The AI Workflow I Killed Last Week

A workflow I built eight months ago is now dead. I killed it on April tenth. Writing this as a post mortem so I do not build its cousin by mistake.

The business problem was legitimate. A client needed weekly content summaries pulled from ten sources and delivered in a slide deck. I built an agent that did the crawling, summarization, and deck assembly. It ran every Friday.

What went wrong

Nothing broke. That was the problem. It ran for eight months, nobody looked at the deck, and we only noticed when the client asked to unsubscribe from their own report.

The failure

I optimized the wrong thing. The deck was beautiful. The pipeline was reliable. The audience was empty. No amount of engineering fixes that.

What I should have done

Ship a manual version first. A shared doc. Three bullet points. Zero automation.
Wait for the first complaint that it was too slow or too inconsistent.
Automate only the parts people asked for.

The rule I added to my notes

Do not automate a thing until someone has complained about doing it manually. Automation without a complaint is a solution looking for a problem.

The cost

Thirty hours of build time
Six dollars a week in API costs
One embarrassed conversation

What replaced it

A weekly email. I write it in ten minutes using Claude to draft a first pass. No pipeline. No scheduler. The client reads every one.

The lesson

Engineering quality does not create audience. It only amplifies what is already being pulled for. See Shreyas Doshi on pull vs push for the framing I now use. Kill your darlings. Keep the newsletter.