All field notes
// how we work·2026-05-19·4 min read

Your system gets better while you sleep

The strange economics of owning an AI system in a year when the models keep improving underneath it.

Here is a strange property of the systems we build, one that has no equivalent in any other capital purchase a small business makes: they improve after you've paid for them.

Your delivery van depreciates. Your POS system gets slower with every update. But the AI system reading your Bills of Lading this morning is running on a model that is measurably better than the one it launched with — better at handwriting, better at weird table layouts, better at the messy scan from that one supplier who photographs documents at an angle.

You didn't pay for that improvement. Nobody scheduled it. The model underneath got better, and your system inherited the gains.

We design for this deliberately. Every Quietwork build separates the plumbing (the integrations with your inbox, your Xero, your carrier portals — stable, boring, ours to maintain) from the reasoning (the model calls — swappable). When a better or cheaper model ships, upgrading is a config change and an afternoon of testing, not a rebuild.

This is most of what the Tune-Up retainer actually is, by the way. Not "maintenance" in the gym-membership sense — it's us re-testing your system against new models as they release, taking the accuracy gains, and often cutting the running cost at the same time, because newer models tend to be cheaper per task than the ones they replace. One system we run today costs 60% less per month than at launch and makes fewer errors.

The practical advice, whether you build with us or anyone else: ask your builder where the model lives in the architecture. If the answer is "everywhere" — if swapping it means rewriting the system — you're buying a depreciating asset. If the answer is "behind one interface," you're buying the only machine in your business that gets better while you sleep.

— THOMAS CHAN · FOUNDER, QUIETWORK