An AI content pipeline that publishes — audited, rebuilt, and handed back working

A cash home-buying business had commissioned three interconnected automation workflows meant to run a content pipeline from news ingestion through to blog drafting and social publishing. They believed it was nearly finished.

It wasn't running at all. Two of the three workflows had never once completed successfully — ten executions each, zero successes. One workflow's schedule trigger was blank and its wait step had no parameters, so any run that reached it hung indefinitely. Another crashed on every run from twelve expressions pointing at a node that had been renamed. The third ran but silently lost data: its category filter was disabled and inverted, and its parser only accepted a format its own prompt told the model not to use. Across all 57 nodes there was no error handling of any kind — no retries, no alerts, no failure logging.

We ran a formal technical audit of all 57 nodes first, delivered as a client-facing report, then rebuilt the pipeline around a spreadsheet acting as control panel and approval gate.

Each morning the first workflow pulls a real-estate news feed, has a model read and score every article for relevance to the business, and writes the qualifying ones into an asset library for a human to skim. Marking a row approved wakes the second workflow, which writes a full 1,800–2,200 word SEO blog post, generates three matching images, uploads them to the site's media library, creates the post as a draft, and stamps links back into the sheet. Marking a row approved for social triggers the third: concept, image and caption, image uploaded to obtain a public URL, published, and the row flipped so it can never post twice.

Nothing goes live without a human changing a status value. The automation produces drafts and queued work, never unreviewed output.

Model selection was treated as engineering rather than defaulting to the largest available: moving routine categorisation down a tier cut cost sharply with no measurable quality loss.