This week in numbers (real, from our system)
- 🤖 AI agents running: 19
- 📝 Content published: 53 (blog RU 14, EN 21, Altezza 18)
- ⚙️ Generated programmatically: 4
- 📥 Leads in the system: 246 (+0 in the last 7 days)
Short answer: We run a content factory at Arxitek almost entirely on AI agents — research, writing, SEO optimisation, internal linking and publishing are all handled by automated pipelines. Building in public means we share exactly how it works, what breaks, and what we fix week by week, so other small and mid-sized businesses can replicate the approach without the trial-and-error cost.
Here's the thing: most businesses treat content as a manual chore. Someone writes a brief, a copywriter drafts something, an editor rewrites half of it, an SEO specialist adds keywords, and then it sits in a queue for two weeks. By the time it publishes, the topic has moved on.
We decided to do it differently. We built a content factory powered by AI agents, and we're documenting the whole process in public — the wins, the failures, the messy middle. This is not a polished case study. It's a living diary. If you run a small or mid-sized business and you're trying to figure out whether content automation is worth your time, this is the article you actually need.
What Does "Building in Public" Actually Mean for a Business?
Building in public is a commitment to radical transparency about your process. Instead of publishing a glossy retrospective once a year, you share the work as it happens — the architecture decisions, the tools you're testing, the things that didn't pan out. It's a practice that originated in the startup and indie-developer world, but it translates perfectly to any business running an operational experiment.
For us, it means this weekly diary. Every week we document what our AI agents did, what changed in the pipeline, and what we learned. We share it on the Arxitek blog because we believe the audience we want to reach — owners and marketing directors of small and mid-sized businesses — deserves honest, engineering-level transparency rather than marketing copy.
Why does this matter beyond the content itself? Because building in public creates a feedback loop. Readers spot edge cases we missed. They ask questions that expose gaps in our thinking. They share their own experiments. The diary becomes a community resource, not just a brand asset.
There's also a practical SEO angle. A regularly updated diary signals freshness to search engines. Each weekly entry adds long-tail keyword coverage organically. And the honest, first-person format tends to earn backlinks from people who reference real-world experiments rather than theoretical guides.
Let's be real about the vulnerability, though. Sharing what breaks is uncomfortable. There's a temptation to only publish the success stories. We resist that temptation deliberately, because the failures are where the actual learning lives. If you're considering building in public yourself, that's the mindset shift you need to make first.
The format also forces discipline. When you know you're going to write about your process every week, you document it properly as you go. You keep better logs. You write clearer internal notes. The act of narrating the work improves the work. That side effect alone is worth the effort.
How Does Our AI Agent Pipeline Actually Work?
Let me walk you through the architecture without the jargon. Think of it as an assembly line where each station is an AI agent with a specific, narrow job.
Stage 1: Topic and Keyword Research
The first agent monitors search trends, competitor content gaps, and our own analytics to surface topic candidates. It doesn't just pull keywords — it clusters them by intent and maps them to stages of the buyer journey. The output is a prioritised content brief with the target keyword, secondary keywords, search intent, recommended structure, and a short competitive analysis. A human (me, usually) reviews the brief and approves or adjusts it. This is the only mandatory human checkpoint before production begins.
Stage 2: Research and Source Gathering
Once a brief is approved, a second agent goes to work gathering source material. It searches for recent industry data, relevant studies, and authoritative references. It does not fabricate statistics. It flags when data is absent and suggests qualitative framing instead. This is a non-negotiable design principle in our pipeline: honesty over plausibility. The agent produces a structured research pack that the writing agent draws from.
Stage 3: Writing and SEO Structuring
The writing agent takes the brief and the research pack and produces a draft in our defined editorial voice. It applies the GEO (Generative Engine Optimisation) structure we've developed — short-answer blocks, question-format headings, FAQ sections — because we're optimising not just for Google but for how AI-powered search tools surface answers. The draft lands in our CMS as a review document, not a published post.
