Watching the AI inside your tools

Two things are happening to business software at once, and today they turned up side by side. The AI now built into everyday tools is doing more on its own, and the way companies decide what to buy has grown harder to read — which puts a plainer question in front of every buyer: what is your software actually doing, and who can see it.

Most of the items worth your attention today sit on one side or the other of that question. Some are about the AI inside the tool. Some are about the market the tool is sold into. Both end up at the same place: you need to be able to inspect what you are running and trust that the company behind it will still be there next year.

Seeing what the AI does inside your tools

When a feature stops being something a person clicks and becomes something a model does on their behalf, a gap opens in your records. The old assumption — that every action in a system was taken by a named human — no longer holds, because now a person and a model act together. That makes an audit trail harder to read and more important to have.

Google Workspace said administrators can now reach audit logs for its Gemini Notebook feature in the Admin console [1]. An audit log is simply a record of who did what and when. For an AI feature, that record answers questions a business will eventually need to answer: which people used the tool, on what, and how often. Without it, an AI feature is a black box — useful, perhaps, but impossible to govern, review, or explain to anyone who asks.

If you are evaluating any tool with AI baked in, treat visibility as a requirement, not a nicety. You want to know what the model touched, on whose behalf, and whether you can export that record later. We have written before about the audit trail nobody thinks about, and the arrival of AI features only sharpens the point. The same discipline you would apply to a human action — log it, attribute it, keep it — now has to cover the model too. This is also why it pays to decide in advance what AI should and should not do in your business: a boundary is only real if you can check whether it held.

Buying software in a market that keeps moving

TechCrunch reported research arguing that a startup's recurring revenue is less secure than it has been, and pointed at the cause: new buying behaviour in the AI era [2]. For a buyer, this is not gossip about vendors. It is a warning about the ground under the tools you depend on.

Recurring revenue — the subscriptions a software company counts on month after month — is what keeps that company funded and its product maintained. When buying patterns shift quickly, as the research describes, a vendor's income becomes less predictable, and a less predictable vendor is a riskier dependency. That does not mean avoiding newer companies. It means doing the homework: can you get your data out, is the pricing stable, and would your business survive if this particular tool changed hands or changed course.

The practical response is the one we have argued for before — choose software worth using on evidence, not atmosphere. Ask what happens to your records if you leave. Ask whether the tool locks you in or lets you go. A healthy market is churning right now, and the only protection a buyer has is to assume any single tool might not last and to keep the exit open.

Running more than one AI agent at once

GitHub published a guide for beginners on running several agents at the same time in its Copilot app, framing it as the moment the idea stops feeling frightening and starts feeling capable [3]. An agent, in this sense, is software that carries out a multi-step task on its own rather than waiting for a click at each step. Running several in parallel means more work happening at once, with less of your attention on each thread.

That is genuine leverage, and it carries a genuine cost: review. One agent produces one piece of work you can check. Several agents produce several, and the bottleneck moves from doing the work to verifying it. A business that adopts parallel agents has to decide, before it turns them on, how the output gets reviewed and who is accountable for it. The capability is real. So is the new habit it demands. The teams that gain from this will be the ones that treat the extra output as something to be checked, not something to be trusted on sight.

A stack of tools that works together

Xero announced the winners of its 2026 App Awards, and opened with a line that is worth sitting with: behind every thriving small business is a stack of tools cutting through the noise [4]. The word that matters there is *stack*. Almost no business runs on one application. It runs on several, and the value is in how well they pass information between each other.

An ecosystem award is, underneath the celebration, a map of what connects to what. For a buyer, that map is useful information. A tool that plugs into the systems you already use saves you the quiet, recurring tax of copying data by hand and reconciling two versions of the truth. A tool that stands alone makes you the integration — and you are the most expensive and least reliable connector in the building. Before you add anything to your stack, it is worth asking the question we have asked before about why your tools do not talk to each other, because an isolated tool rarely stays cheap.

How a large vendor's numbers read

SaaStr looked at Salesforce at roughly $45 billion in recurring revenue, noting that it reported its quarter on 26 August and the stock rose 23 per cent [5]. A small business is not Salesforce, but reading a large vendor's results is still a skill worth having, because the tools you buy are often sold by public companies whose numbers are in the open.

What you are looking for is not the headline but the direction. Is revenue still growing, and growing from the core product rather than from one-off effects. A single quarter's share-price jump tells you how investors feel on one day. The underlying growth rate tells you whether the company funding your tool is on firm ground. For a buyer, the lesson travels downward in scale: judge any vendor, large or small, on whether its business is sound enough to keep building the thing you are paying for.

The common thread across all five is control. You want to see what the AI does, know the vendor is solid, keep your tools talking to each other, and stay able to leave. None of that is new. What is new is how fast the AI era is testing each one at the same time.