The bottleneck that was not headcount
This media company had two senior writers and a content director. They were producing one excellent weekly newsletter with a 71% open rate, genuinely strong for a B2B audience. The business wanted to launch four more vertical publications to serve adjacent audiences. The content director’s position: we cannot do this with two writers. It is not about speed, it is about depth. Each newsletter requires real editorial judgment about what is worth covering and how.
The content director was right about the judgment part. She was wrong about whether the other 60% of the work, the research, the source gathering, the first draft, the formatting, the scheduling, needed to come from the writers too.
The pipeline design
We designed a five-role agent pipeline:
Role 1: Researcher: Given a topic brief and a source list (RSS feeds, industry newsletters the editors trusted, Perplexity search for fresh developments), the researcher agent identifies the 12 most significant developments of the week in the vertical. Output: a structured brief with each item, its significance, and a source link.
Role 2: Editor: Takes the research brief and applies editorial judgment rules from a voice document the content director wrote. What kinds of stories fit this vertical? What does not? This is still an AI step, but it is constrained tightly by the editorial guidelines. Output: a prioritized list of 4 to 5 stories, with notes on why each was selected and a suggested angle.
Role 3: Drafter: Writes the newsletter body using the prioritized story list and the brand voice document. The drafter uses Claude Opus 4.5 (better voice consistency on long-form copy) rather than Sonnet. Output: a full draft in the newsletter format with headlines, intros, story sections, and a CTA.
Role 4: Human checkpoint: The draft goes to a writer for 45 to 60 minutes of editing. This is the only human step in the pipeline. The writer adds their perspective, catches any errors or tone misses, and makes judgment calls on items the AI flagged as uncertain (marked with [REVIEW] inline). No content publishes without this human sign-off.
Role 5: Publisher: Formats the approved draft for Beehiiv, schedules it, and generates the social variants for LinkedIn and X. Automatic after human approval.
The whole pipeline runs on a Monday trigger for each vertical, producing five drafts-ready-for-human-edit by Tuesday morning.
What the writers actually said
The content director was skeptical. After seeing the first draft outputs, her assessment was: “These are about what I would expect from a strong freelancer who had a good brief but had not read enough of the vertical.” That is a fair assessment, which is why the human editing step exists.
After three months, her revised assessment was: “The editing time has come down to 45 minutes from 80. The researcher is better than I expected. It finds things I would have missed.” The researcher agent’s ability to systematically cover more sources than a human would check in a research pass was a consistent surprise.
One writer went from producing one newsletter per week to handling two verticals. The other handles two verticals. The content director handles one (the flagship) plus editorial oversight. That is five newsletters with the same team.
The open rate question
The most common objection to AI-assisted content is that readers will be able to tell and will stop reading.
Month 1: open rates averaged 66% across the four new verticals, versus 71% for the flagship, which had the most established audience. Month 3: 68% across all five verticals.
The flagship maintained its 71% rate. The content director’s read on the new vertical numbers: “68% for a newsletter in its third month of existence is good. Better than we would have done launching manually with how stretched we were.”
One of the new verticals hit a 74% open rate in month 3, above the flagship. The reason: the researcher agent was systematically finding underreported stories in that vertical that the audience had never seen covered elsewhere.
The detail that keeps it honest
The pipeline catches factual errors about as well as a tired writer does, meaning not perfectly. We added a verification step between the Drafter and the human checkpoint where the agent checks every specific claim (numbers, dates, quotes) against its sources and flags any it cannot verify. This reduced the number of corrections the human writer needed to make.
The biggest miss in the first month was an agent-drafted headline that was slightly misleading about the scope of a regulatory change. The writer caught it before publication. The system now includes a “headline accuracy check” prompt that specifically tests whether the headline is supported by the body text.
Tech used
Mastra (multi-agent orchestration) · Claude Opus 4.5 (drafting) · Claude Sonnet 4.6 (research, editing, publishing) · Perplexity MCP · Beehiiv MCP · Buffer MCP (social scheduling) · n8n (trigger and scheduling) · Slack (draft delivery to writers)