Sixteen Workflows, One Small Server
I spent the summer automating my own operation before automating anyone else's. Sixteen workflows, a €8 box, a 0.9% failure rate — and one rule that turned out to matter more than any of the engineering.
There is a question I would put to anyone selling automation, so it seemed fair to put it to myself first: does your own operation run on the thing you are recommending?
Since the start of the summer, mine does. Sixteen workflows now run the administrative half of my working life — lead research, follow-ups, content distribution, invoice extraction, monitoring — on a single small server that costs about €8 a month and also hosts the database. Across 213 production executions the failure rate is 0.9%, and the average run takes 1.46 seconds.
Those are small numbers and I want to be careful not to dress them up. This is not a platform. It is one person’s back office, instrumented.
But the exercise answered a question I have been circling for two years, which is which work is actually automatable — and the answer turned out to have almost nothing to do with the difficulty of the task.
The cheap part was never the compute
The AI usage is close to embarrassing to report. A daily digest costs roughly a cent. A researched dossier on a new contact, a few cents. Extracting the fields from an invoice, about $0.011.
Eleven-tenths of a cent. That is the marginal cost of the thing that, five years ago, would have been a junior person’s Tuesday morning.
The expensive ingredient is knowing what to build, in what order, and which failure modes to handle. That is not a technology cost. It is a judgment cost, and it has not fallen at all — if anything the collapse in the price of execution has made judgment scarcer, because there is now nothing stopping you from building the wrong thing very quickly.
The rule that mattered more than the engineering
Everything that goes outward passes a human gate.
The machine researches, drafts, summarises, proposes. A person clicks send. Not one of the sixteen workflows publishes, emails, or commits anything to another human being without me looking at it first. The follow-up after a call arrives in my drafts, in something close to my voice, and I still read it before it goes.
This started as caution and turned into the design principle. Once you decide the machine prepares and a person commits, a whole class of questions resolves itself — what to do about tone, about errors, about the thing an automated system says on your behalf while you are asleep. You are never in that position, so you never have to answer for it.
The corollary is less comfortable. If a workflow’s only value came from removing the human at the end, it was not a good candidate for automation. It was a bad process wearing a costume.
What I would build first, again
The monitoring. Failures alert my phone within seconds, definitions back up nightly off-site, and a daily check confirms the containers are up, the disk has room, the certificate has not expired, and last night’s backup actually landed.
Boring, and the first thing I would rebuild. Automation you are not watching is not leverage. It is a liability with a cron schedule.
The full teardown — every workflow, the costs, the design rules, and the mail provider that silently ate my approval emails because of a brand name in the subject line — is on the Marain site.