Why automate content at all
According to FirstPageSage, the average SEO ROI is 748%. But to get that ROI, you have to publish regularly. One or two articles a month won't do anything.
Let's do the actual content marketing math for 2026:
- Freelance article, 1,000-1,500 words: $80-$250 per article (Stacc, 2026)
- Agency article: $625-$1,250 per article (Stacc, 2026)
- Cover from a designer: $30-$100 on Fiverr
- Time for one article with a manual rewrite: 6-8 hours
Plus the hidden cost — managing freelancers takes 10-15 hours a week (Feedbird, 2026). Briefing, reading drafts, sending edits, publishing. For a small agency, that's a non-starter.
Either write everything yourself and burn a week on 4 articles. Or pay $300-$1,000/month for 4 freelance articles and put up with quality swings. Neither worked for us.
Build in-house automation. AI does the grunt work, I just approve the result with a Telegram button. One pass from idea to published article — 5 minutes of my time and $0.30 in API costs.
What we're building: 7 n8n workflows
The idea is simple. A Telegram bot takes decisions from me, n8n runs the chain of generation, publishing, and notifications.
What the system does:
- Every morning it scans 6 RSS sources (MarkTechPost, VentureBeat AI, Google AI Blog, n8n Blog, DEV n8n Community, Hugging Face Blog)
- Scores topics via Gemini 3.1-flash-lite for relevance to business automation
- Sends the top-10 to Telegram with "🗂 Select" buttons
- After I pick a topic — generates 2 cover variants via OpenAI Images in HAIQ's brand style
- Writes a bilingual article (UA + EN) via GPT-5 with facts only from the source
- Sends a preview with 5 buttons: publish (3 categories), edits, rewrite, discard
- Publishes to the site via the HAIQ REST API with a full SEO package and cover
- Catches errors and pings Telegram with an explanation of what broke and where
n8n 2.31.5 self-hosted on Railway ($5-10/month), OpenAI API (GPT-5 + Images), Gemini API, Telegram Bot, Google Sheets, your site with a REST API. Build time — 2-3 weeks for the full version.
Architecture: how 7 workflows pass an article along the pipeline
The pipeline is a conveyor belt. Each workflow handles one stage and calls the next. State lives in Google Sheets: candidates, drafts, published, errors.
WF1 — Discovery
Cron trigger at 09:00 daily. Reads 6 RSS feeds. Deduplicates against past candidates in Sheets. Passes candidates to Gemini for scoring on a 0-10 scale for relevance to automation and business. Sends the top-10 with the highest scores to Telegram with "🗂 Select" and "⏭ Skip batch" buttons.
WF2 — Selection
Router for all callbacks from Telegram. 9 Switch branches: select topic (select), skip batch (skip), pick cover (cover), regenerate covers (regen_cover), publish as Case/AI News/Guide (publish with kind_code), edit (edit), rewrite article (regen_article), discard (discard), handle text messages for edits (text_message).
WF3 — Cover Generation
Downloads the source article through the r.jina.ai proxy (bypasses Cloudflare), extracts the main text via GPT-4o-mini, analyzes the content. Generates 2 image prompts in HAIQ's brand style: first isometric 3D, second flat vector illustration. OpenAI Images creates 2 covers at 1536×1024 JPEG. Arrives in Telegram with pick buttons.
WF4 — Article Generation
Receives the chosen cover and source_text. Calls GPT-5 with a bilingual prompt. The model returns JSON with UA and EN versions at the same time: same structure, same length 1,200-1,500 words, same facts, different language. Plus SEO slugs, meta-title, meta-desc, short-desc, tags. The article gets rendered to Telegraph for preview. I get a message with the first paragraph and 5 buttons.
WF5 — Publish
When I pick a category (Case, AI News, or Guide), the cover uploads to the site via /api/upload, the article gets published via /api/articles with the full payload (langs.ua + langs.en, tags, meta), sitemap updates, confirmation lands in Telegram with a link to the live article.
WF6 — Regen and Edit
Two-mode workflow. In regen mode: GPT-5 gets the same source and the previous version of the article, generates a substantially different one — different H2s, different tone, different angle. In edit mode: I write to Telegram in plain text what edits to make ("add an H2 about security", "shorten the intro"), and GPT-5 applies exactly those edits, no more. In both modes — it edits the existing Telegraph page instead of creating a new one.
