Agents become buyable and the cost of running them lands

Two threads run through today's announcements, and they are really one thread. The AI agent is turning from a demo into a product a business can actually buy, and every vendor is now working out how context reaches that agent and who pays for the compute when it runs.

The agent moves from promise to product

An AI agent is not the same thing as a chatbot. A chatbot answers a question. An agent is given a goal, decides on the steps, uses tools to carry them out, and reports back. For the past two years that distinction lived mostly in keynotes and waitlists. A review of the field this week frames the shift plainly: "AI agents were the promise of 2024, the hype of 2025, and the expectation of 2026 and beyond" [1].

For a business choosing tools, the practical change is that you can now judge an agent the way you judge any other piece of software. Ask what tools it can reach. Ask what tasks it finishes without a human in the loop. Ask what guardrails sit around your data when it does. Those are procurement questions, not research questions, and that is new. If you kept agents at arm's length last year because none could be trusted with real work, that reasoning is worth revisiting now. We wrote about the general move from pilots to purchases in yesterday's briefing, and this is the same current running under a specific category.

The caution that matters is matching the agent to the task. An agent that books a meeting and an agent that issues a refund carry very different risk. The first is reversible in seconds. The second touches money and a customer relationship. A sensible buyer draws a line between what an agent may complete on its own and what it must hand back to a person for approval.

When the buyer is an agent, not a person

There is a sharper version of this argument doing the rounds. It starts from Jeff Lawson's 2021 book *Ask Your Developer*. The summary of that argument is that "developers had become the real buyers of a huge slice of B2B software" [2]. The engineer picked the tool, wired it in, and brought the budget along behind them. The new claim is that the agent is becoming the developer: the thing that decides which tool to call when a task needs doing.

If that holds even partly, it changes who your software has to please. A human buyer reads a landing page and watches a demo. An agent reads your documentation and your interface and makes a choice in milliseconds. So the question for a vendor becomes whether their product is legible to a machine as well as to a person. Clear and stable interfaces, predictable behaviour, and honest descriptions of what a tool actually does start to matter in a way marketing copy does not. For a buyer, the mirror image is useful too. If your own staff are starting to let an assistant pick tools for them, you want tools that behave the same way every time, because an agent cannot improvise around a surprise the way a person can.

Why a small group of users drives most of your AI cost

The cheerful version of agents skips the bill. The honest version does not. A long piece on pricing this week comes from someone the writeup describes directly: "Ulrik Lehrskov-Schmidt has worked on pricing full time for about ten years" [3]. The point of the piece is that AI usage is not spread evenly across a customer base. A thin slice of heavy users can account for the large majority of what it costs a vendor to run the feature.

This matters to you as a buyer, not only to the companies selling to you. If a vendor charges a flat price per seat but a few of your people run the AI constantly, one of two things happens over time. Either the vendor absorbs the cost and the price rises later, or they introduce usage caps and metering. Neither is wrong, but both change what you signed up for. The practical move is to read how each tool prices its AI today. Flat, metered, or capped are three different contracts, and the one you accept now shapes your bill as your own usage grows. Ask the question before you depend on the feature, not after.

Moving from a meeting to the task that follows it

Most of the value of a meeting is lost in the gap between the meeting ending and anyone acting on it. Decisions get made, and then they sit in someone's memory until the next meeting. Google Workspace is pushing at that gap. In its own words, "To make it easier for teams to move from planning to action, we're introducing meeting recaps in Google Chat" [4].

The idea is to put the summary where the follow-up work already happens, rather than in a separate document nobody reopens. The trade-off is the same one every summary feature carries. A recap is only as reliable as what it captured, and a tidy summary can give false confidence. Treat it as a prompt to check, not a record to trust blindly. For small teams in particular, the real win is reducing the number of status meetings, a theme we covered in keeping a small team aligned without meetings.

Giving an assistant access to your own notes

An assistant is only useful if it can see your material. A general model knows the world. It does not know your notebook. Google is connecting those two with its notebook tool, where "This new integration allows you to manage and query notebooks directly within your chats using natural language" [5].

The concept worth understanding here is the connected app. Instead of copying your notes into a chat, you grant the assistant access to them and ask questions in place. That removes friction, and it also moves a data boundary. Once an assistant can read a body of your work, the sensible checks are the ordinary ones: who authorised the connection, what it can reach, and how to switch it off. Convenience and access are the same decision viewed from two sides.

A note on the contact centre

One recognition is worth logging for anyone shopping in this area. Zoom reports that its contact centre product appears in this year's Gartner analysis, and invites readers to "See how we believe AI-first design and unified communications define our approach to CX" [6]. An analyst placement is one input among several, not a verdict. It tells you a product is being tracked seriously. It does not tell you whether it fits the way your team actually handles customers.

The common lesson across all six items is that the agent era is now a buying question, not a watching question. Decide what AI may do on its own, read how each tool prices it, and keep a person on the decisions that are hard to undo.