Choosing tools by the details that matter

The productivity-software news on 14 July shares one theme: the value of a tool is decided less by its headline feature than by the small choices around it. Who sets the exact words a model writes, which model runs at all, and what a listed price actually means are the details that separate a tool you can live with from one you cannot.

The AI race moves from the frontier to the model you can own

For two years the story of business AI was a race to the single most capable model. That framing is now being questioned. According to a report on 14 July, "Hugging Face CEO Clem Delangue says enterprises increasingly want open models, due to cost, accessibility, and ownership." [1]

Three words there carry the weight. Cost, because a model you host does not bill you per request. Accessibility, because an open model can run where your data already sits, rather than requiring every record to travel to someone else's service. Ownership, because a model you control does not change underneath you when a vendor revises its terms or retires a version you had built around.

The same day showed how much capital stands behind the frontier. Reflection AI "has signed a $1 billion deal to access Nebius' compute." [2] A single billion-dollar contract for computing power is a useful reminder that the leading edge is expensive to reach and expensive to hold. Part of the appeal of open models, for a buyer adding up total cost over several years rather than judging a demo, is precisely that the bill does not scale with every question asked.

The lesson for a business choosing tools is not to pick a side in the frontier-versus-open debate. It is to ask any AI vendor a plain question: if this model changed, or its price rose, what would happen to the work I have built on it. That is the same discipline we set out in what AI should and should not do in your business. A model is a supplier like any other, and a good supplier relationship survives the supplier changing its mind.

The exact words, not a preset

Most AI writing help offers a fixed menu: make it shorter, make it more formal, make it more polished. That is easy to use and easy to outgrow, because a menu can only ever offer what its author imagined. Google Workspace is loosening that constraint in Gmail. "Users can now edit and revise their email drafts in Gmail via the prompt bar, using custom refine instructions in Help me write." [3] The change matters only against what came before it: "Previously the refines were limited to preset options like Polish, Formalize, and Shorten." [4]

The move from a preset to a free-form instruction is a small change with a large consequence. A preset applies the same transformation to every draft. A custom instruction — "make this sound like the way I wrote to this client last month", or "keep it factual and drop the apology" — carries your judgement into the edit. That is the useful part. The cost is that judgement now sits with you. A preset is predictable; a free-form instruction does what you asked, which is only as good as the asking. Anyone rolling this out to a team should treat the prompt as a skill to be learned, not a button to be pressed.

The pricing gap that ends a deal before it starts

A software purchase can fail long before anyone talks to a salesperson. SaaStr described one this week, recounting a real attempt to switch tools: "We went to buy Riverside, a podcasting tool, after years of being on Zoom." [5] The piece frames the sticking point as a pricing gap — the space between what a buyer is ready to pay for and what the vendor has decided to charge for, or how it has drawn the line between plans.

A pricing gap is not the same as a high price. It is a mismatch in shape. A buyer who wants one capability finds it bundled into a tier full of things they will never use, or split across an add-on they did not expect. The deal does not end in a negotiation; it ends in a quiet decision not to proceed, and the vendor often never learns why. This is why a pricing page is a sales tool in its own right, and why an unclear one costs revenue that never shows up as a lost deal. We went into the mechanics of this in how pricing pages fail. If you sell software, the exercise is to read your own pricing as a stranger would and count how many questions it leaves unanswered. If you buy software, the exercise is to name the exact capability you need and check whether the plan you would pay for actually contains it.

What to expect when a vendor's system breaks

Every tool you depend on will fail at some point. What separates vendors is not whether that happens but what they do in the hour it does. Intercom set out how its team handles this for its AI agent, opening with the plain truth of it: "When customers rely on you, minutes matter in an incident." [6] The company describes a defined sequence — detect, mitigate, and learn — that its engineers run every time.

The word worth underlining is learn. Detection and mitigation get a service back on its feet; only the learning stops the same failure recurring. A buyer evaluating a critical tool should ask the vendor to describe their incident process in these terms, and be wary of a vendor who cannot. A defined process is a sign that outages are treated as events to be studied rather than embarrassments to be buried. The record such a process leaves behind is exactly the kind we described in the audit trail nobody thinks about: boring until the day it is the only thing that tells you what happened.

Compliance tools shaped by the people who use them

Some of the most valuable product work is invisible from the outside, because it removes a cost the user had stopped noticing. Xero published an update to its Workpapers feature and named that cost directly: "If you’re a Xero partner in Australia, you know compliance season has a cost that doesn’t always show up in timesheets or recoverable time billed to the client." [7]

That is the interesting claim, quite apart from the feature. A cost that does not appear on an invoice — re-keying figures that already exist, checking numbers by hand, the friction of a busy season — is real money spent, and it is easy to leave unaddressed precisely because nobody bills for it. A tool that removes such work returns time that was never being counted. When you assess any tool aimed at a repeated, seasonal task, the honest measure is not the licence fee but the unbilled hours it gives back, set against that fee.

The thread, pulled tight

The five items differ in subject but agree on a stance. A business is better served by a tool whose model it can reason about, whose edits it can direct, whose price it can read, whose failures it handles openly, and whose value it can measure in hours rather than adjectives. None of that is about the newest capability. It is about control, clarity, and knowing what you are actually buying — the same standard we hold every tool to in choose software worth using.