The Cost of AI-Generated Content for SaaS Marketing (Real Numbers)

We publish content for four SaaS products plus the studio site. Most of what you read on our blogs is drafted by an LLM and edited by a human (us). The version of this post you are reading right now was produced exactly that way. Every founder writing about AI content marketing cost dances around the actual numbers. We are going to put them on the page, including the embarrassing parts. Here is what we spend, what we earn, and what we still pay humans to do.

What this post covers

Three things:

1. Real per-article costs: LLM spend, human time, and what counts as overhead.
2. The conversion outcomes we have measured in 12 months of running this hybrid content engine.
3. The threshold where AI-assisted content beats both pure-human and pure-AI alternatives.

We are not selling a content service. We have no incentive to inflate or deflate any of these numbers.

The unit economics of an AI-assisted article

A typical 1,800 to 2,200 word article on one of our blogs has the following cost shape. We are using rounded numbers that are within 10% of the real ones.

LLM costs:

• A draft generation with a strong model (we use Claude through OpenRouter for most posts) burns roughly 80,000 to 120,000 input tokens (we provide context: brand voice, product references, prior articles, and the spec) and emits 4,000 to 6,000 output tokens.
• At our blended rate that works out to roughly $0.40 to $0.80 per article for the draft.
• Light revision passes (asking the model to tighten a section or rewrite a weak intro) add maybe $0.10 to $0.20.
• Total LLM spend: $0.50 to $1.00 per article.

Human time:

• 25 to 45 minutes of editing per article. We read every line, change roughly 20 to 30% of the text, kill paragraphs that feel filler, fix factual errors (the model occasionally says something subtly wrong), and add specific numbers we know but the model does not.
• At our blended rate (we count our own time at 70 EUR per hour for cost analysis), that is 30 to 50 EUR per article in time cost.

Image generation:

• A hero image generated with our preferred image model costs around $0.04 per image, plus 10 minutes of selection and minor adjustment time.

Total cost per article:

• LLM: ~$0.80
• Human edit: ~40 EUR
• Image: ~10 EUR (mostly time)
• Hosting and publishing: rounding to zero (we publish on our own infrastructure)

Round number: 50 EUR all-in per article, with about 1% of that being the AI cost and 99% being human time.

The AI cost is a rounding error. The human cost is the entire budget. That changes how you think about the work.

Comparing to pure-human and pure-AI

Let us put this against the alternatives.

Pure-human article, written by us from scratch with no AI assist:

• 3 to 5 hours per article at the same rate, so 210 to 350 EUR per article.
• Better in some dimensions (we write what we know cold), but slower and more painful.
• We do this for the most important pillar pieces or for posts where we want a specific personal narrative the LLM cannot fake.

Pure-AI article, generated by the LLM with no human edit:

• $0.50 to $1.00 in LLM cost. Zero human time.
• Lower quality. Generic phrases ("in today's fast-paced world"), missing specifics, hallucinated details.
• We tested this honestly for a month. The articles ranked worse, converted worse, and one of them said something factually wrong about Carriva that we caught only because a friend emailed us.

AI-assisted hybrid (what we do):

• 50 EUR per article.
• Quality close to pure-human on most posts, with the personal narrative replaced by the LLM's draft.
• Throughput around 4 articles per day if we focused on it (we do not, because we ship products too).

The hybrid is the obvious win on cost-per-quality. The pure-AI option saves money but does not save reputation.

What we still pay a human to do

This is where the post gets honest. The LLM does not do these things well, and we still do them by hand.

Source-of-truth fact checking

The model can confidently state numbers that are wrong. It is fine on general facts. It is unreliable on specific, recent, niche facts. Every article that mentions a number gets the number checked. When we write about Carriva and the 18 months we spent on it, every concrete number (534 prospects, 20 contacted, 6 bugs in 36 minutes) is something we verified, not something we trusted the model to know.

Voice and personal narrative

The model can imitate "founder writing" but it cannot tell the story of the night Jean-Luc called us about the cohort 1965 bug. We add those stories by hand. They are what makes the post feel real and not sloppy.

Internal links and product context

We link articles together by hand. The model proposes links, we accept the good ones and reject the off-topic ones. The same goes for product mentions. The model knows we have a product called Carriva but it does not always know which post is the right one to link to in this paragraph. That is human work.