Stage 4: Internal Linking and Metadata
A separate agent scans the draft against our existing content library and suggests internal links. It also generates the slug, meta description, and Open Graph tags. This sounds minor but it used to eat a surprising amount of editorial time. Automating it means the human reviewer focuses on substance, not mechanics.
Stage 5: Review, Publish, and Monitor
A human editor — again, often me — does a final read. We're not proofreading for typos at this point; the agents handle that. We're checking for voice consistency, factual integrity, and whether the piece actually serves the reader. Once approved, a publishing agent schedules and deploys the post, pings our distribution channels, and sets up a monitoring task to track ranking movement over the following weeks.
What Breaks (and How We Fix It)
I'd be lying if I said the pipeline runs smoothly every week. Here's an honest log of recurring failure modes and our current fixes.
Voice drift is the most common issue. When the writing agent processes a brief that's slightly outside our usual topic range, it sometimes defaults to a more generic, corporate tone. Our fix is a voice-calibration prompt that references a set of approved exemplar paragraphs. It's not perfect, but it catches the worst drift before review.
Hallucinated specificity is the one we take most seriously. Occasionally an agent will generate a sentence that sounds like it contains a real statistic but is actually a plausible-sounding fabrication. We've built an explicit instruction into the writing agent's system prompt: if you don't have a sourced figure from the research pack, describe the phenomenon qualitatively. We also do a manual spot-check on any sentence containing a number. This adds a few minutes to review but it's non-negotiable.
Internal linking loops happen when the linking agent recommends circular references — page A links to page B which links back to page A in an unhelpful way. We added a deduplication check and a rule that a link must add genuine navigational value, not just keyword proximity.
Scheduling conflicts occasionally cause two posts to publish on the same day when the queue is busy. Simple fix: the publishing agent now checks the existing schedule before confirming a slot.
The broader lesson here is that AI agents are not set-and-forget systems. They're more like junior team members who are extremely fast and extremely literal. They do exactly what their instructions say, which means the quality of your instructions determines the quality of the output. Invest in your prompts the way you'd invest in an onboarding process.
What Content Automation Actually Frees Up (and What It Doesn't)
Let me be direct about this, because there's a lot of hype around AI replacing content teams. That's not what's happening here, and I think it's important to say so plainly.
Content automation frees up the mechanical, repeatable parts of the content process: keyword research, first-draft writing, metadata generation, internal linking, scheduling, basic distribution. These tasks are time-consuming, low-creative-value, and error-prone when done manually under deadline pressure. Handing them to AI agents means the humans on the team spend their time on the things that actually require judgment.
What it doesn't replace: editorial taste, strategic thinking, relationship-driven content (interviews, partnerships, community pieces), and the final quality gate. A human still decides what the business should be known for. A human still decides whether a piece is honest, useful, and worth the reader's time.
The framing I use internally is this: the AI agents handle the factory floor. The humans handle the design studio. The factory produces volume and consistency. The design studio produces direction and quality standards. Both are necessary. Neither replaces the other.
For small and mid-sized businesses, this matters enormously. You probably don't have a six-person content team. You might have one person wearing three hats. Content automation doesn't give you a team — it gives that one person the leverage to produce at team-scale without burning out.
How Does This Approach Compare to Hiring a Content Agency?
This is the question I get most often, so let me address it directly.
| Factor | AI Agent Pipeline | Traditional Content Agency |
|---|---|---|
| Setup time | Weeks to months (pipeline build) | Days (onboarding call) |
| Ongoing cost | Tool subscriptions + oversight time | Monthly retainer (varies widely by market) |
| Voice consistency | High once calibrated | Depends on writer turnover |
| Speed | Hours per post once running | Days to weeks per post |
| Strategic input | Requires internal ownership | Often included in senior packages |
| Transparency | Full (you own the pipeline) | Varies; often a black box |
| Scalability | High — volume scales without linear cost increase | Linear — more posts = higher cost |
| Failure mode | Technical drift, hallucination risk | Quality inconsistency, writer churn |
Many of the businesses we work with at Arxitek end up doing a hybrid: they use an agency for high-stakes, high-visibility content (thought leadership, major campaign launches) and run the AI pipeline for the high-volume, long-tail SEO content that would be uneconomical to commission manually.
