What I Chose Not to Automate

When building is cheap, the answer to "should we automate this?" drifts toward yes. Claude Code can write the integration, the email sequence, the auto-resolve. Three of the better calls I've made went the other way.

Kivo: a returned kit doesn't trigger a reship

The shipping queue on Kivo's kit-operations dashboard treated every returned kit as a reason to ship another. The query was doing exactly what it was told, and automating the reship would have been the easy fix. But a return almost always means the patient's program is over. So I took returns out of the queue, and a reship is now something the team decides on purpose. The queue answers who needs a kit. Whether someone gets another one is a call a person makes.

Lingisa: the agent doesn't get the last word

In Lingisa, you pin a comment to part of a 3D model and an AI agent makes the change. Early on, when the agent finished, its comment marked itself resolved. Technically that was accurate, since the work was done. But the person asked, so the person should decide whether it's right. Done work now waits in "awaiting review" until someone chooses Accept or Not quite… It costs one click, and it's what lets someone trust the agent with the next request.

Orchid Finance: no bank APIs

At first I wanted Orchid Finance to do everything: Fidelity, Plaid and bank APIs pulling transactions in automatically. Those integrations need real backend infrastructure, security compliance and ongoing maintenance, which would have eaten all the time I wanted to spend on design. So I cut them. Balances, income and expenses are entered by hand, and no financial data passes through a third party. The constraint made the app clearer: it only has to answer where the money is going and whether we're hitting our goals.

The test I use

These aren't the same decision three times, but they pass the same test. Automate the part that's a data problem: finding who needs a kit, applying the edit, adding up the month. Keep a person where the answer depends on context only they have, like whether a patient should get another kit, or whether a change is actually right. And be wary of automation whose real cost is maintenance you don't want to own.

Claude Code makes automation nearly free to build. It doesn't make it free to be wrong, and it doesn't decide where a person belongs. That part is still the design work.