The day AI inside your tools became a line item

A single thread runs through today's product announcements: the AI built into everyday business software is now something you pay for, budget for, and set limits on, rather than a free extra that arrives with an update. The same day brought agents that touch real money and real customers, which sharpens the second question every buyer now has to answer — what these agents are allowed to do, and what it costs when they do it.

The price of AI becomes a number you can control

For most of the past two years, AI features inside business software were priced the way early mobile data once was: you turned them on and hoped the bill at the end of the month made sense. That model works while usage is small and experimental. It stops working the moment a team starts leaning on AI for daily work, because the cost moves with usage and usage is hard to forecast.

HubSpot's announcement today speaks directly to that problem. It introduced spend caps, the ability to swap seats for credits, and more flexible payment options, framed around predictability rather than new capability [1]. A spend cap is the simplest and most useful of the three: it sets a ceiling, so a team can try something without the risk that an experiment quietly runs up a bill. The seat-to-credit swap matters for a different reason. Software has traditionally been sold per person, but AI is consumed per task, and those two units do not line up. Letting a buyer move budget from unused seats into usage credits is an admission that the per-seat model alone no longer describes how the product is used.

This is worth understanding before you choose any tool with AI inside it. Ask how the AI is metered, whether you can cap it, and whether the vendor shows you what you were actually charged rather than an estimate. A plan is a promise about what you will pay for what you use, and the clearer that promise, the easier the tool is to live with — a point we have made before in what a subscription plan is really selling. 360REV meters AI by the work it delivers and shows the charge against the answer, so the number on the invoice is the number you can check.

AI in the inbox moves behind the paid tiers

Email is where most knowledge work still begins, and it is where AI has the clearest short-term job: deciding what you need to read first. Google expanded AI Inbox in Gmail to paid Workspace users today, describing a feature that highlights actionable items and summarises priority work so people can catch up faster [2]. Alongside it, Gmail Live is also rolling out to paid Workspace users [3].

The useful detail here is not the feature itself but where it sits. Both capabilities are landing on paid tiers. That tells you something about the shape of the market: the AI layer is becoming part of what distinguishes a paid plan from a free one, not a universal default. For a business choosing an email and documents suite, this changes the comparison. You are no longer only comparing storage and seats. You are comparing which AI functions come with which tier, and whether the summarising and triage you would come to depend on are included or sold separately.

An agent that answers customers, with numbers attached

Most claims about AI agents are promises. Today brought one with figures. Salesforce announced that Tottenham Hotspur launched an AI agent called Ask Spurs on its Agentforce platform, reporting that the club unified over 4.6 million fan accounts across more than ten systems and that the agent reached an 86% engagement rate [4].

Two parts of that are instructive for any business, regardless of whether football is involved. The first is the unification step. The agent is only useful because the data behind it was pulled together from many systems into one place first. An agent answering customers is only as good as the records it can reach, and most organisations underestimate how much work that consolidation takes. The second is the engagement figure. An 86% engagement rate is a usage number, not a resolution number — it tells you people interacted, not that their problems were solved. When you read vendor metrics, separate the two. We wrote about choosing which numbers actually change a decision in numbers that change a decision.

Agents that touch the money need the whole picture

Xero's post today makes an argument that applies well beyond accounting. It describes how every transaction a business makes sets off a chain of obligations — a bill to approve, a payment to make, a reconciliation to do, a cash position to update — and argues that an agent acting on your behalf in commerce needs the complete cash flow picture to do so safely [5].

The principle underneath is simple. An agent that can act, but cannot see the full state of your finances, is more dangerous than no agent at all, because it will act confidently on a partial view. This is the same caution that applies to any automation you let near a decision. The more an agent can do, the more complete and current its data has to be before you trust it. We set out where automation should stop in what AI should and should not do in your business.

Keeping your own copy of the data in sync

The last item is quieter but matters to anyone building on top of a vendor. Stripe introduced a way to sync its data to a Postgres database in real time, noting that applications rely on current payment data to fulfil orders, update accounts, and manage access to paid features [6]. Until now that syncing was custom work each team had to build and maintain.

The lesson generalises. If your business depends on data held inside someone else's service, you want a supported, current copy of it in a place you control — for reporting, for backup, and for the day you might need to move. That is not a lack of trust in the vendor. It is ordinary prudence, and the same instinct behind being able to export your records on any ordinary day, which we covered in the data you should be able to export on any Tuesday.

What to take from the day

Read together, these announcements describe a market that has stopped treating AI as a novelty and started treating it as infrastructure with a running cost. The features are moving behind paid tiers, the bill is becoming something you cap and swap and forecast, and the agents are being pointed at real money and real customers. The questions that follow are the practical ones: what is it allowed to do, what can it see, and what will it cost. If you compared tools yesterday on seats and storage, the comparison this quarter also includes how the AI is metered and whether you can keep your own copy of what you put in. For more on the week's thread, see yesterday's daily briefing.