This week in numbers (real, from our system)
- 🤖 AI agents running: 19
- 📝 Content published: 69 (blog RU 23, EN 25, Altezza 21)
- ⚙️ Generated programmatically: 4
- 📥 Leads in the system: 264 (+0 in the last 7 days)
Short answer: At Arxitek we run a content factory powered by AI agents that handle research, drafting, SEO optimisation and analytics — all in a transparent, building-in-public loop. No magic, no hype: just a set of specialised bots, clear handoffs between them, and a human editor who keeps the whole thing honest. The result is a scalable content operation that any small or mid-sized business can replicate.
Let's be real: most companies talk about AI automation in the abstract. "We're exploring AI." "We're piloting something." What they almost never do is show the actual plumbing — the agents, the prompts, the failures, the weekly rhythm. That's exactly what this diary is.
Every week we publish a transparent look at how our own content factory runs. What the AI agents did, where they got stuck, what a human had to fix, and what the output looked like. If you're a business owner or a marketing director trying to figure out whether content automation is real or just vendor hype, read on. This is the honest version.
What Does "Building in Public" Actually Mean for a Business?
Building in public is a practice borrowed from the indie-maker world: you share your process, your wins, your mistakes, and your metrics as you go — not in a polished retrospective six months later, but in real time. For a software startup, that might mean tweeting your MRR every month. For us at Arxitek, it means documenting exactly how our AI-driven content pipeline works, week by week, warts and all.
Why bother? A few reasons, and they're all practical.
First, it keeps us honest. When you commit to showing your work publicly, you can't quietly sweep a broken workflow under the rug. If an AI agent produces garbage output three weeks in a row, that goes in the diary. That kind of accountability forces faster iteration.
Second, it builds trust with the exact audience we want to reach — small and mid-sized business owners and marketing directors who are sceptical of AI hype. Showing the real process, including the rough edges, is more convincing than any polished case study.
Third, it's genuinely useful content. The most-read posts in our blog aren't the theoretical ones. They're the ones where we say: "Here's the exact agent setup we used this week, here's what broke, here's how we fixed it." People can take that and apply it to their own operations.
Building in public isn't a marketing stunt. It's a discipline. It requires that you actually have something real to show — a working system, not a demo. That's the bar we hold ourselves to.
Why Transparency Matters More Than Polish
There's a temptation, especially in B2B, to only share success. To wait until everything is clean and the numbers look good. Resist that. The businesses that follow our diary tell us the most valuable entries are the ones where something went sideways. A prompt that produced off-brand copy. An SEO agent that over-optimised a headline into nonsense. A scheduling bot that published a draft instead of a final.
Those failures are the curriculum. They're what help a marketing director at a mid-sized company decide whether to invest in content automation — and how to avoid the obvious mistakes.
How Are Our AI Agents Structured Inside the Content Factory?
Here's the thing: "AI agents" is a term that gets thrown around loosely. Let me be specific about what we mean.
We run a multi-agent pipeline where each bot has a single, well-defined job. There's no one mega-prompt that does everything. That approach fails. Instead, we have a chain of specialised agents, each handing off to the next, with checkpoints where a human can intervene.
The Research Agent
This agent's job is to gather context before any writing happens. It pulls current search trends for a given topic, identifies the questions people are actually asking (the kind that show up in "People Also Ask" boxes and AI-generated answers), and maps out the competitive landscape for that keyword cluster. It doesn't write a single word of the article. Its output is a structured brief: target keywords, angle, key questions to answer, approximate word count, and a list of facts that need to be verified by a human before they go into the piece.
The research agent is the one that saves the most time. Manual keyword research and competitive analysis used to take a significant chunk of a content strategist's week. Now it's a fraction of that, and the brief is more thorough than what most humans produce under time pressure.
The Draft Agent
The draft agent takes the brief and produces a first-pass article. It writes in the house voice — direct, first-person, no corporate fluff — because that voice is baked into its system prompt. It follows the structural template: short answer block at the top (for AI search visibility), intro, H2 sections, FAQ, conclusion.
The draft is never published as-is. That's a rule we don't break. The draft agent's output is raw material. It's good raw material — structured, on-topic, the right length — but it needs a human pass for accuracy, voice calibration, and the kind of judgment calls that bots consistently get wrong (knowing when a claim needs a caveat, when an example is too generic, when a section is technically correct but misses the point).
