The day software started talking to software

The announcements this morning share one subject: tools built for people are now being used, in part, by software acting on a person's behalf. The practical questions that follow are not about what these systems can do but about governance, cost and who is accountable when an automated actor does the work.

If you buy software this year, this is the lens to carry into every demo. A product that assumed a human at the keyboard now has to account for a non-human one, and that changes what you should ask before you sign. We keep a plain record of what each tool does to a customer so that when an automated step runs, you can still see who or what touched the account and when.

When half the traffic is not a person

Most communication tools were designed around a simple assumption: a message has a human sender. Permissions, notifications and audit trails all lean on that idea. When a share of the messages in a channel start coming from automated agents, the assumption quietly breaks. You are no longer only managing conversations between colleagues. You are managing a space where code posts, reads and reacts alongside people.

That shift is now measurable. Slack reported that half of the messages on the platform are no longer from people [1]. For a business, the useful response is not alarm but a few concrete checks. Can you tell a human message from an automated one at a glance. Is an agent's activity logged in a way you can review later. Who approved the agent, and who can switch it off. These are the same questions we raised in decisions automation should never make: the point of automation is to remove effort, not to remove the record of what happened.

The trade-off is real. More automated traffic can mean faster coordination and fewer manual updates. It also means more noise, and a harder time reconstructing a decision after the fact. Before you expand an agent into a live channel, decide how you will read it back.

Governance of what a meeting leaves behind

A video call used to be ephemeral. Now a single meeting can produce a recording, a transcript and a set of AI-generated notes, and each of those is a data object that lives somewhere, under some policy, visible to some people. Many organisations have never written that policy down, which means the default behaviour of the tool becomes the policy by accident.

Google Workspace is adding a way to display custom policy text in the green room before a Google Meet call [2]. The green room is the screen you see before you join, and showing policy there is a small design decision with a large governance effect. It puts the rules in front of a participant at the one moment they are about to generate the artefacts those rules govern.

For a buyer, the lesson travels beyond this one product. Ask of any meeting tool: what does it keep by default, for how long, and who can see it. If the answer is a buried setting, you have a governance gap. If the tool surfaces the policy to the people it applies to, you have a smaller one. The feature matters less than the habit it encourages, which is telling people what is being recorded before it is recorded.

The cost sitting behind the agents

Automated agents and AI notes do not run on nothing. They run on model providers, and the economics of those providers are now part of your supply chain whether you contract with them directly or not. When a vendor you depend on raises prices or changes terms, the cost flows down to you.

That is why the scale of the model business is worth watching even if you never buy from it directly. Anthropic's quarterly revenue rose to $16.96 billion from $11.56 billion in the previous quarter [3]. Numbers of that size tell you two things at once. The demand for these models is deep, which suggests they are not a passing experiment. And the amounts being committed are large enough that pricing and availability will be shaped by forces well outside your own usage.

The practical takeaway is modest. When a feature in your stack is powered by an external model, treat that model as a dependency with its own cost curve. Know whether your vendor passes model price changes through to you, and whether you could switch providers if the terms moved. A capability you cannot afford next year is not a capability you own.

When commercial demand overtakes the original customer

A company's revenue mix tells you who it is really built to serve, and that mix shifting is a signal about the wider market. TechCrunch reported that Palantir's U.S. commercial revenue surpassed its U.S. government revenue for the first time [4]. A firm long associated with the public sector now earns more from ordinary businesses.

For a buyer, this is a demand reading rather than a recommendation. When vendors that grew up serving one kind of customer start earning more from private companies, it tells you that enterprise appetite for these tools is broad enough to reshape established businesses. That is useful context when you are judging whether a category is early and risky or settling into something durable. It does not tell you the tool is right for you. It tells you that you are not the only one being asked to decide.

One maintainer, one assistant

Not every item today is about scale. GitHub described a project that turns any place on Earth into a Minecraft world, built by one person, where GitHub Copilot now helps him maintain it alone [5]. On its own it is a small story. As a pattern it is worth noticing.

Many of the tools a small business relies on are maintained by very few people, sometimes one. An AI assistant can extend what a single maintainer keeps alive, which is good news for the long tail of useful software. It also concentrates the knowledge in one place. When you adopt a tool kept going by one maintainer and one assistant, weigh the convenience against the question of what happens if that person steps away. The assistance that keeps a project alive today does not, by itself, answer who keeps it alive tomorrow.

What to take from the day

The common thread is that the actor on the other side of your software is increasingly not a person. That raises governance questions in your chat tool, in your meetings, in your cost base and in the small projects you quietly depend on. None of it argues against adopting these tools. It argues for adopting them with the record, the policy and the exit intact. For the previous day's reading, see yesterday's briefing.