AI lands inside the tools, but the plumbing decides

The announcements that mattered today share one thread: artificial intelligence is arriving inside the tools businesses already run, at the exact point where the work happens. The lesson underneath them is older than the AI, though — value comes from the plumbing around a feature, not the feature itself: current data, reliable backups, and a clear view of where growth already sits.

Incremental backups change what a good export looks like

A backup is a copy of your data you can restore from when something goes wrong. For years the standard shape of a business backup was the full export: dump everything, every time, and store it. That works, but it is expensive in time and storage, and the cost grows with the organisation. The larger you get, the more it hurts to re-copy data that has not changed since yesterday.

An incremental export changes the arithmetic. Instead of re-exporting the whole estate on every run, you capture only what moved since the last snapshot. That makes frequent backups affordable, which matters more than it sounds — the value of a backup is not just that it exists, but how recent it is. A backup from last quarter is a very different safety net from one taken this morning.

Google Workspace announced today that administrators can now use incremental exports when backing up organisational data [1]. The practical effect is that a business can take snapshots often without paying the full-export tax each time. When you are choosing tools, this is the kind of capability worth reading for: not the headline feature, but whether you can get your own data out cleanly and regularly. That principle sits behind the data you should be able to export on any Tuesday and what happens to your data when you leave — a tool you cannot back up on your own terms is a tool that owns you, not the other way round.

AI arrives where the work already happens

Most early AI features asked you to leave what you were doing, open a separate box, describe your problem, and paste the answer back. The more useful pattern is the opposite: the help appears where the error is, in the tool you were already using, aimed at the specific thing in front of you.

Google's other announcement today is a small but clear example. Gemini in Google Sheets can now diagnose and fix formula errors in one click [2]. A formula error is a well-defined problem — the spreadsheet already knows something is broken and roughly where — so it is a sensible place to let a machine suggest the fix. The task is bounded, the context is right there, and a wrong suggestion is cheap to reject.

That boundedness is the point. AI earns its place when the task is specific, the input is clean, and a human can glance at the result and accept or discard it. It earns much less when it is asked to reason across messy, half-recorded context. We wrote about drawing that line in what AI should and should not do in your business, and a one-click formula fix is squarely on the useful side of it.

An AI agent is only as current as its knowledge

If you put an AI agent in front of customers, you have taken on a maintenance job most teams underestimate. The agent answers from a knowledge base, and a knowledge base goes stale the moment the product changes. Ship a new feature, change a price, retire an option — and unless the agent's knowledge moves with you, it will confidently give last week's answer to this week's customer.

Intercom wrote today about the process it uses to keep its agent, Fin, ready for every release. Its framing is blunt: every time you launch a product, your agent's knowledge base needs to keep pace [3]. The useful idea here is procedural, not technical — agent readiness becomes a step in every release, not a clean-up job someone remembers later. That is the same discipline we described in keeping customer data current: automated systems are only as good as the information feeding them, and the information decays unless someone owns keeping it fresh.

The wider takeaway for anyone weighing whether to automate a customer-facing task is that the automation is never free. It carries an ongoing cost of keeping its inputs accurate. If you cannot commit to that upkeep, the honest answer may be not to automate the task yet — a question worth working through with how to tell whether a task should be automated.

The cheapest growth is the customer you already have

Most businesses spend their growth budget at the top of the funnel: more leads, more advertising, better conversion on strangers. There is a cheaper source of growth that most teams barely touch — the customers already paying them. Expansion revenue, the extra value an existing customer buys over time, tends to cost far less to win than a new logo, because the trust and the relationship already exist.

SaaStr made this point today through HappyFox, whose CEO described closing a million dollars in expansion on a twenty-dollar AI agent spend. The framing is worth quoting directly: the cheapest growth a company has is already sitting inside its own customer base, and almost nobody is mining it [4]. What makes that mining possible is not the AI on its own — it is having a record of who your customers are, what they use, and where the next natural step for them sits.

This is why the quality of your customer records is a growth question, not just an admin one. You cannot expand a relationship you have not written down. We made the operational case in follow-up is the whole job: the money is usually in the second conversation, not the first, and it goes to whoever remembered to have it.

Buying tools in a down market

The backdrop to all of this is a hard year for public software. SaaStr's look at Navan opened by noting that a chart of public B2B software in 2026 is a sea of red [5]. When markets tighten, software budgets tighten with them, and buyers rightly become more careful about what they commit to.

That caution is healthy, and it changes how you should choose. In a generous market you can afford a tool that mostly works and swap it later. In a lean one, the cost of a wrong choice — migration, retraining, lost data — is heavier, so the questions that matter move from the demo to the fundamentals: can you get your data out, does it hold accurate records, and does it do the boring things reliably. Those are exactly the tests in choose software worth using, and a down market is precisely when they earn their keep.

Read across the five items, the pattern is consistent. The AI is the visible part; the durable value is in backups you can take often, knowledge you keep current, records complete enough to grow from, and the plain reliability that survives a bad year.