← All articles

Mikhail Polyakov

Automating content creation with AI: six steps of a pipeline and the numbers from ten articles

Level: BeginnerReading time: 15 min

Topics: Automation, AI, Small business

A small conveyor on a wooden desk in the evening: blank sheets of paper pass under a turquoise scanning beam, most go on to a neat stack and one is diverted into a side tray; next to it a desk lamp, a pile of folders and house plants

Short answer: you can have a model write the draft, gather material, make the image and translate, but the decision to publish is better left to a review and a person. On my article pipeline for this site, every one of the ten articles published from October 1 to 5, 2026 was rejected at the first review that produced an answer: from 2 to 8 issues per article. Google writes that it judges the quality of a text regardless of how it was created, and among the violations it names mass creation of pages with no value for readers in order to rank in search. Yandex, in its help pages, names automatically generated content that is useless for the user among the reasons for restrictions. From March 1, 2027, Russia's AI law gives those who create audio and visual materials with large AI models the possibility to label them; the text of the relevant article of the law contains no obligation to add a label. A bill that proposes to oblige the developers of such models to label materials is under consideration as of October 5, 2026. Below: what content automation is, what search engines and the law say about it, how my two pipelines work and what their logs show, six steps for your own pipeline, and a calculation of when it pays off.

This article was made by the same pipeline it describes: the draft was written by the Claude model from Anthropic, it was reviewed by the GPT model from OpenAI, and I set the topic and the rules.

What content creation automation is

Content creation automation is a chain of steps where a program does part of the work on a text, image or video: it gathers material, writes a draft, checks it, draws a cover, translates, and posts it to a website or a channel. A language model (a neural network that writes text from instructions) can write the draft in such a chain, but it does not have to be responsible for everything.

Examples for a small company: a weekly newsletter to customers on industry news, blog articles based on questions that come in to support, product descriptions for a marketplace, a summary of a webinar recording for social media. The pattern is the same everywhere: material goes in, a text that customers will see comes out, and in between there is room for mistakes.

What Google, Yandex and the law say

Google. In February 2023, the Google Search team wrote that its systems aim to reward high-quality content "however it is produced". In the same post: using automation, including AI, to generate texts with the primary purpose of manipulating search rankings is a violation of its spam policies. In the spam policies, updated on August 28, 2026, this is called scaled content abuse: many pages generated to rank in search rather than to help people. The first example on the list is using generative AI to produce many pages without adding value for users.

Google's guidance on helpful content (updated on October 1, 2026) suggests asking three questions about your text: who created it, how, and why. If a text is mostly generated by automation, Google suggests asking yourself whether it is clear to readers that automation was used, and whether you explain how and why. According to the same guidance, an AI disclosure is useful where readers would reasonably expect one. Fabricating an author, for example using an AI-generated photo and a made-up expert name, is described in the same guidance as deception.

Yandex. In the Yandex Webmaster help pages, Yandex names, among the sites to which its algorithms may apply restrictions, sites that copy or rewrite information from other resources and sites "with automatically generated content that is useless for the user" (my translation). In my view, the key word here is "useless".

Russian law. Russian Federal Law No. 243-FZ of July 26, 2026, on supporting the development of artificial intelligence technologies in the Russian Federation (its official title is in Russian), entered into force on September 1, 2026, except for certain articles. Its Article 9 deals with labelling materials created using large foundation AI models and enters into force on March 1, 2027. Under the law's definition, a large foundation model has at least 1 billion parameters, that is, adjustable numbers inside the model, serves as a basis for other software and handles a large number of different tasks. Under the text of the article, a person who uses such a model to create material in audio and (or) visual form "is provided with the possibility" of placing a notice that AI was used. The format and procedure of the label are set by an agreement with whoever provides the model. Large platforms where users run their own pages (the law lists criteria, including more than 500,000 users from Russia per day) must give users the possibility to label information created with such models. The text of Article 9 contains no obligation to add a label for the person who creates the material. On August 17, 2026, bill No. 1317978-8 amending this article with respect to labelling was registered in the State Duma, the lower house of the Russian parliament. According to ComNews, the bill proposes to oblige companies that develop large foundation models to label materials created with these models, and to have the government set the form of the label. Until the bill is passed, there is no such obligation. As of October 5, 2026, according to its page on the State Duma website, it has not been passed and is listed as under consideration: on that day the responsible committee proposed accepting it for consideration. If you publish images, audio or video made by AI, check the status of these rules closer to March 2027, and if in doubt, ask a lawyer.

