The cost and control of AI-assisted work

Today's announcements circle one question that every buyer of business software now has to answer: what does it cost to run AI inside a tool, and who watches what that AI does. The vendors are pulling in two directions at once — making each unit of AI work cheaper to run, while quietly raising the amount of oversight the software now needs from you.

That tension is worth understanding before you sign anything. A tool that bundles AI is not a fixed purchase in the way a spreadsheet licence was. It has a running cost that moves with how you use it, and it takes actions on your behalf that you may not see. Both of those change how you should evaluate the tool, and today gives us concrete examples of each.

Making each token count is now a product feature

GitHub published a piece on how Copilot handles context and routes requests between models, framed explicitly around getting more useful work out of each token you spend [1]. The interesting part is not the engineering. It is that the efficiency of the machine has become something a vendor markets to you directly, because you are the one paying for the inefficiency.

When a tool charges by usage — credits, tokens, actions, whatever the unit — the vendor's decisions about which model to use and how much context to load are decisions about your bill. Load too much context and you pay for words the model did not need. Route a simple request to an expensive model and you pay for capability you did not use. GitHub's framing, that the work is about making "more of each session go toward useful work, so your credits go further" [1], is a signal that these choices are now visible enough to compete on.

For a business choosing tools, the lesson is to ask how a usage-based product decides what to charge you for, and whether that decision is made in your favour or the vendor's. A metered feature is only fair if the meter counts your work and not the machine's overhead. We take that view seriously in how we charge for AI actions, and it is worth demanding from any vendor. If you are weighing what a plan actually sells you, the difference between paying for your work and paying for someone else's plumbing is the whole thing — a point we made in what a subscription plan is really selling.

Software that acts for you needs to be watched

SaaStr published a piece today with a blunt title: monitor your agents, both the AI ones and the human ones [2]. It opens with a story about a PR agency telling them never to contact their client again, after an executive had a bad experience — the point being that an agent acting in your name can damage a relationship before you know anything happened.

This is the other half of the day's thread. As tools take actions on your behalf — sending messages, updating records, reaching out to people — the question stops being "can it do the work" and becomes "can I see what it did." An AI that drafts a reply is a tool. An AI that sends the reply is an actor, and an actor needs a record.

The practical requirement is an audit trail: a plain log of what was done, by which agent, to whom, and when. Without it, a mistake is invisible until a customer complains, which is exactly the situation the SaaStr piece describes. With it, you can find the action, understand it, and correct the behaviour rather than apologising blind. This is not a new idea, but AI makes it urgent, because software now takes actions that used to require a person to type them. We wrote about where that line should sit in decisions automation should never make, and the general case for a record you can read afterwards in the audit trail nobody thinks about.

How the best software companies actually sell now

SaaStr also covered a set of sessions from its 2026 event on how companies including Stripe, Google, Canva and Cloudflare are selling this year, with the promise that these sessions "actually showed the data, the org charts, the pricing models, and the failures behind" the usual abstract talk about AI and go-to-market [3].

Why does a buyer care how vendors sell? Because pricing models are converging on usage, and the way a company sells tells you how it will bill you. When the pricing model is the story — as SaaStr says it was on that stage — the shape of your invoice is being decided in those rooms. A per-seat plan and a per-action plan feel similar on a pricing page and behave completely differently once your team grows or your usage spikes. Reading how a category prices itself is part of choosing well, and the failure modes are consistent enough that we catalogued them in how pricing pages fail.

Generated video moves further into everyday tools

Google announced new video capabilities in Google Vids, letting users create longer videos and generate several at once using its Veo model [4]. The announcement describes "new ways to create and iterate on video content in Google Vids using Veo" [4], with longer runtimes and consistent characters across a clip.

The reason this belongs in a buying briefing is not the feature itself but what it signals: content generation is being folded into the general productivity suite, not sold as a separate creative product. When video generation lives inside the same tool as your documents and slides, the cost of trying it drops to nearly nothing, and the volume of generated material your team produces goes up. That is worth planning for. More generated content means more to review, more to store, and more to keep consistent with what your brand actually says. Capability that arrives quietly inside a tool you already pay for still has a cost of use.

The quiet usability change worth noticing

Google also expanded event colours in Google Calendar, moving past its previous limit of eleven predefined colours [5]. On its own this is small. In the context of the rest of the day, it is a reminder that not every useful change is an AI change.

A broader palette lets a team encode more meaning into a calendar at a glance — client work in one colour, internal in another, travel in a third — without opening anything. The value of software is often in exactly this kind of small affordance, the thing that removes a moment of friction you have stopped noticing. When you evaluate a tool, the headline capabilities get the attention, but the daily-use details are what决定 whether people actually adopt it. That is the unglamorous half of choosing software that is worth using, which we set out in choose software worth using.

What to take from the day

The through-line is that AI has changed what you are buying. A tool with AI in it has a running cost that moves with use, so ask how the meter is set. It takes actions in your name, so ask what record it keeps. And its pricing model is being shaped by how the vendor has decided to sell, so read that as carefully as the feature list. The video and calendar updates round out the picture: capability keeps arriving, some of it loud and some of it quiet, and the work of choosing is deciding which of it actually earns a place in how your business runs.