Building Better Workflows with Automation

Most automation advice focuses on what tools to use. The more useful question is usually structural: which parts of a workflow are actually safe to automate, and which parts only look mechanical from a distance but are quietly doing more work than they appear to.

Fact: the workflows that automate cleanly tend to have a specific shape — a clear trigger, a deterministic or near-deterministic set of steps, and a low cost if something goes slightly wrong. Sending a routine notification, moving a file between systems, generating a first-draft summary, flagging an item that matches known criteria — these hold up well under automation because the downside of an occasional miss is small and easy to catch.

Where automation quietly breaks things

The failure pattern that shows up most often isn't automating something that obviously required a human — it's automating a step that looked mechanical on the surface but actually contained a judgement call nobody noticed was happening. A "simple" approval step that occasionally required someone to weigh context not captured anywhere in the system. A categorisation rule that worked for the vast majority of cases and silently mishandled the remainder in a way that compounded over time. These failures are dangerous precisely because they're invisible until something downstream goes wrong.

Analysis: a useful practical test before automating a step is asking how often a human doing that step actually deviates from the "obvious" rule, and why. If the answer is essentially never, and when they do it's for a documented, specific reason, it's a strong automation candidate. If the answer is more often than you'd think, usually for reasons that aren't written down anywhere, that's a signal the step is quietly doing more judgement work than it looks like, and automating it without addressing that will cause real problems eventually, even if it looks fine at first.

Opinion: the workflows that hold up best over time aren't the ones automated most aggressively — they're the ones automated deliberately, with a clear, honest answer for where the remaining judgement calls live and who's responsible for them. "We automated everything" is a less impressive claim than it sounds if nobody can say confidently what happens when an edge case shows up.

Prediction, held loosely: as AI makes it easier to automate increasingly judgement-adjacent steps, the practical skill that matters more, not less, is being precise about where genuine judgement is still required and building deliberate checkpoints for it — rather than assuming a system that handles the common case well has actually handled the whole workflow.