Nobody demos a nightly batch job. There is no conference talk where someone opens a terminal, shows a scheduled task that has run without incident for nine years, and receives applause. The software that actually carries the world — payroll runs, settlement files, backup rotations, the integration that moves lab results to the doctor before morning rounds — is aggressively, deliberately boring, and it is the best software I know.
I say this as someone whose job title has the word automation in it, which means I spend my days in rooms where excitement is being sold. Every vendor deck promises transformation. Every proof of concept sparkles. And after fifteen years of building the things, I have arrived at an unfashionable conviction: excitement is a cost, not a benefit. The goal of everything I build is to become so predictable that everyone forgets it exists.
What boring is made of
Boring is not the absence of engineering; it is the visible result of an enormous amount of it. A boring integration is one where somebody thought about what happens when the same file arrives twice, so the job is idempotent and the duplicate is a log line instead of a double payment. Boring means the retry has a backoff, the backoff has a ceiling, and the ceiling has an alert. Boring means the error message contains the identifier you will actually need at 04:13 when it finally fires.
Exciting software, by contrast, is software that still has surprises in it. The demo is exciting precisely because nobody has yet paid the cost of discovering its edge cases. Every system is exciting when it is young; the discipline is in how fast you can make it boring, and the tragedy of my industry is that we reward people for producing excitement and then transfer the systems to someone else before the bill arrives.
Every system is exciting when it is young. The discipline is how fast you can make it boring.
Boring also has a sound, and the sound is silence of a very particular kind: the silence of alerts that were tuned until they only fire when a human should actually stand up. A monitoring channel that cries wolf forty times a day is exciting, and its excitement trains everyone to ignore it, which is how the forty-first alert — the real one — scrolls past unread. Making a system trustworthy and making its silence trustworthy are the same project. You are not done when the system works; you are done when nobody checks whether it worked.
The best operations engineer I ever worked with had a habit that taught me more than any book. Whenever we finished a build, she would ask one question: “What will wake me up?” Not what does it do, not how fast is it — what will wake me up. We would then spend a day removing reasons, one by one, until the honest answer was almost nothing. That day never appeared in any project plan, and it was the most valuable day of every project.
Boring in the age of probabilistic machines
This is the lens through which I look at the current wave of AI, and it makes me both more optimistic and more careful than the discourse around me. A large language model is the most exciting component I have ever been offered. It is also, by construction, a machine that does not do the same thing twice. Putting one into a business process is putting a brilliant, tireless, slightly unpredictable colleague into the middle of your payroll run. The question is not whether it is impressive. The question is my colleague’s question: what will wake me up?
The teams getting durable value from AI right now are, without exception, the ones doing the boring work around the exciting part. They write evaluations before they write prompts, so that “it seems better” becomes “it scores better on the four hundred cases we care about.” They put the model behind the same contracts we have always used — schemas on the way in, validation on the way out, a deterministic fallback when confidence is low. They log the model’s account of itself next to its output. They cap the blast radius, then widen it as evidence accumulates, exactly the way you would extend trust to any new operator.
None of this photographs well. All of it is the difference between a pilot that dazzles a steering group and a system that a nurse, a caseworker, or an accountant leans on at volume without thinking about it. The highest compliment a user can pay an AI system is the same one they pay a good dial tone: nothing. It just worked, again, so they thought about their actual job instead.
A defence of the metronome
I keep a metronome on my desk, an old mechanical one, as a small argument with the industry I love. It does one thing. It does it at the tempo you set, every time, for as long as you wind it. Nobody has ever been impressed by it, and every musician who has practised against one owes it a debt they rarely articulate. Boring instruments are how skill compounds; boring infrastructure is how societies compound. The glamorous things are built on top of ten thousand metronomes, ticking.
So this is my toast, raised in an industry that will never toast it back: to the batch jobs that ran last night. To the queue that drained, the certificate that rotated, the file that arrived twice and was counted once. To every engineer who made something so reliable that their own name faded from it. May your pager stay silent, may your dashboards stay green, and may the exciting people upstream of you never quite learn what carried them.