The running cost of agents, and the quiet admin changes

The thread running through today is cost — not the sticker price of software, but the standing cost of the AI agents now built into it, and the question of who absorbs that cost over time. Running alongside it are two quieter changes to everyday tools, a reminder that the decisions which affect a business most are often the dull ones about compliance and moving data between systems, not the headline features.

The margin bill behind agent features

When a vendor adds AI agents to its product, something changes in the economics that a buyer rarely sees. Traditional software has very low marginal cost: once it is built, serving one more customer costs almost nothing. Agents break that pattern. Every time an agent runs, it consumes compute, and that compute is a real, recurring expense the vendor pays whether or not the customer notices. The industry has started calling this the margin bill for agents — the gap between what AI features cost to run and what customers currently pay for them.

Databricks gives a sense of the scale at which this is now playing out. The company reported that it "crossed a $7 billion revenue run-rate in Q2, growing more than 80% year over year, and closed a $5 billion strategic round at a $190 billion valuation led by Coatue" [1]. Growth and funding on that scale buy a company time to absorb the cost of agents while it works out the pricing. A smaller vendor cannot. For a business choosing tools, the practical lesson is to read AI features as a cost the vendor is carrying for now, and to ask what happens to your bill when that cost is eventually passed on. A feature that is free during a land-grab is not necessarily free forever. This is the same reasoning we have written about before in what a subscription plan is really selling: the price today tells you less than the direction of travel.

Agents inside the software delivery workflow

The second strand of the agent story is about where the work happens. GitHub published a walkthrough of using agents to carry a feature through the stages of building software — described as four agent apps that "can help you scope, secure, roll out, and ship a feature across the SDLC–all without leaving GitHub" [2]. The phrase that matters there is "without leaving." The appeal of putting agents inside the tool a team already uses is that nobody has to switch context, copy work between systems, or reconcile two sources of truth.

That convenience carries a trade-off worth naming plainly. The more of a workflow a single platform absorbs, the more a business depends on that one platform, and the harder it becomes to move later. This is not an argument against consolidation — switching between disconnected tools has its own well-known costs, which we covered in why your tools do not talk to each other. It is an argument for going in with open eyes. Before you let one platform own a whole process, confirm you can still get your work out of it. The audit trail, the history, and the artefacts should be yours to export on any ordinary day, not held hostage to the convenience that drew you in.

What "built with agents" looks like from the inside

It is useful to see how the vendors themselves build with this technology, because it tells you how mature their claims really are. Klaviyo's co-founder spoke at an industry event about exactly this. By his account, he "came to SaaStr AI to walk through how a 2,300-person public company builds AI products" [3] — and the framing was explicitly about the build system rather than the vision.

For a buyer, there is a signal in that distinction. A vendor that can describe the plumbing — how agents are actually wired into the way the company works — is further along than one that can only describe the ambition. When you evaluate an AI feature, ask the supplier to walk you through how it runs day to day: what it does on its own, what a person checks, and what happens when it is wrong. A clear answer is a good sign. A deck full of adjectives is not. We set out which of those judgements a machine should and should not be trusted with in what AI should and should not do in your business.

A compliance change that hits the bottom line directly

Not every important change is about AI. Xero flagged a regulatory shift for Australian businesses: "From 1 October 2026, Australian businesses will no longer be able to add a surcharge to eftpos, Visa, Amex or Mastercard payments" [4]. This is a small, concrete rule with a direct effect on margins for any business that currently passes card fees on to customers. Those fees do not disappear; they simply stop being something you can add at the till.

The wider point for choosing tools is that your accounting and payments software is where changes like this land first. A rule change dated months ahead gives you time to model its effect — to work out what absorbing the fee does to your prices before the date arrives, rather than discovering it in a month-end figure. The value of the tool here is not a feature. It is that it tells you about the change early enough to plan, and it holds the numbers you need to plan with.

Moving data between systems without losing it

The last item is the kind of improvement that gets little attention and saves a great deal of frustration. Google announced that it is "introducing two key improvements in Google Sheets that preserve formatting and linked data when converting files from Microsoft Excel" [5]. The reason this matters is that most businesses do not live inside one vendor's world. Files move between spreadsheet programs constantly, and every conversion is a chance to lose formatting, break a formula, or quietly corrupt a figure.

When you assess a tool, interoperability of this kind is worth more than it looks. A product that reads and writes other formats cleanly reduces the cost of changing your mind later, and it lowers the risk that a routine export quietly damages your records. The test is boring and effective: take a real file, move it in, move it back out, and check that nothing changed. The tools worth keeping are the ones that pass it.

The through-line for a buyer

Four of today's five items point the same way. Agents are no longer a demo; they are a standing cost inside the products you buy, a question of who pays and when. The other two — a surcharge rule and a cleaner file conversion — are reminders that the unglamorous parts of a tool, compliance timing and data portability, often decide whether it is worth keeping. Choosing software well means weighing both at once, which is the habit we keep returning to in choose software worth using.