The takeaway for business from these rules, in my view, is this: the common feature for Google and Yandex is that a text is useless for the reader. Google separately names as a violation many pages created to rank in search without value for people, and Yandex separately mentions copying and rewriting other people's material. AI as such, as a way of writing, is not named as a violation in the rules cited, and Google states directly that appropriate use of AI or automation is not against its guidelines.

How my article pipeline works

Since October 1, 2026, the articles in this section go through the following path.

  1. I choose a topic or ask for an article in a chat.
  2. Claude searches the internet for primary sources itself, opens them and writes every statement of the article into a separate file next to a verbatim quote and a link. I call this file the claims register.
  3. Claude writes a draft, and a program runs mechanical checks: banned words, long sentences, overlap with already published articles, em dashes.
  4. The draft, the brief and the claims register go to OpenAI's GPT model with instructions for an independent reviewer: check each fact against the register, look for violations of the law, dangerous advice, and terms the reader will not understand. The answer follows a strict format: accept or reject, plus a list of blockers, that is, issues that must be fixed before publication. The reviewer does not open the source pages themselves; it checks the text against the quotes written out in the register. If a quote was written out incorrectly, it will not notice: this is the weak spot of my pipeline, and the second of the six steps below deals with it.
  5. After a rejection, Claude fixes every blocker and sends a new version. The limit is three reviews; beyond that only by my decision.
  6. GPT draws the cover from a description that forbids lettering and logos. Under the pipeline's rules, Claude opens the finished image and, if it sees letters, a logo or a face, edits the description and redraws it. Claude viewed the cover of this article on October 5, 2026, and no redraw was needed. The accepted text is translated into English and goes through the same review.
  7. The publishing program takes exactly the version the reviewer accepted, checks the file's fingerprint and, after posting, compares the live page with the accepted text.

What the logs of the first ten articles show (October 1-5, 2026, each with a Russian and an English version):

  • The first verdict received for the Russian version was "reject" for all ten. The first verdict contained from 2 to 8 blockers, 51 in total.
  • Accepting the Russian version took from 2 to 6 verdicts, with a median of 4. In seven articles out of ten, more than three verdicts were needed, and I allowed reviews beyond the limit.
  • In total there were 67 review runs across the two languages; in three of them no answer in the required format came back, and such a run does not count. Reviews with a recorded time took from 0.4 to 12.3 minutes, with a median of about one minute. The 12.3 minutes belong to a run that never produced an answer.
  • In the receipts where usage was recorded, one review took from 26 to 41 thousand input tokens (a token is a piece of a word; the amount of work a model does is measured in them).

What errors the review caught. Examples from real verdicts:

  • In the fintech article, the conditions for mandatory acceptance of the digital ruble were retold incorrectly, and different parts of an article of Russia's Code of Administrative Offences on fines for personal data violations were mixed up.
  • In the medicine article, a median from a study was called an "average", and a conclusion about all medical AI was drawn from an incomplete list kept by the US regulator.
  • In the article on website security, a piece of advice to an accountant on checking a suspicious request could, if followed literally, have ended in a real payment.
  • In the transport article, a trial reported in a February news item was presented as ongoing.
  • In several articles, advice to adopt something came without a calculation of whether it would pay off, and technical words were left unexplained.

Errors like these show up when you check against the source. A business owner who relies on a wrong fine amount or an outdated condition set by law may make a wrong decision.

Second example: a news digest

The AI assistant Hermes prepares a selection of news about AI, robots and history for me and sends it to Telegram. Until September 23, 2026, the model itself searched for news for the selection. In my working note, the reason for the change is recorded like this: the previous agent kept breaking and made up news. Since September 23, the news has been gathered by an ordinary script from 12 RSS feeds (a machine-readable feed of a site's updates) from publications such as Ars Technica, The Verge and MIT Technology Review, and the model only selects and retells what was gathered. A manual run on September 23 produced 9 news items, all links from the feeds. According to the task log, the selection was prepared every day from September 24 to October 5.

The conclusion I drew from both pipelines: where material can be gathered by a program, let the program bring it, and leave selection and retelling to the model. Then every link in the finished text can be checked against the list of what was gathered. Where the sources are not known in advance, as with articles, the model searches for them itself, and then you need written-out quotes and spot checks by a person.

