What counts as proof when you pick a tool

The announcements worth your attention today share one question. When you choose a tool, what counts as proof that it works, and who gets to do the judging. A public benchmark for code-review software, an AI interviewer that has already sat through half a million conversations, and an accounting product handed a customer-satisfaction award all turn on the same decision a buyer keeps making: which signal do you trust, and how much weight should it carry.

A benchmark for AI code review changes the conversation

Most buying decisions about AI tools are made on anecdotes. Someone tried a product on a few examples, it looked impressive, and that story becomes the basis for a purchase. The trouble is that a handful of good demonstrations tells you almost nothing about how a tool behaves on the hundreds of ordinary cases you will actually feed it. A benchmark is the attempt to replace the anecdote with something comparable: a fixed set of tasks, an agreed definition of the right answer, and a scoring method that treats every tool the same way.

GitHub published one today for AI code review. It describes the project as "a benchmark for code review agents built on representative GitHub pull requests, multi-source ground truth, calibrated evaluation, and production-aligned metrics" [1]. Each of those phrases is doing work. Representative pull requests means the test cases resemble the work your team does, not toy problems. Multi-source ground truth means the correct answer was not decided by one person with one opinion. Calibrated evaluation means the scores are meant to line up with reality rather than inflate. For a business weighing whether an AI reviewer earns a place in its workflow, the arrival of a shared yardstick matters more than any single vendor's claim. You can start to ask how a tool scores against others, rather than how good its own marketing says it is.

An AI interviewer at scale, and the cost of getting it wrong

Scale changes the stakes of any decision. A process that makes a small error once is a nuisance. The same error repeated across a large population becomes a pattern, and in hiring a pattern can be unfair in ways that are slow to notice and hard to undo. That is the lens to use on the news that an AI interviewer has moved from pilot to volume.

According to a report on the tool, it "has already conducted more than 500,000 interviews, with Snowflake, Snorkel, and Capgemini among its early testers" [2]. Half a million interviews is not an experiment; it is an established channel through which candidates now pass. There are real attractions for an employer. An automated interviewer is consistent, available at any hour, and never tired on a Friday afternoon. Those are genuine gains over a rushed human screen. The trade-off is that consistency is only a virtue when the thing being applied consistently is correct, and a judgment applied half a million times is a judgment you want to have examined carefully. A business considering a tool like this is not only buying efficiency. It is deciding where human judgment stays in the loop and where it does not, which is a question every organisation should answer on purpose rather than by default. We have written before about which decisions automation should never make without a person signing off.

Markdown in Docs, and why formats decide whether tools talk

A file format is a quiet decision with loud consequences. When two tools store information in formats that understand each other, work flows between them without friction. When they do not, someone spends an afternoon copying, reformatting, and checking that nothing broke on the way across. The less glamorous half of any software stack is the plumbing that moves content from one place to another, and formats are that plumbing.

Google is now treating Markdown as a native citizen of its document tools. The company said it is "introducing the ability to view, edit, and collaborate on Markdown (.md or .markdown) files directly in Google Docs, as well as view rendered previews in Google Drive" [3]. Markdown is the plain-text format that developers, documentation writers, and a growing number of AI tools produce by default. Until now, bringing a Markdown file into a collaborative editor usually meant converting it and losing something in the process. Editing it in place removes a conversion step and the small risks that come with every conversion. The broader point for a buyer is that interoperability is a feature in its own right, and it is worth checking before you commit. A tool that reads and writes the formats your other tools already use will cost you less in friction over its lifetime. This is the same concern we raised in why your tools do not talk to each other: the gaps between systems are where time quietly disappears.

Stock charts in Sheets, and reading a number the right way

A chart is an argument about what a number means. The same figures drawn two different ways can lead a reader to two different conclusions, which is why the choice of visualisation is not a cosmetic one. A price that moved within a wide range during a period looks calm if you plot only its closing value and volatile if you show the full span. The chart decides which story the reader sees first.

Google added support for this kind of nuance in Sheets today. It said that "Google Sheets now supports new types of stock charts, making it easier to visualize price movement over time" [4]. A stock chart condenses several readings from each period into a single marker rather than a single point, so a reader can see the range rather than only the endpoint. For a finance team, the value is not the novelty of the chart type. It is that the right chart stops a reader from drawing the wrong conclusion. Choosing how to present a figure deserves the same care as choosing the figure itself, a point we developed in numbers that change a decision.

Customer satisfaction as a buying signal, and what it measures

Awards and satisfaction scores are evidence, but they are a particular kind of evidence, and it helps to know what they do and do not tell you. A satisfaction award reflects how existing customers feel, which is useful precisely because those customers have lived with the product rather than watched a demonstration. It does not tell you whether the product fits your specific needs, and it should sit alongside your own trial rather than replace it.

Xero reported winning Canstar's 2026 Most Satisfied Customers Award for small business accounting software in New Zealand. The post is worth reading for one line about why people buy accounting software at all: "Every small business owner I talk to has a different reason for taking the leap – but not one of them say it's for the paperwork" [5]. That is the real measure behind a satisfaction score. Owners are not chasing features for their own sake; they are chasing the hours and the calm that good tools give back. A high satisfaction score is a signal that a product delivered some of that to people who already use it. Treat it as one input among several, weigh it against a trial on your own data, and you have turned an award from a slogan into a piece of evidence. We set out how to combine signals like this in how to choose software worth using.

The thread, pulled tight

Four of today's five items are, underneath, the same story told in different registers. A benchmark asks what evidence should decide a purchase. An AI interviewer asks who should be doing the judging and at what scale. A format change asks whether your tools can carry evidence between them without loss. A satisfaction award asks how much weight a particular kind of evidence should carry. The useful habit for any buyer is to notice which kind of proof is in front of you and to ask the next question rather than the first one. A demo is a start, not a verdict. A benchmark score is a comparison, not a guarantee. A satisfaction award is a report from people who are not you. Keeping a record of how a tool behaves on your own work, and being able to export that record when you move on, is the quiet discipline that keeps all of these signals honest. It is also the reason we design 360REV so the audit trail and the export belong to you rather than to us.