tied to once you choose. The cloud bill is the vendor's to carry, but it reaches your invoice eventually [1][2]. The strategy is yours to set, task by task [3]. The place the work happens is changing [4]. And the tools that last are the ones your data settles into [5]. A business choosing software this week does not need to pick a side in any of these. It needs to read each one for the trade-off it names, and decide with its own work in front of it.