
A hundred and fifty writers produced every piece by hand. The system drafts, illustrates and narrates it now, and a person signs off before anything publishes.
A marketing business sells the judgement of its writers. The writing itself was never the problem. What cost them was the number of hours between a brief and something ready to publish.
Years of habit ran the pipeline, and it ran a little differently for every client. The people who knew it were the people doing it. So the first job was finding out what actually happened, one client at a time.
Their customers buy writing that a person wants to read. A draft that announces itself as machine written loses the reader and the search result in the same moment. The value of the whole system depended on output nobody could pick out as generated.
A piece is not finished when the text is. It needs images for the blocks, a narrated version, and video cuts at the lengths a social post takes. A tool that only writes solves one quarter of it.
The writers are what the client sells, and their judgement is what customers pay for. Remove them and the product goes with them. Approval had to be part of the design rather than a setting somebody turns on.
The team works in several languages and across a spread of time zones, some in the office and most not. Anything built for them had to be one system rather than a tool per office. A hundred and fifty people were never going to be retrained into one house method.
Output ran at three or four pieces a week and it had run there for years. A customer asking for more meant hiring somebody, which is the point at which growth costs more than it returns.
The proposal was full automation with the people left in it, which sounds like a compromise and is the only version that works here. The answers run in the order the problems were listed.
We sat with the team before proposing anything. Every stage was written down as it actually ran, client by client, including the parts nobody thought were worth mentioning. That document became the specification, and it was theirs whatever they decided to do next.
The model was trained against the way this team already writes rather than against the internet. A draft arrives in their voice and a writer edits it, which is a different job from rewriting a stranger's work. The output does not announce itself, and that was the entire point of the exercise.
Images for the blocks, a narrated audio version and video summaries all generate inside the editor. Video runs at two minutes for the long cut, and at fifteen, eight or five seconds for the short ones. The writer clicks rather than briefs four different suppliers.
The system stops and waits for a person. A writer verifies the text and the images before anything moves on, and nothing publishes without that. It runs the whole pipeline and then stops, one step short of publishing, on purpose.
Everybody works in the same tool, in the office or a long way from it. Language is a property of the piece rather than a separate installation, so a brief in one language travels the same rails as the next.
A writer starts from the brief and the system produces the piece with its media attached. What is left for them is the judgement, which is the part they were hired for. The piece goes out on a click.
A month of content now takes a week, and the people who used to produce it are doing work that needed them more.
Three or four pieces a week was the ceiling. What used to be a month of work for a customer now lands inside a week. Nobody was hired to make that happen.
The saving is theirs to measure and they put it in the thousands of hours. The people it freed were moved onto work the business had been putting off rather than shown the door.
A writer signs every piece before a customer sees it. That was the condition the whole design was built around, and it survived contact with the schedule.
Tell us about the app you want to build. We send back a costed plan with a fixed price and a launch date, free and yours to keep.

