The Cost of Looking

9 July 2026 — Another week of releases, another round of price cuts, and for most small businesses the same unanswered question sitting underneath all of it: how would you know if any of this were working?

That question is worth more attention than the news, because the news is now abundant and cheap while evidence remains scarce. Model capability improves monthly. Inference costs fall. Tools proliferate. None of it tells a fifteen-person firm whether the invoicing process got faster, or whether the customer response got better, or whether the thing they deployed in March is still doing what they think it is.

There is a reason this gap persists and it is not laziness. Measuring an automation properly requires a baseline taken before you started, and by the time you think to measure, the before-state is gone. The workaround is unglamorous: for every process you automate from now on, record three numbers first — time taken, error rate, and who does it. They take an hour to gather and they are the only thing that will let you answer the question honestly later.

The regulatory thread kept its shape. High-risk obligations arrive in August with no small-business exemption, and the classification exercise depends entirely on an inventory most organisations have not built. Compliance and measurement turn out to be the same discipline viewed from two angles: both require you to know what you are running and be able to prove what it did.

What this means if you run an SME. Pick the process you are most tempted to automate. Before you touch it, spend an hour writing down how long it takes, how often it goes wrong, and what it costs when it does. That page is worth more than any tool evaluation, because it is the only thing that will tell you — in six months, when the novelty has worn off — whether the decision was right.

Weekly review from the mid-week signal archive.