Can AI run a daily newsletter with zero human input?

This one started with a conversation at Capital Factory, a co-working space in downtown Austin. I was talking to a biotech investor who wanted a way to get an automated daily digest of the biggest news and deals in the space. I'd built email newsletters before and was getting deep into AI automation at the time, so the question stuck with me — could I build a system that handles the entire editorial pipeline? Research, curation, writing, delivery. No human touching any of it.

The pipeline I put together works like this: every morning, an AI research agent uses deep research APIs to discover and summarize around twenty of the most important biotech stories published that day. It filters for the stuff that's actually interesting — not just press releases, but stories people would genuinely want to read. Then it formats everything into a clean email and delivers it through Resend. Each day's report feeds back as context for the next one, so the system learns not to repeat itself. A self-correcting loop.

The part that surprised me most had nothing to do with the newsletter itself. I wanted to figure out which AI research provider actually surfaced the best, most current news while following specific editorial instructions, so I built an eval system — a structured rubric that scored research outputs across multiple dimensions. I ran it against Perplexity, OpenAI, and a few others. OpenAI's deep research came out on top. Building the eval was honestly more engaging than building the newsletter.

The landing page is live at biotechmorning.com and the subscription flow works, but I haven't launched the actual newsletter yet. The reason is something the project taught me: "biotech" as a category is way too broad. Some people care about pure research and science. Others want pharma business deals. Others are tracking specific sub-fields like gene therapy or diagnostics. Making one daily email that serves all of those audiences well turned out to be the real problem — not the automation. Automating content generation is the straightforward part. Figuring out what's actually valuable to a specific reader is where it gets hard.

I still think there's a version of this that works, and the bones are solid. If I crack the targeting problem, the whole system is a template — swap out "biotech" for any industry and the same pipeline runs. That's what keeps me coming back to it.