When AI tools get practical about cost and control

The productivity-software news on 18 June 2026 shared one thread: the industry has largely stopped talking about what artificial intelligence might do one day and started dealing with what it costs, how it behaves, and who controls it. For a business choosing tools, that shift is welcome, because cost, behaviour and control are things you settle with a budget and a written policy rather than with a demo.

What follows is the handful of announcements that matter if you are deciding what to pay for, in order of how much weight they should carry.

What spending data says about AI budgets

When you are deciding which tools to fund, one of the quieter signals is where other buyers are already putting their money. Individual case studies are easy to cherry-pick. Aggregate spending is harder to argue with, because it reflects thousands of separate decisions made with real budgets rather than a survey answer. Payment processors sit on exactly that kind of data, which is why their observations are worth reading even when you are not their customer.

Stripe published an analysis of spending patterns across the customers paying with its Link product and reported a clear direction of travel. The company said it found that those customers are spending more on AI than they were three months earlier, and putting money into platforms that let them build with AI. [1]

Treat a rising line as information, not as instruction. A category that is maturing tends to get cheaper per unit of work over time, which can be a reason to wait rather than to rush. What the trend does tell you is that AI has moved from experiment to line item for a lot of businesses, and that budgeting for it as a recurring cost — not a one-off pilot — is now the normal case. If you want a discipline for reading figures like this before they change a decision, we have written about that separately in numbers that change a decision.

Building spreadsheets by describing them

The spreadsheet is still the tool most businesses reach for first, and the reason is familiar: it asks nothing of you except that you already know how to use it. The barrier has always been the formulas and the structure. Natural-language features aim at that barrier by letting you describe the result you want in ordinary words and having the software assemble the sheet.

Google reported that the Gemini capabilities that let people build and edit entire spreadsheets using plain language are being extended to more of the world. The company noted that earlier in the year it introduced Gemini in Sheets features that allow you to build and edit entire spreadsheets using simple natural language, and that support is now widening. [2]

The practical point is about who on your team can now do this work. When the interface is a formula, spreadsheet work concentrates in the few people who are fluent. When the interface is a sentence in your own language, that pool widens. The trade-off is the one every generated artefact carries: you still have to check the output, because a sheet that looks right and is wrong is more dangerous than one that is obviously broken. And there is a point at which describing a spreadsheet is a sign you have outgrown the spreadsheet itself — a subject we covered in when a spreadsheet stops being enough.

Deciding how your AI talks before it talks to customers

An AI agent that answers customers is, in effect, a member of staff working from a script. The difference is that nobody hired it, nobody trained it, and unless someone decides otherwise, nobody wrote the script. That gap is the subject of conversation design.

Intercom made the case plainly. The company wrote that if nobody on your team owns how your AI Agent communicates, it defaults to sounding like an LLM, and described conversation design as the discipline that fixes that. [3]

The idea generalises past any one vendor. Before an automated agent speaks for you, someone should own its tone, the moment it hands a conversation to a person, and the things it is not allowed to say or promise. Those are business decisions, not engineering ones, and they are cheaper to make on purpose than to discover after a customer has been given the wrong answer in a confident voice. This is really a specific case of a larger question about drawing lines around automation, which we set out in what AI should and should not do in your business.

Limits that reduce noise

Most software is built to accept more input, not less. So it is worth noticing when a widely used platform adds the ability to accept less on purpose. GitHub described new pull request limits and framed them around volume rather than capability. The company said the feature helps manage contribution volume in your repositories, and pointed to more on the roadmap. [4]

The lesson travels beyond code. Any channel that accepts inbound work — support tickets, form submissions, review requests — reaches a point where more throughput stops being a benefit and starts being noise that buries the items that matter. A cap is not a sign of weakness in a tool; it is a recognition that attention is the scarce resource. When you evaluate software, it is worth asking not only how much it can take in, but whether it gives you any way to slow the flow down when you need to.

The cost floor under the tools you buy

The price of an AI feature eventually traces back to the hardware it runs on, and that hardware market has been unusually concentrated. So competition at that layer is relevant even to a buyer who will never touch a chip. TechCrunch reported that Amazon is looking to compete more directly in that market, writing that AWS is in talks to sell its chips to other data centres, and noting that its chief executive has described this as a $50 billion opportunity for the company. [5]

Do not overreact to infrastructure news. A development at this layer takes years to reach your invoice, and the direction is not guaranteed. But over a long horizon, more suppliers of the underlying hardware tends to mean more pricing pressure on the AI features you pay for downstream. It is a reason to keep your commitments flexible rather than a reason to change a tool today.

When a vendor says autonomous

Finally, a note on the language of AI sales. The word doing the most work in a lot of pitches is autonomous, and it deserves scrutiny. SaaStr published a discussion of Artisan's fully autonomous AI business development representative that opened by acknowledging how similar most such pitches sound. The piece observed that a founder gets on stage, shows a slick demo, and promises the death of the BDR — and that the interesting question is what happens when you press on the claim. [6]

That is the right instinct for any buyer. When a product is described as autonomous, the useful follow-up is narrow and concrete: autonomous within what boundaries, reporting to whom, and stopping when what happens. A tool that can act on its own is only as safe as the limits around it, which brings the day's thread back to where it started. Cost, behaviour, and control are the three things worth pinning down before you sign.

At 360REV we build on the assumption that these are the buyer's decisions, not the software's, which is why automation and AI in the platform run inside limits you set rather than defaults you inherit. The broader point stands whatever you use: the announcements that should move you are the ones about how a tool behaves under your control, not the ones about what it can do in a demo.