WF7 — Error Handler
A separate workflow with an Error Trigger node. Wired as errorWorkflow to WF1-WF6. Classifies errors into 8 categories: Telegram API, OpenAI, HAIQ API, Google Sheets, Telegraph, n8n Task Runner, Cloudflare, custom. Logs to Sheets and sends to Telegram in plain-English text with "what happened" and "what to do".
Everything lives in Google Sheets: 5 tabs (sources, candidates, drafts, published, errors). Every workflow reads and writes through the Sheets API. No separate database. No admin panel to stand up. The Sheet is storage, UI, and analytics all in one.
What AI does and what a human does
The key decision — split the process into 3 jobs AI does well, and 1 job a human does.
| Step | AI | Human | Time |
|---|---|---|---|
| RSS scoring | Gemini picks top-10 from 30-40 | — | Automatic |
| Topic pick | — | Click from 10 options | ~30 sec |
| Cover generation | 2 variants from OpenAI Images | — | ~45 sec |
| Cover pick | — | Click from 2 variants | ~20 sec |
| Article generation | GPT-5 UA + EN in parallel | — | ~90 sec |
| Preview reading | — | Read in Telegraph | ~2-3 min |
| Decision | — | Click: publish / edit / rewrite | ~5 sec |
| Site publish | Upload + Insert + Sitemap | — | ~15 sec |
All decisions go through Telegram inline buttons. I never type. I only click.
Timeline of a single publication
Here's what a single publication looks like from my side, with real timings.
09:00. Morning Telegram message lands: "10 topics for today". Each topic — headline, source, Gemini score, "🗂 Select" button.
09:15. 30 seconds of reading headlines, I pick one. "📝 Preparing article…" lands.
09:16. After 45-60 seconds, 2 covers arrive with "✅ Pick this" buttons under each. One isometric 3D, one flat vector. I pick the one that fits the topic better.
09:17. "📝 Preparing article for the chosen cover…" arrives. GPT-5 generates the bilingual article in 60-90 seconds.
09:18. Preview lands: headline, TL;DR, first paragraph, tags, UA/EN slugs, Telegraph link, and 5 buttons. "🗂 Publish Case", "📰 Publish AI News", "📖 Publish Guide", "✏️ Edits", "🔄 Rewrite".
09:19. I read the preview in Telegraph in 2-3 minutes. All good — I tap one of the "Publish" buttons. Need edits — I tap "✏️ Edits" and type in chat: "shorten the first paragraph" or "add an H2 about security". GPT-5 rewrites the article with exactly those edits.
09:20. "✅ Published on HAIQ.agency" arrives with a link to the live article. Google is already indexing it.
5 minutes of my time from the first click to a published article. From there — it lives on the site, works for SEO, brings traffic. Manually the same work would take 6-8 hours.
How much one cycle costs
Breaking down the cost by API calls. Prices as of August 2026.
| Stage | API call | Model | Cost |
|---|---|---|---|
| WF1 Discovery | 1× Gemini scoring | gemini-3.1-flash-lite | ~$0.007 |
| WF3 Extract | 1× GPT-4o-mini | gpt-4o-mini | ~$0.002 |
| WF3 Image prompts | 1× GPT-4o-mini | gpt-4o-mini | ~$0.001 |
| WF3 Cover ×2 | 2× OpenAI Images | gpt-image-1-mini medium | ~$0.043 |
| WF4 Article UA+EN | 1× GPT-5 (bilingual) | gpt-5 | ~$0.25 |
| WF5 Publish | HAIQ API + sitemap | — | $0 |
| Total | ~5 API calls | — | ~$0.30 |
If I need to rewrite the article via WF6 — that's another +$0.25. If I regenerate covers — +$0.04. Max cost per cycle with all regens — $0.60.
Plus infrastructure: Railway VPS for n8n ~$10/month, Google Sheets free, Telegraph free, HAIQ backend we already have.
ROI: when it makes sense to build
Comparing manual and automated processes on concrete numbers for 8 articles a month.
| Metric | Manual rewrite | HAIQ pipeline | Difference |
|---|---|---|---|
| Cost per article | $110 ($80 copywriter + $30 cover) | $0.30 | −99.7% |
| Human time per article | 6-8 hours | 5 minutes | −98.9% |
| Bilingual UA+EN | +$40-$60 for translation | Included | −100% |
| SEO meta, slugs, tags | +30 min manual work | Automatic | — |
| CMS publishing | +15 min | Automatic | — |
| Month (8 articles) | $880 + 60 hours | $2.40 + 40 min | −99.7% |
One-time investment for the basic build — from $500. Payback for a business with 8 articles/month — less than a month. Yearly savings — ~$10,500 plus 60+ hours of my work freed up.
