The day AI stopped being a line on the pricing page

Across today's announcements a single thread runs: artificial intelligence is no longer a feature a vendor bolts on and charges for, it is becoming part of the cost of running the software and part of the controls an administrator has to manage. Box put a number on what AI does to a margin, Google added a security lever for who may open a product and from where, and two separate pieces argued that the hard work is rebuilding the business underneath, not shipping the feature on top.

If you are choosing tools, that thread changes what you should look at. The question is no longer "does this product have AI" — almost all of them now do. The useful questions are what the AI costs the vendor to run, what it costs you to govern, and whether the company selling it has actually rebuilt its work around it or simply added a tab. The items below are ordered by how much they should weigh on a buying decision. We build 360REV to be honest about these same trade-offs, which is the only time our own product appears here.

What AI costs the company that sells it

Box reported its quarter today, and the detail worth your attention is not the revenue line, it is the margin line. The report notes "20 Basis Points of AI Margin Cost" alongside revenue of $321.1 million, up 9 percent [1]. Twenty basis points is two tenths of one percent — small, deliberately disclosed, and a signal that running AI features costs the vendor real money on every use.

Why should a buyer care about someone else's margin? Because that cost has to go somewhere. A vendor absorbing it today is making a bet that the feature earns loyalty or expansion; a vendor that cannot absorb it will eventually move it onto your bill, meter it, or quietly cap it. When you evaluate an AI feature, ask how it is priced and whether the price is stable. A flat, predictable cost means the vendor has solved its own economics. A usage meter with no ceiling means you are carrying the vendor's variable cost, and your monthly figure will move with how much your team uses the thing. This is the same reasoning we set out in choosing software worth using: the price you see on day one is only useful if you understand what moves it on day ninety.

The modest acceleration Box described — the report calls it "real if modest acceleration" [1] — matters too, because it shows AI features can be shipped without wrecking a mature software business. That is reassuring for anyone worried that every vendor is about to raise prices to pay for a model.

Who may open the product, and from where

Google Workspace announced that Google Classroom now supports Context-Aware Access controls. The update, in Google's words, "allows Google Workspace administrators to set granular security parameters for Classroom access directly from" the admin console [2]. Context-aware access is a concept every business buyer should understand, because it is becoming standard across serious tools.

The idea is simple. A login proves who you are. Context-aware access adds conditions about the circumstances: which device, which network, which country, whether the device is managed and up to date. A correct password from an unmanaged laptop on an unknown network is treated differently from the same password on a company device in the office. This is the difference between "are you allowed on the platform" and "are you allowed in, right now, from here."

For a small business this is not an enterprise luxury. It is the control that stops a leaked password from being enough on its own. When you compare tools, look for whether an administrator can set these conditions centrally rather than trusting each person's good habits. That central, provable control is part of what we mean in what enterprise-grade should mean to a small business: the grown-up features should be available to you at your size, not reserved for the largest accounts.

A compliance deadline that has already passed

Xero published a reminder that the first quarterly deadline for Making Tax Digital for Income Tax has, in its words, "been and gone" [3], affecting sole traders and landlords earning over £50,000. The reason this belongs in a tools briefing is that regulatory deadlines are one of the few forces that should override your normal pace of software change.

Most tool decisions can wait. A statutory filing requirement cannot. If your accounting or bookkeeping tool does not support a mandated digital filing format by the date the rule takes effect, no amount of other features makes up for it. When a compliance change is on the horizon, the first question for any tool you rely on is not "is it nice to use" but "will it keep me legal on the date." Xero's framing — the deadline is already in the past for this first quarter — is a useful prompt to check that your own tools have caught up rather than assuming the rule is still coming.

Automating the small, repetitive decisions

GitHub published a walkthrough of using its Copilot app to triage Dependabot pull requests — the routine stream of library-update proposals a codebase generates. GitHub frames it plainly: "Learn how the GitHub Copilot app can handle this type of repetitive task" [4].

The wider lesson is about where automation earns its place. The best candidates are high-volume, low-judgement tasks: sorting, labelling, routing, flagging the ones a human should actually look at. Dependency triage fits because most updates are routine and a few are not, and a tool that separates the two saves attention for the ones that matter. The trap is letting automation make the decision rather than prepare it. The line we draw in what AI should and should not do in your business applies here: let the machine gather, sort and recommend, and keep the merge — the irreversible step — with a person.

Training the agents, not just buying them

Arga Labs, covered by TechCrunch today, raised money to build "a better way to train enterprise AI agents" — specifically "$10 million in a seed funding round that was led by General Catalyst" [5]. The existence of a company funded to do only this is itself the signal.

An AI agent that acts inside your business — reading records, drafting replies, updating systems — is only as good as what it has been taught about your business. The emergence of tooling aimed squarely at training agents tells you that buyers are learning the agent is not a finished product you switch on. It is a capability you shape with your own data and rules, and that shaping is where the value and the risk both sit. When you evaluate an agent-based tool, ask how it learns your context and how you correct it when it is wrong.

The spending behind all of it

Underneath every item above is the raw cost of compute. TechCrunch reported that "Amazon is adding another 2 million Nvidia GPU chips to its data centers over the next two years" [6]. You do not buy chips, but you feel their price through every AI feature you pay for. The scale of this spending is why Box's twenty basis points matters: the hardware is expensive, someone pays for it, and the vendors who have worked out how to absorb that cost are the ones whose pricing will stay still. That stability is worth more to a business choosing tools than any single feature.

The practical takeaway for today is narrow. When a tool shows you an AI feature, look past the demo to three things — what it costs to run, who can govern it, and whether the company has rebuilt its work around it. Those are the differences that will still matter a year after you sign.