What We Automated This Year, and What We Left Alone

We run a content operation across a number of brands, several platforms and a weekly cadence that does not pause. Over the past year we have automated a significant amount of it. We have also deliberately declined to automate several things that would have been straightforward to build, and those choices have been more consequential than the ones we made.

Publishing and verification

Scheduling content across platforms was the first thing to go. It is repetitive, the rules are fixed, and a machine does not get bored at the eleventh upload.

More valuable, and less obvious, was automating the check afterwards. Scheduling something is not the same as it publishing, and the gap between the two is where quiet failures live: a token expires, an upload silently fails, a post goes to the wrong account. We now verify every scheduled item on the platform itself, after the fact, rather than trusting the scheduler's own report. That single habit has caught more problems than the scheduling saved time.

Finding things

The other clear win was retrieval. Across years of production, the question is rarely whether an asset exists. It is whether anyone can find it before it is faster to reshoot. Indexing footage, transcripts and deliverables so they are searchable has repeatedly turned a half-day hunt into a minute, and it required no clever technology, only the discipline of consistent naming.

What we would not hand over

Anything that reaches a client goes out only after a person has read it. Not because a machine cannot write a competent message, but because the cost of a wrong one is not symmetrical with the time saved. The same applies to anything that publishes publicly under a brand we do not own.

We also keep a person on final creative judgement. Automation is genuinely useful at generating options, at rough assembly, at the first ninety percent of a repetitive edit. Deciding which option is right, and whether the piece actually works, has not been a good candidate.

The rule we settled on

Automate the step, keep the human on the gate. A process can be entirely automated up to the point where something becomes irreversible or public, and then it should stop and wait for a person. In practice that means most of our pipelines end in a draft, a queue or an email rather than in a published thing.

It is slower than full automation by a few minutes a day and has saved us from every category of error we would have been embarrassed by. That trade has been worth making every time.