How do you calculate whether an AI automation is worth building?

The calculation people do has two terms: how long the task takes and how much the build costs. That version makes almost every automation look worthwhile, which is why so many get built and so few are still running a year later. The missing term is the recurring obligation — someone has to notice when it stops, notice when its output degrades, and update it when something upstream changes — and it does not go away.

So the honest form has four terms. Time saved per occurrence, which should be measured rather than estimated, because estimates of how long a routine task takes are consistently too high. Occurrences per month, which is the term that decides everything. Build cost, spread over the period you expect the automation to remain relevant. And the monthly cost of keeping it working, which for a settled automation is perhaps an hour and is considerably more in the first quarter.

Frequency dominates, and this is the most useful thing to understand about the arithmetic. The same build effort pays back several times faster on a daily task than a weekly one, and a monthly task almost never justifies a build no matter how long it takes. This inverts the usual instinct, which is to automate the task that feels most painful — and the painful task is often the rare one, because rare tasks are the ones nobody has a routine for.

There is a second benefit that belongs in the calculation and resists measurement: consistency. An automated step happens the same way every time, does not get skipped when someone is busy, and does not stop when a person is away. For processes where the real cost is the occasion it did not happen — the follow-up nobody sent, the check nobody ran — this can be worth more than the time saved, and it should be stated in the case rather than left as an unquantified bonus.

Against that, count the failure cost honestly. An automation that is wrong occasionally imposes a cost per error, and if nobody is checking, that cost is paid by whoever receives the output. Multiplying an error rate by a consequence is uncomfortable and it is the difference between a calculation and an argument for something already decided.

The practical shortcut for most small businesses: if the task happens daily and takes more than a few minutes, it is probably worth automating. If it happens monthly, it probably is not, whatever it costs. Between those, the arithmetic above is worth doing properly, and the term to be honest about is the recurring one.