The AI bill at the top, finished work at the bottom

The announcements worth your attention today split neatly into two layers. At the top, the companies that rent out computing are describing spending on a scale that will eventually show up in what you pay; at the bottom, the tools you open every morning are quietly learning to produce finished work rather than just suggestions.

That gap is the story. It is easy to read the trillion-dollar headlines and conclude that nothing has changed for a small team choosing software. The opposite is true. The infrastructure numbers tell you where pricing and lock-in are heading, and the feature releases tell you what your staff will actually be doing next quarter. A sensible buyer reads both.

Files you can edit inside the assistant

Google Workspace said it is adding ways to create, preview, and edit files and code directly inside the Gemini app, including the ability to expand to full screen and highlight text to ask questions [1]. This is a smaller change than it looks, and a more important one.

Up to now, the usual pattern with a chat assistant has been a two-step one. You ask for a draft, the assistant returns text, and you copy that text into a document, a spreadsheet, or your codebase, where the real work lives. The artefact and the conversation stay separate. What this release describes is the artefact moving into the conversation: you generate a file and keep editing it in the same place you asked for it.

The convenience is obvious. The trade-off is where your source of truth ends up. If a quotation, a policy, or a script now lives as an object inside a chat thread, you need to know whether it leaves a durable, governed file that your team can find in six months, or whether it stays bound to one person's chat history. For a business choosing tools, that is the question to ask a vendor, not whether the editing feels quick. At 360REV we treat the saved record as the thing that matters and the conversation as the thing that produced it, so that the output outlives the session that made it.

If you want the longer version of this argument, we set it out in what AI should and should not do in your business.

There is no single AI strategy, even inside one company

Salesforce published a piece making a point that is easy to miss in a year of confident pronouncements: that AI is still jagged, and that your enterprise strategy should be jagged too [2].

"Jagged" is the useful word here. It means capability that is uneven rather than uniformly good or bad. A model can be strong at summarising a long document and weak at a closely related task, such as reconciling two documents that disagree. The danger is to treat AI as a single switch you flip on across the whole business, because the quality you see in a demo of one task does not carry over to the next one.

For a buyer, the practical consequence is to stop looking for the one tool that does everything and start matching tools to tasks where you can actually check the output. That means picking jobs with a clear right answer, keeping a person on the tasks where being wrong is expensive, and expecting your mix of tools to look untidy for a while. An untidy map is not a failure of planning. It is an honest reflection of where the technology is reliable today.

The bill behind the AI boom

At the infrastructure layer, Amazon's chief executive said on Thursday that Amazon wants to be the king of AI cloud computing and become the first $1 trillion revenue business, and that getting there means continuing to spend heavily on capital expenditure [3].

This is not gossip about other people's balance sheets. The money that cloud providers commit to data centres and chips is the raw cost of the services almost every piece of software you use now sits on top of. Over time, that cost flows downstream. Sometimes it arrives as a new premium tier for the AI features; sometimes it arrives as a quiet change to usage limits. When a provider signals years of heavy spending ahead, the reasonable assumption is that it intends to recover that spending from customers, and that the customers include the vendors you buy from.

So the useful way to read a capex announcement is as a forecast of your own future unit costs and of how much a vendor will want you to standardise on its stack. Neither is a reason to avoid these platforms. It is a reason to keep your own data portable, so that a future price change is a decision you can act on rather than one you have to accept.

What booming revenue cost in cash

The same set of results carried a second number worth holding onto. Amazon reported strong cloud revenue but otherwise missed expectations, and the report included negative free cash flow of $2.4 billion [4].

Free cash flow is the cash a business has left after it pays to run and expand itself. It can go negative even while revenue grows, precisely because the company is spending ahead of the income to build capacity. That is a defensible choice for a business this size. The point for a buyer is the general one: the financial decisions a vendor makes shape its roadmap and its pricing far more than any feature page does. A supplier spending aggressively to win a market will behave differently from one protecting margin, and both patterns eventually reach your renewal. Reading a vendor's own numbers is part of choosing it.

Automation as the layer that connects the rest

Finally, Zapier published a guide to using Gemini through its integration, opening by noting that anyone who lives in the Google ecosystem has probably used Gemini for a quick draft or a fast answer [5]. The piece itself is practical, but the category it sits in is the one that ties today's two layers together.

Automation is the connective tissue between the assistant that drafts something and the systems where your work actually lives. On its own, a model that can write a reply or summarise a thread is a convenience for one person. Wired into the tools your team already uses, the same capability becomes a step in a process that runs without anyone copying and pasting. That is where the value stops being personal and starts being operational.

The caution is the mirror of the Salesforce point above. Because capability is jagged, the steps you hand to automation should be the ones with a predictable outcome, and the steps that need judgement should still pause for a person. An automation that confidently does the wrong thing at scale is worse than no automation at all.

What to take from the day

The thread running through all five items is the distance between what is being spent and what is being shipped. The spending is enormous and will, in time, reach your invoice. The shipping is modest and immediately useful: a file you can edit in place, an integration you can wire up this afternoon. A business choosing tools this week should let the small changes guide what it adopts and let the large numbers guide how carefully it keeps its data its own. We covered a similar split in yesterday's briefing, and the pattern is holding.