In the first 3 weeks after launching the pipeline, we published 15 articles. Manually that would take 100+ hours. Automation took ~2 hours of my time and $4.50 in API costs. Source: HAIQ Agency internal analytics, August 2026.
When it makes sense and when it doesn't
| Publishing volume | Verdict | Why |
|---|---|---|
| 1-3 articles/month | Write by hand | Build investment won't pay off |
| 4-8 articles/month | Automation justified | Payback 1-2 months |
| 10+ articles/month | Automation mandatory | Manual process burns budget |
Key takeaways from the project
Bilingual is practically free for AI. GPT-5 generates UA and EN in the same API call. With freelancers it costs double. An underrated advantage for brands going after the English-speaking market.
Approval via Telegram saves hours on day-to-day management. No opening CMS, no reading drafts in email, no chasing edits. Everything — phone buttons. Publishing one article fits inside a coffee break.
State in Google Sheets — simple and reliable. No separate database. The Sheet is storage, UI, and analytics in one place. Scales to tens of thousands of rows without issues.
Errors are normal — a system for catching them matters. Every pipeline error lands in Telegram with a category ("🔴 OpenAI: out of credits" or "🟡 Sheets: matching column empty") and an "Open execution" button. Saves hours on debugging.
Regen and Edit matter more than initial generation. GPT-5 doesn't always nail it on the first try. The ability to rewrite in 90 seconds or make text-based edits ("shorten the intro") — that's what makes automation actually usable day-to-day.
Frequently asked questions
Will this approach work for a small company with 1-2 people?
Yes, this is exactly for small and mid-sized business. We're a team of 3 at HAIQ Agency. The pipeline was built by me alone. Minimum API cost per month is $2-$5 with 6-10 articles. For a business with a $50/month automation budget, it pays off from month one.
How much does it cost to build such a pipeline turnkey?
The basic version with 3-4 workflows (no regen/edit, no covers) — from $500. The full version like ours (7 workflows, bilingual, covers, text-based editing, error handler) — $1,500-$3,000 depending on your CMS integration complexity. Delivery time — 1-2 weeks.
Won't readers notice it's AI content?
GPT-5 in 2026 writes naturally. The pipeline layers on HAIQ's style rules, bans typical AI phrases ("in today's world", "revolutionary"), uses rhetorical questions and active voice. I personally read every article before publishing. If something reads off — I tap "🔄 Rewrite" and get a new version in 90 seconds. Out of 15 published articles, we haven't gotten a single comment suspecting AI.
Can I connect my own CMS instead of your HAIQ backend?
Yes. The pipeline builds against any CMS with a REST API or GraphQL: WordPress (via the wp/v2 API), Ghost, Sanity, Strapi, Contentful, Webflow, Shopify. Only WF5 (Publish) changes — the other workflows stay the same. Building the adapted WF5 takes 3-5 hours of work.
What if I don't want to pay for GPT-5 and want something cheaper?
You can swap in GPT-4o (~$0.10 per article instead of $0.25), Claude 3.5 Sonnet (~$0.15), or Gemini 2.0 Pro (~$0.08). Quality is slightly lower, but for news rewrites the difference isn't critical. The minimal config with gpt-4o-mini + Gemini covers — ~$0.08 per article.
How long can one pipeline run without maintenance?
Our pipeline has been running unattended for 3+ weeks. Once a week I check the errors tab in Sheets — mostly transient failures (429 rate limits, connection aborted) that the system retries through. Real breakages that need manual work — once every 2-3 weeks. Main causes: OpenAI API changes, n8n updates, or an RSS source moving to a new domain.
Ready to build this for your business?
Publishing 4+ articles a month? Know the pain of "waiting weeks for a freelancer"? Time to think about automation.
We at HAIQ Agency build these pipelines for clients. Project scoping is free. We audit your current content process and show you exactly where automation will save the most.
Delivery — from $500 for the basic version with 3-4 workflows. Timeline — 1-2 weeks. Payment — 50% at kickoff, 50% at acceptance.
Based on HAIQ Agency's own experience and analytics (August 2026). External data: FirstPageSage, Stacc, Feedbird.
