From The Work
How I Built an AI-Human Creative System (And What Actually Worked)
What actually worked when building an AI-assisted creative workflow — the setup, the failures, and the principles that made it sustainable.

The conversation about AI in creative work has a problem. It's almost entirely about tools. Which model is best for copy. Whether to use Midjourney or Firefly. How to write a better prompt. These are useful questions but they're the wrong starting point.
The right starting point is the system. Not which tool to use but where in your workflow AI judgment should substitute for human judgment, where it should augment human judgment, and where it should stay out entirely. Get that architecture right and the tools almost pick themselves. Get it wrong and you end up with a faster way to produce work that still isn't any good.
I built an AI-human hybrid system from scratch at Moonbeam, a podcast discovery platform that was acquired by Audacy within twelve months of launch. The core performance metric we were optimizing for was session length — how long someone stayed in the app listening to content they hadn't specifically searched for. A 23-minute average session time, sustained at scale, is not a number you reach by accident. It's the output of a curation system that's been architecturally designed to surface the right content at the right moment to the right listener.
Here's how we built it, and what it taught me.
The Architecture Decision
The first question wasn't "how do we use AI?" It was "what decisions actually benefit from algorithmic judgment versus human judgment?"
At Moonbeam, the content discovery challenge had two distinct layers. The first was pattern matching at scale: given a listener's behavior history, surface content that matches their demonstrated preferences. This is a computation problem. A human curator cannot hold ten thousand listener profiles in their head simultaneously. An algorithm can. This layer belonged to the machine.
The second layer was quality and taste: not all content that matches a listener's pattern is worth their time. A podcast episode might be topically relevant but poorly produced, or accurate to a listener's history but not worth elevating. This is a judgment problem. An algorithm in 2021 — and honestly in 2026 — cannot reliably distinguish between content that's technically appropriate and content that's actually good. This layer belonged to humans.
The system we built reflected that distinction exactly. The algorithm surfaced candidates. Humans filtered for quality. The algorithm learned from the human decisions over time. Neither layer could have produced the result alone.
The Human Layer
The human curation layer is the part that gets cut first when organizations try to scale AI systems cheaply. It's also the part that determines whether the system produces good outcomes or just fast ones.
At Moonbeam, I designed the human layer around a specific question: what would a trusted friend who knew your taste recommend right now? Not a recommendation engine. Not a "users like you also listened to" algorithm. A friend.
The editorial criteria we developed for human curators were specific and teachable: production quality, narrative coherence, information density, distinctiveness of perspective. Each piece of content that advanced through the system had been evaluated against those criteria by a person, not just matched by a machine. That human filter is what created the session time. Listeners stayed because the quality threshold was real, not simulated.
What This Model Looks Like in Practice
Since Moonbeam I've applied the same architecture to creative production workflows in other contexts. The principle translates: identify the decisions that benefit from computational scale, identify the decisions that require human judgment, design the handoff between them deliberately, and build the feedback loop that lets the system learn.
In a content production context this usually looks like: AI handles first-draft generation, format variation, metadata creation, and distribution scheduling. Humans handle strategic judgment (does this serve the brief), brand voice (does this sound like us), and quality assessment (is this actually good). The workflow is designed so AI output feeds into human review rather than bypassing it, and human decisions feed back into the AI layer as training signal.
The mistake most teams make is using AI to replace the human review step rather than to make the human review step faster and more scalable. That's how you get content volume without content quality. The 23-minute session time at Moonbeam happened because the quality floor was maintained by humans even as the volume scaled algorithmically. Remove the humans and the floor collapses.
The Current Toolkit
The specific tools I'm using now are less important than the architectural principles, but for context: prompt engineering for brief development and first-draft copy; custom workflow automation (using local inference via Ollama for internal tasks that shouldn't touch third-party APIs); structured AI review protocols for content QA; and human-in-the-loop checkpoints at every stage where brand voice or strategic judgment is involved.
The stack changes. The architecture doesn't.
What Actually Works
Three things I'd tell anyone building this for the first time:
Define the handoff before you pick the tools. The most important design decision in an AI-augmented workflow is where the human takes over. Design that boundary explicitly, not by default.
Protect the quality floor. AI is very good at producing volume. Volume without a quality filter produces noise. The human layer isn't a bottleneck in your workflow — it's the mechanism that makes the workflow produce outcomes worth having.
Build the feedback loop. An AI system that doesn't learn from human decisions is a static system. Static systems have a ceiling. The compounding value of an AI-human workflow comes from the machine getting smarter over time based on what the humans are approving and rejecting. Design for that from the beginning.
The tools are interesting. The system is the thing.