The SEO Agent
Once the human editor has revised the draft, the SEO agent runs a pass focused purely on technical optimisation. It checks keyword density and placement, ensures the target keywords appear naturally in the title, at least two H2s, and the first paragraph. It reviews the meta description for length and click-worthiness. It flags any heading that's too long, too vague, or unlikely to surface in a featured snippet.
Critically, the SEO agent does not rewrite for SEO at the expense of readability. Its job is to flag, not to override the editor. If a keyword placement would make a sentence awkward, the agent flags it and offers an alternative — the human decides.
The Analytics Agent
This one runs after publication. It monitors performance data — organic impressions, click-through rate, average position, time on page — and surfaces anomalies. If an article that was ranking well drops suddenly, the analytics agent flags it for review. If a piece is getting impressions but low clicks, it suggests headline variants to test. It doesn't make changes autonomously. It generates a weekly report with prioritised recommendations.
The Scheduling and Distribution Agent
The final agent in the chain handles distribution: posting to the blog CMS, formatting for the email newsletter, creating social media snippets for LinkedIn and Telegram. It works from a content calendar that a human sets at the start of each month. It doesn't decide what to publish or when — that's a strategic call that stays with the team. It just executes the logistics.
What Does a Typical Week in the Content Factory Look Like?
I want to give you a concrete picture of the weekly rhythm, because "we use AI agents" is meaningless without the operational reality.
Monday: The research agent runs briefs for the week's planned articles. A human content strategist reviews each brief, adjusts the angle if needed, and approves it. This takes a fraction of the time it used to.
Tuesday–Wednesday: Draft agents produce first-pass articles from approved briefs. The human editor works through the drafts — this is still the most time-intensive part of the week, and intentionally so. Good editing is not a task you hand to a bot.
Thursday: SEO agent reviews edited drafts. Editor reviews SEO flags and makes final calls. Articles are finalised and handed to the scheduling agent.
Friday: Scheduling agent publishes according to the content calendar. Distribution snippets go out. Analytics agent pulls the previous week's performance data and delivers a report.
Ongoing: The analytics agent monitors live articles and surfaces anything that needs attention — a ranking drop, an unusually high bounce rate, a piece that's gaining traction and could be expanded.
The whole system runs on a combination of n8n for workflow orchestration, a set of large language model APIs for the actual generation tasks, and a human layer that sits at every critical decision point. The humans aren't managing the bots — they're doing the work that requires judgment. The bots handle the work that requires volume and consistency.
What Does Content Automation Actually Solve — and What Doesn't It Fix?
This is the question I get most often from marketing directors at SMBs. Let me answer it directly.
Content automation solves:
- Volume without proportional cost. Producing a consistent stream of well-structured, SEO-ready articles used to require a team. Now it requires a smaller team and a well-designed pipeline. The cost per published piece drops significantly as the system matures.
- Consistency. Bots don't have bad days. The research agent produces a thorough brief whether it's Monday morning or Friday afternoon. The SEO agent checks every article against the same criteria every time. Human consistency at that level is expensive and hard to sustain.
- Speed from brief to draft. The time between "we should write about X" and "here's a first draft" is a fraction of what it used to be. That speed matters when you're trying to respond to a trending topic or a competitor's move.
- Operational visibility. Because every step is logged — brief approved, draft generated, edits made, SEO pass completed, published — you have a clear audit trail. You know exactly where a piece is in the pipeline at any moment.
- Strategy. No agent decides what topics matter for your business, what angle will resonate with your specific audience, or how content fits into the broader marketing plan. That's human work.
- Original insight. Bots synthesise existing information well. They don't generate genuine new perspectives, proprietary research, or the kind of contrarian take that makes a piece genuinely worth reading. That comes from the humans in the loop.
- Brand voice at depth. You can get close with a well-crafted system prompt, but the subtleties of a brand's voice — the specific humour, the cultural references, the way it handles sensitive topics — require ongoing human calibration.
- Relationship-driven content. Interviews, expert roundups, community-driven pieces — anything that requires a human relationship to produce is outside the scope of automation.
Comparison: Manual Content Production vs. AI-Assisted Content Factory
| Dimension | Manual Production | AI-Assisted Factory |
|---|---|---|
| Brief preparation | Hours per article | A fraction of the time, with human review |
| First draft | Hours per article | Generated from brief, human-edited |
| SEO review | Manual checklist, inconsistent | Systematic, every article, same criteria |
| Publishing & distribution | Manual, error-prone | Automated, logged, consistent |
| Performance monitoring | Periodic, often reactive | Continuous, proactive flagging |
| Cost per article | Higher at scale | Reduces significantly as pipeline matures |
| Human involvement | High across all tasks | Concentrated at high-judgment steps |
| Consistency | Variable | High for process; variable for voice (needs calibration) |
How Do We Keep the Quality Bar High Without Slowing Down?
