Most companies talk about AI. We decided to use it for our own blog. Not to save money, but to show in practice what a working AI agent looks like.
How the Content Agent Works, Step by Step
The process starts with us. We hand the agent a topic, a short brief, and a few parameters: category, date, UTM tags. Everything else is its job.
The agent runs these steps in order:
- Receives the inputs as a simple JSON object.
- Generates a full HTML article following the blog's style guide.
- Fills in every SEO field: title, og:title, meta description, og:description.
- Updates the blog's index file with the new article.
- Refreshes sitemap.xml with the new URL and date.
- Commits with a descriptive message and pushes to the GitHub repository.
The whole process takes under 90 seconds. From topic to GitHub push.
What It Costs Per Month
Transparency matters. Here are the real costs at 8 articles per month:
That's what it costs to generate one article through the language model's API. At 8 articles per month, that's under 1 lev for content.
The main cost isn't generation. Editorial review takes 15-20 minutes per article. That's where the real time investment goes.
For comparison: a good SEO copywriter for Bulgarian runs between 80 and 200 BGN per article. For 8 articles that's 640 to 1,600 BGN a month. The AI agent produces the same volume for the price of one coffee.
What the Team Checks Before Deploy
The agent doesn't deploy on its own. Every article goes through human review. We check three things:
- Factual accuracy: numbers and examples have to be correct for the BG market.
- Tone and style: the agent is sometimes too formal. We correct it.
- Structure: we verify the CTA is clear and the headlines walk the reader through.
On average we make corrections in about 30% of generated content. The other 70% goes through unchanged.
Why We Do It Publicly
Many agencies sell AI solutions without using them themselves. We think that's a mistake.
When a client asks us whether an AI agent will work for their store, we can point to a live example. Not a pitch deck, not a demo environment. A real product running in production.
That builds trust faster than any brochure.
Three times faster than writing by hand at the same quality bar. Measured across 8 articles in the March–May 2026 period.
There's another reason too. The blog is our proof of concept. If the agent doesn't work well here, we won't sell it to clients. Simple rules.
What the Agent Can't Do
Honesty requires us to name the limits too.
- The agent doesn't know news from the last few days unless we feed it the context.
- It can't interview customers or collect primary data.
- It sometimes generates examples that sound plausible but need to be verified.
- It doesn't replace strategic thinking about brand messaging.
That's why the team stays in the loop. The agent is a tool, not a replacement.
If you run an online store and wonder whether a similar agent would save you work, the answer depends on your content volume. Above 4 articles or product descriptions per month, the investment pays off quickly.