An AI agent that saves ~20 hours/week
It doesn’t just write, it questions itself
For a research-focused YouTube channel I built an AI system that proposes, weighs and prepares content, saving about 20 hours of work every week. The hard part wasn’t speed, but reliability: an answer that looks right but isn’t is more dangerous than an obviously wrong one, because it slips through unnoticed. That’s why the agent questions its own work, and the human keeps the final say.
The context
A client runs a YouTube channel with research-focused content. Every week, just to produce one video, dozens of small decisions had to be made before the first line was even written. Which topic is worth it? Which isn’t? What would genuinely interest people? And then the heavier work began: research, script, narration, visuals, sources. Whole hours every week that ate up time and energy.
I built him a digital assistant powered by Artificial Intelligence that takes over this flow: it finds topics, helps with the selection, and prepares the material: the script, the filming instructions, the title, the description, the video’s thumbnail, even the sources. The goal wasn’t “something that writes instead of him”, but something that spares him the repetitive work and the small, tiring decisions.
The challenge
At first the output was impressive: fast, well-made, often very useful. Until, while testing it, it prepared a script about a Meta court case. Everything looked right, even the confidence of the text. With just one problem: the information wasn’t true. It wrote that the trial would take place, when the case had closed with a settlement, shortly before it.
The danger wasn’t that it said something plainly wrong. It was that it said something almost right, and that is more insidious, because it doesn’t put you on guard. That’s when I understood what the system really needed: not something that merely writes, but something that questions. Because Artificial Intelligence always sounds certain, even when it doesn’t know. And certainty doesn’t mean it’s right.
The solution
I built a system that doesn’t just offer ideas, but weighs them. Each week it proposes topics and scores them on axes that matter in practice, from how strong the story is to how reliable the sources are, and surfaces the best. (The exact way it scores stays my own know-how.) And I added a missing step: before anything is delivered, the information is automatically checked for currency, automatically cross-checked across multiple sources, and finally passed through human eyes. A little slower, far more certain.
The results
- Today the assistant saves the client about 20 hours of work every week, steadily, again and again: research, writing and a host of small decisions.
- Every human correction becomes a new criterion: the agent tunes ever closer to what the channel wants.
- The errors that “look right” are caught before they go live, not after.
- Time and energy now go to judgement and creation, not to drudgery.
What this means for you
The hard part wasn’t technical but building something that speeds up the work, without the human stopping to think. That’s what I bring to an organisation: not a machine that just churns things out, but a way for Artificial Intelligence to work alongside people, carrying the load off them. Because Artificial Intelligence can produce; judgement, though, stays human.