What's on the Roadmap for the Pipeline?
Building in public means sharing what's next, not just what's done. Here's what we're actively working on.
Multimodal content generation is the next frontier. Right now the pipeline produces text. We're testing agents that can brief and prompt image generation for featured images and social cards, reducing the design bottleneck without removing human creative oversight entirely.
Automated performance feedback loops are something we're prototyping. The idea is that the monitoring agent doesn't just track rankings — it feeds performance data back into the topic-research agent, so the pipeline learns over time which content types and structures perform best for our specific audience. This is the move from automation to genuine machine learning within our own content system.
Personalised content variants are on the longer-term roadmap. The same core article, automatically adapted in tone and depth for different audience segments — a technical version for developers, a plain-language version for business owners. The infrastructure for this is mostly in place; the challenge is quality control at scale.
Sales content integration is something we're piloting right now. AI agents that generate personalised outreach sequences, proposal summaries, and follow-up content based on CRM data. This is where content automation starts to blur into sales automation, and it's genuinely exciting territory.
Frequently Asked Questions
What tools do AI agents typically use for content automation?
Most production-grade content automation pipelines combine a large language model (such as GPT-4o, Claude, or Gemini) with an orchestration layer that manages the workflow — popular choices include n8n, Make, and custom API integrations. On top of that, you'll usually see SEO data connectors (Semrush, Ahrefs, or similar), a CMS integration (WordPress, Webflow, Contentful), and a monitoring layer for post-publish analytics. The specific stack matters less than the architecture: clear agent responsibilities, explicit quality gates, and human review at the right checkpoints.
Is building in public a good strategy for small businesses?
It can be, but it requires commitment to honesty. Building in public works best when you have a genuine operational experiment to document — a new process, a new tool stack, a new market approach. It builds trust, generates organic content, and creates a feedback loop with your audience. It's less effective if used purely as a marketing tactic without real substance behind it. For small businesses with limited content budgets, it's an efficient format: the work you're already doing becomes the content.
How long does it take to set up an AI agent content pipeline?
Realistic timelines vary considerably depending on your technical resources and the complexity of your existing content workflow. A basic pipeline — brief approval, first draft, metadata, publish — can be functional within a few weeks if you have technical support. A fully optimised pipeline with performance feedback loops, voice calibration, and multi-channel distribution typically takes several months of iteration. The first version will be imperfect. That's expected. The goal is to get something running and improve it continuously, which is exactly what building in public documents.
What are the main risks of automating content with AI agents?
The three risks we encounter most often are: hallucinated specificity (agents generating plausible-sounding but unsourced facts), voice drift (output that doesn't match your brand tone), and over-automation (removing human judgment from checkpoints where it's genuinely needed). All three are manageable with well-designed system prompts, explicit quality rules, and mandatory human review gates. The risk of under-automation — spending hours on mechanical tasks that agents could handle — is equally real and often underestimated.
Can content automation work for businesses without a technical team?
Yes, increasingly so. The no-code and low-code automation platforms available in 2025-2026 have made it possible to build functional AI agent workflows without writing code. Tools like Make and n8n offer visual workflow builders with pre-built AI integrations. The limiting factor is usually not technical skill but process clarity: you need to know your content workflow well enough to describe it precisely before you can automate it. If your current process is ad hoc and undocumented, start there before reaching for automation tools.
Conclusion
Building in public is uncomfortable in the best possible way. It forces honesty, discipline, and continuous improvement. Our AI agent content factory is not finished — it's a living system that we're iterating on every week, and documenting that iteration is part of the work.
If you're a business owner or marketing director trying to figure out whether content automation is worth the investment, I hope this diary gives you something more useful than a polished pitch deck: a real picture of what the work actually looks like.
We're not selling a perfect solution. We're sharing a real process, openly, because we think that's more valuable.
If you want to talk through how an AI agent pipeline might work for your specific business — your content volume, your team size, your existing tools — get in touch with the Arxitek team. We're happy to have a practical conversation, no sales script required.