AI on top, but the value is in the data underneath

Almost every announcement worth reading today mentioned AI in some form, from new models to interactive workspaces built on top of a chatbot. The thread running through them is quieter than the headlines: the value a business gets from any of these tools still comes from the structured data underneath — the records, the evidence, the resources you can find and produce — not from the model layered on top.

When AI agents still need real software

There is a tempting argument going around: once you have capable AI agents doing the work, you no longer need purpose-built business software. Give the agents a database and let them run. It is worth understanding why that argument is seductive and where it breaks down, because it changes how you spend money this year.

The seductive part is real. A general-purpose agent plus a raw store of data can answer a lot of ad-hoc questions, and for a small enough operation that can feel like enough. The break comes when the work has to be repeatable, auditable, and shared across people who were not in the room when a decision was made. A raw database has no opinion about what a customer record should contain, what states a deal can be in, or who is allowed to change what. That opinion — the structure — is most of what you are buying when you buy business software. SaaStr, writing about its own operation of more than twenty AI agents and three humans, reached the same conclusion: the clean story that agents remove the need for a CRM "sounds clean" but, in their words, "It's also wrong, I think."[1]

The practical reading is not "ignore agents". It is that agents work best on top of structured systems, not instead of them. If you are weighing a tool, ask what structure it imposes and whether that structure matches how your business actually operates. We cover the underlying idea in what a CRM is actually for and, on the automation side, in what AI should and should not do in your business.

Winning a dispute is an evidence problem

If you take payments online, you will eventually face a "product not received" dispute — a customer, or their bank, claiming an order never arrived. What determines whether you keep the money is rarely persuasion. It is whether you can produce the right evidence quickly and in the right form.

Stripe published an analysis of what actually moves win rates, based on real data rather than folklore. They looked at "evidence packets from one million disputes over a 16-week period" to see which pieces of evidence correlate with businesses keeping their revenue.[2] The specific findings matter less, for a buyer, than the shape of the lesson: disputes are won by records you captured at the time of the transaction, not by arguments you construct after the fact. Delivery confirmation, timestamps, customer communications, and the trail of what was promised and when — these either exist in your systems or they do not.

This is why the boring question about any tool — what can I export, and how completely — is not boring at all. A system that quietly discards the shipping timestamp or the message where the customer confirmed the address has cost you a dispute you did not know you were going to have. It is the same discipline we describe in the data you should be able to export on any Tuesday: if the evidence only lives in someone's memory or a screenshot, it is not evidence you can rely on.

Choosing which AI, not just whether

Google released three new Gemini models today. Notably, none of them was the next flagship: TechCrunch reported that "Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber, but the continued absence of Gemini 3.5 Pro raises fresh questions about its AI strategy."[3]

For a business, the useful takeaway is not the individual model names. It is that "which AI" is now a real decision with real trade-offs, not a single switch you flip on. Vendors increasingly ship families of models rather than one model — lighter, faster variants for high-volume tasks and heavier ones for harder reasoning. That is good for cost control, because you can match the model to the job instead of paying flagship rates for a task that a smaller model handles. It also means the tool you buy should let you see and, ideally, choose what is running under the hood, so you are not silently upgraded or downgraded in ways that change your bill or your output quality.

The general rule holds regardless of provider: a model is a component, and components change. Build on tools that treat the model as a configurable part rather than a permanent fixture, and you avoid being trapped when the next family ships.

AI that shows its work

GitHub described a feature it calls canvases, aimed at making AI output something you can work with rather than just read. In its words, "Canvases turn AI into interactive workspaces where you can visualize information, explore workflows, and take action across complex tasks."[4]

The idea worth extracting is the direction of travel. A plain chat reply is a dead end — you read it, and any next step is manual. An interactive surface, where the AI's output becomes something you can inspect, adjust, and act on, keeps the human in control of the result. For a buyer, the question to ask of any AI feature is whether it produces something you can verify and change, or only something you have to trust. The former fits into a real workflow. The latter tends to sit unused after the novelty fades, because people cannot check it.

Meeting resources in one place

Google also revamped the Google Meet homepage on the web. The stated purpose is organisation: "We're introducing a revamped Google Meet homepage on the web to help you stay organized and prepared throughout your entire meeting workflow."[5]

The concept behind this is worth naming, because it recurs across tools. Most wasted time in meetings is not the meeting itself; it is the scramble beforehand to find the right notes, the last thread, the file someone shared. Pulling those resources into one predictable place is a small structural fix with an outsized payoff, and it is the same principle as keeping a customer's quotes, invoices and conversation together rather than scattered across apps. When you evaluate a collaboration tool, look at how much hunting it removes, not how many features it adds.

The common lesson

Four of today's five items are, on the surface, about AI. Read together, they point somewhere unglamorous. The dispute analysis wins on records captured at the time. The agents story ends at the need for real, structured software. The interactive workspace matters only because it lets a person check the work. Even the meeting redesign is really about putting the right material where you can find it. The model on top gets the attention; the data and structure underneath decide whether any of it is worth paying for. When you choose a tool this quarter, spend most of your judgement below the AI layer, where the durable value lives.