Agentic Marketing Pipeline

I wanted to see how far an agent could take go-to-market — not a chatbot that answers questions, but a system that does the work. Point it at a company's website and get back the scaffolding a marketing team spends its first month building: who to sell to, how to reach them, and what to say.

The bet

Most "AI marketing" is a text box that writes a tweet. The interesting problem isn't the writing — it's everything upstream of it. Who is this company's ideal customer? Which real people match that profile? What's the brand's voice, and what should it actually publish? Get those wrong and the content is confident nonsense.

So the constraint was simple: one input — a URL — and a complete, grounded starting point out, with every step inspectable rather than a black box.

How it worked

The agent runs as a tracked, multi-step job — not a single prompt. Each stage writes its progress to the database so the UI can show the work as it happens, and each stage's output becomes the next stage's context:

• Read the business. Fetch and extract the site (or a pasted description / uploaded deck).
• Generate the ICP. An ideal customer profile — industries, company sizes, seniorities, job titles, pain points, tech stack — inferred from what the company actually does.
• Turn the ICP into people. Translate the profile into concrete buyer personas, then into real prospect-search filters, with an estimate of how many people each persona matches.
• Build the strategy. Content pillars, keywords, brand voice (tone, phrases to use and avoid), and differentiators — each grounded in the steps before it.
• Generate the content. Draft and repurpose social posts that fit the voice and the pillars, ready to review and post.

Because some of these calls are slow and the whole thing has to survive failures, it runs on a job queue with a background worker, structured-and-validated output at every step, and per-run cost metering.

What I learned

The model was the easy part. The engineering was making it reliable and inspectable:

• Orchestration beats one big prompt. Splitting the work into discrete steps — each with its own schema, validation, and retry — made the output far more dependable than asking for everything at once, and let the UI show real progress instead of a spinner.
• Ground every step in the last. Personas are only as good as the ICP; content is only as good as the voice. Passing structured context forward is what keeps the chain coherent.
• The hard part is the seam to the real world. Turning a fuzzy "ideal customer" into an actual prospect search meant mapping model output onto an external provider's exact filter schema — that unglamorous translation layer is where it gets useful instead of just impressive.
• Long-running AI needs real infrastructure. Queues, workers, progress events, and cost metering matter as much as the prompts once a job takes minutes and costs money to run.

What transfers

This is the shape of most useful AI work: a capability wrapped in orchestration, validation, and a clean seam to the systems that already run the business. The agent is only interesting because every step is grounded, inspectable, and cheap enough to run for real.