Killing filler

LLMs love filler. Long transitions, "in conclusion" paragraphs, three-bullet summaries that say nothing. The single biggest improvement on every draft is cutting the bloat. We are ruthless about this. Better 1,800 honest words than 2,200 padded ones. That cut is human work.

The conversion data

Here is the embarrassing-honest part: we did not have great content analytics before we migrated to Umami self-hosted. We discussed that migration in our PostHog vs Umami self-hosted breakdown. So the numbers we trust are post-migration.

For the studio site (draftedby.com) over the last 90 days:

• Articles published: 14
• Cost: 14 × 50 EUR = 700 EUR
• Organic clicks per day: trending from 25/day in week one to roughly 95/day at the time of writing, with two articles driving 60% of the traffic.
• Sign-ups attributable to blog content (last-touch in Umami): 38 over the period.
• Conversion to paid: too early to measure cleanly, but 4 of those signups have already converted on Carriva and 2 on a lesson-planning product.

If we attribute even a conservative LTV per converted customer (say, 200 EUR over the next 12 months on average), the 6 conversions covered the content cost roughly 1.7 times over in 90 days. The trajectory suggests it gets better as the catalog of articles grows.

The honest caveat: blog content compounds slowly. Most of the traffic gains come from articles that have been live for 4 to 8 months. The articles we published last week are not driving most of today's traffic. If you want fast results, content is the wrong channel. If you want a moat, content is one of the few channels with real durability.

Where AI content fails

Three failure modes we have hit:

1. Niche, expert subjects. The model knows the general shape of "running RAG in regulated industries" but it does not know the specific failure modes we hit. For our piece on RAG in regulated industries, the human edit was closer to 90 minutes because so much of the content was domain-specific.
2. Topical news. The model is months out of date. Anything that requires current awareness of an industry release or a regulation change is human-first.
3. Listicles people want to copy. "Top 10 X" articles tend to be generic and the AI version reads like every other AI version. We avoid these. They do not differentiate the studio.

The math that decides whether to use AI for content

For us, the decision tree is roughly:

• If the article is about a topic we know cold and want to be the canonical voice on, we draft pure-human.
• If the article is informational with a clear structure and we have specific data to insert, we go AI-assisted hybrid.
• If the article is a recap of the news, we do not write it. We are not a news outlet.

The dollar question is "does the marginal article pay for itself?" Our threshold for that has been "if a 50 EUR article generates one signup over 12 months, it pays for itself many times over." That bar is low. We hit it.

A note on why we use Claude Code, not ChatGPT

Worth a paragraph because it shapes the workflow. We draft articles using Claude Code with a persistent skill that knows our brand voice, our product context, and the spec for the article we are writing. The 1M-context window matters because we feed it our prior articles, our brand guide, and the topic notes in one shot. We tried ChatGPT and we do not use it for code or for content. We covered the broader reasoning in our piece on why we use Claude Code for everything.

The cost per article is not meaningfully different across providers. The workflow quality is.

What we would change

Two things we have not optimized:

1. The image generation pipeline. Right now hero images are a manual step. We could automate the prompt-to-publish flow but we have not. The 10-minute manual step is worth catching the occasional bad image.
2. The post-publish promotion. We write the article and we do not do enough to push it. If you have organic traffic, content is great. If you do not yet, content alone is slow. We accept that for now and we are explicitly not running paid campaigns to amplify content because we have not crossed the 100 daily organic clicks threshold per product yet.

What the numbers mean for someone starting fresh

If you are weighing AI content marketing cost for a new SaaS:

1. Do not go pure-AI. The reputational risk on factual errors is a real expense not on the spreadsheet.
2. Do go hybrid. The cost-per-quality is genuinely the best we have found.
3. Budget human time at your real rate, not zero. Your time is not free.
4. Measure conversions, not just traffic. Traffic without conversion is vanity.
5. Be patient. Content compounds in months, not weeks.

What is next

We are looking to publish at a steady cadence of 2 to 4 articles per week across the studio for the rest of 2026. The catalog will compound. We are also experimenting with a small internal evaluation set: we run new article drafts through a quality-checker pass that catches filler, generic phrasing, and missing specifics before a human even reviews. Early days but promising.

AI content marketing cost is real and small per article. The human cost is real and dominant. The combination is the cheapest high-quality content we have ever produced. The catch is the discipline to not phone it in. The minute you publish a 2,200-word article that says nothing, the math stops working because nobody clicks. The math stays working when the article is honest, specific, and worth the reader's time. That is the only durable strategy.