This is the tension at the heart of any content automation system. Speed and volume are easy to optimise for. Quality is harder, and the failure mode is insidious: the output looks fine on the surface but is generic, forgettable, and does nothing for your audience or your search rankings.
Here's how we manage it.
The brief is the quality gate. If the research agent produces a weak brief — vague angle, wrong keyword focus, missing context — the draft will be mediocre no matter how good the draft agent is. The human review of the brief is the most important quality intervention in the whole pipeline. We spend more time here than anywhere else.
The editor has veto power over everything. No agent output goes to the next stage without human approval. This isn't a bottleneck — it's the point. The editor's job is to catch what the bots miss: factual inaccuracies, logical gaps, tone drift, claims that need caveats.
We run a weekly quality retrospective. Every Friday, after the analytics report comes in, the team reviews one published article in detail. Not the metrics — the content itself. Is it genuinely useful? Does it answer the question better than what's already ranking? Would a marketing director at an SMB find it worth their time? This qualitative check catches drift that metrics miss.
We iterate on prompts, not just content. When the draft agent produces something consistently weak in a particular way — say, intros that are too abstract, or H2s that don't answer the question they promise to answer — we update the system prompt. The pipeline improves over time. This is the compounding return of a well-maintained automation system.
Frequently Asked Questions
What tools do AI agents typically use to run a content factory?
Most content automation pipelines use a combination of a workflow orchestration tool (n8n, Make, or Zapier are common choices), large language model APIs for generation and analysis tasks, and a CMS or headless publishing layer for distribution. The specific tooling matters less than the architecture: clear agent roles, defined handoffs, and human checkpoints at every stage where quality or accuracy is at stake.
How much does it cost to set up an AI-driven content automation system?
Costs vary significantly depending on the scale of your operation, the tools you choose, and whether you build in-house or work with an agency. At the low end, a small business running a modest content programme can get started with a relatively modest monthly investment in API costs and workflow tooling. At the higher end, a more sophisticated multi-agent setup with custom integrations and ongoing optimisation represents a more substantial commitment. The better question is cost per published piece at your target volume — that's where automation typically shows a clear advantage over manual production.
Is building in public a good strategy for B2B companies?
For the right type of B2B company — one whose audience is curious about process, methodology, and real-world implementation — yes, it's a strong strategy. It builds credibility in a way that polished marketing copy can't, because it shows rather than tells. The risk is that it requires genuine transparency, including about failures. If your company culture isn't comfortable with that, the content will feel performative and the audience will notice.
Can content automation work for small businesses without a dedicated content team?
It can, but the setup requires upfront investment in prompt engineering, workflow design, and quality calibration. A small business without any content expertise will struggle to configure a pipeline that produces genuinely useful output — the bots amplify the quality of the brief and the editorial judgment you bring to them. The practical path for most SMBs is to either work with a specialist agency to set up and calibrate the system, or to start small with a single automated workflow and expand as you learn.
How do AI agents handle SEO without over-optimising content?
The key is separating the SEO review from the drafting step and giving the SEO agent a narrow, well-defined remit: flag issues, suggest alternatives, but never override the editor's judgment. Over-optimisation — keyword stuffing, unnatural phrasing, heading structures that serve a crawler rather than a reader — almost always happens when SEO logic is baked into the drafting prompt rather than treated as a separate review pass. Keep the two functions distinct and the problem largely solves itself.
Conclusion: The Factory Is Real, and It's Worth Building
I started this diary because I was tired of the AI automation conversation happening entirely in the abstract. Every vendor has a deck showing a beautiful workflow diagram. Almost nobody shows you the actual week-by-week reality of running one.
Here's what I've learned from doing this openly: the technology is genuinely capable, the failures are genuinely instructive, and the humans in the loop are more important than any individual agent. The content factory works because we built it to augment a capable team — not to replace one.
If you're a business owner or marketing director thinking about content automation, the best thing you can do is start small, document everything, and be honest with yourself about where the bots fall short. The compounding returns come from iteration, not from deploying a perfect system on day one.
If you want to talk through what a content automation setup might look like for your business — the architecture, the tooling, the realistic costs and timelines — get in touch. We're happy to walk you through it, no sales pressure, just a practical conversation.