Key Takeaways
An AI content pipeline built for a B2B SaaS client generates 120 content assets weekly — blog posts, LinkedIn articles, email sequences, ad copy, and social content — operated by one part-time content manager. Architecture: GPT-4o for ideation and variation, Claude 3.5 for long-form structured content, automated publishing via Make.com, and human review for brand voice. Cost: £2,100/month in AI and tooling vs £18,000/month prior cost for a 4-person content team. Quality metric: organic traffic up 214% in 6 months. This post details the full pipeline architecture, prompt templates, and quality control system.
The Problem: Content Demand Outpacing Team Capacity
Our client's content team of four was producing 18 assets per week and burning out. Leadership wanted 100+ assets across 7 channels. Hiring would cost £72k/year in salaries alone, before tools, management overhead, and the inevitable churn. We proposed a different architecture: reduce the human team to one operator and build AI systems around them.
The Pipeline Architecture
Layer 1 — Ideation Engine: GPT-4o consumes the client's performance data, competitor content, and keyword targets to generate a weekly content calendar. Layer 2 — Content Generation: Claude 3.5 for long-form (blog, LinkedIn, email). GPT-4o for short-form (social, ad copy). Layer 3 — Brand Voice Gate: human operator reviews 20% of output; AI learns from corrections. Layer 4 — Publishing Automation: Make.com sequences schedule and publish across all platforms.
The Prompting System That Makes It Work
Generic prompts produce generic content. Our system uses three prompt layers: (1) Brand DNA prompt — injected into every request, defining tone, audience, taboo topics, and formatting standards. (2) Content type template — specific structure requirements per format. (3) SEO and entity injection — target keywords, related entities, and competitor gaps pulled from our keyword API, injected per piece.
Quality Control Without a Full Team
The operator reviews 100% of the content calendar (a 2-hour Monday task) and spot-checks 20% of generated assets. Anything the AI produces that the operator corrects gets logged. Every 30 days, we run a prompt refinement sprint using logged corrections. The system improves month-on-month — the operator's corrections become training signal.
The Numbers After 6 Months
Previous cost: £18,000/month (4 FTE content team). Current cost: £2,100/month (AI tools) + £1,600/month (part-time operator). Saving: £14,300/month. Output: 18 → 120 assets/week. Organic traffic: +214%. Lead attribution to content: 34% of pipeline. The compounding effect is real — more content means more data, which means better AI output, which means more traffic.
Ready to implement this in your business?
Book a free AI Audit. 90 minutes. We'll map your highest-value opportunities and hand you a prioritised implementation plan.
Book My AI Audit