Six steps to build your own pipeline

  1. Where possible, a program gathers the material. News feeds, an export of support questions, product cards from your inventory system, a webinar transcript. The model receives ready material and writes from it. If the sources are not known in advance and the model searches for them itself, as with my articles, go straight to the second step and do not skip the spot check.
    • Who does it: a contractor or employee who sets up the export.
    • How to check: every link and every product name in the finished text is present in the gathered material. A program can check this. For sources found by the model, such a list does not work; the second step checks them.
  2. Every statement is recorded next to its source. Numbers, dates, prices, conditions, quotes: in a separate table, each with a line from the primary source.
    • Who does it: the author model while writing, a person selectively.
    • How to check: take three random statements, open the source via the link and find the quote there. If even one of them cannot be found there, send the text back. Statements where an error could cost the business dearly (prices, deadlines, fine amounts, legal requirements) should all be checked, even if the spot check passed. The reviewing model checks the text against the written-out quotes and will not notice a quote that was written out incorrectly, so this step should be done by a person or by a program that opens the page itself.
  3. Mechanical checks before the reviewing model. A list of banned words (promises of results, competitors' names), sentence length, overlap with what has already been published, formatting rules.
    • Who does it: a program, set up once by a contractor.
    • How to check: on a text with a banned word inserted in advance, the check finds it.
  4. Another model reviews, with the right to reject. Instructions for the reviewer: check facts against the source table, look for promises of results, dangerous advice and unclear words; answer in the format "accept or reject, list of issues". A limit on rounds, so that the argument between author and reviewer does not go round in circles forever.
    • Who does it: a contractor sets it up, the owner approves the instructions.
    • How to check: give the reviewer a text with one made-up number. If it accepts such a text, the instructions need more work. I explained why it is not worth relying on a model's own report of its work in the article When an AI agent bends the rules.
  5. Exactly the accepted version is published. The publishing program calculates the file's fingerprint (a checksum that changes with any edit) and posts only the file whose fingerprint matches the accepted one. Any edit after acceptance sends the text back for review.
    • Who does it: the publishing program.
    • How to check: on a separate test copy of the site that is not accessible from the internet, change one letter in the accepted text and try to publish: publishing should be refused. Do not test this on the live site, including hidden sections: if the protection is broken, an unreviewed text will be publicly accessible.
  6. A person chooses the topic and is responsible for what is published. The model does not decide what to write about or what to promise customers. Keep a log in which, for each text, you can see who accepted it and how many reviews there were.
    • Who does it: the owner or the person responsible for content.
    • How to check: for any published article, you can find who approved it and the receipts of all reviews.

About images: if a model draws your covers, explicitly forbid lettering, logos and faces in the description, and look at the result at full size before publishing: are there any letters, numbers or marks resembling logos on it? They are easy to miss in a small preview.

Will it pay off: a calculation with hypothetical numbers

An example with hypothetical numbers; substitute your own. A company publishes 4 articles a month. Writing by hand, an employee spends 6 hours per article. With the pipeline, a person spends 2 hours per article: choosing the topic, reading the finished text, spot-checking sources. An hour of the employee's time costs 1,500 roubles. Subscriptions to two models are, hypothetically, 4,000 roubles a month. Suppose a contractor spends 40 hours building the pipeline at the same rate, that is, 60,000 roubles.

This saves 16 hours a month, which is 24,000 roubles, or 20,000 roubles after subscriptions. The build pays off in about 3 months. This does not include time to maintain the pipeline: editing instructions and dealing with failures, like the three review runs without an answer in my log.

The calculation depends heavily on how much time a person spends per article with the pipeline. At 3 hours, payback is about 4.3 months; at 4 hours, about 7.5 months; at 5 hours, about 30 months. So before building the pipeline, time how long one text takes now, and after launch, time the review. I explained how to measure the benefit of AI in hours saved in the article Developers and analysts working with AI.

When automation is not a good fit

  • The text rests on personal experience, opinion or a relationship with a customer: a column by the head of the company, a reply to a complaint. A model can help with a draft, but the substance should come from a person.
  • An error is costly and there is nothing to check it with: legal, medical or financial advice without sources and without a specialist who will read the text.
  • The goal is to produce lots of pages for search. Google's spam policies directly name generating many pages without value for readers, and Yandex lists sites with automatically generated useless content among those to which restrictions may be applied.
  • There is no material for the pipeline. If the program has nothing to gather, the model writes from what it has learned, and every statement will have to be checked against sources.

Summary

Content creation automation works as a chain: a program gathers material, a model writes, mechanical checks and another model look for errors, and a person decides what gets published. On my pipeline, every one of the ten articles was rejected at the first review that produced an answer, and accepting them took from 2 to 6 verdicts. In the rules of Google and Yandex, the common feature of a violation is that a text is useless for the reader, not that AI wrote it. From March 1, 2027, Russia's AI law provides the possibility to label audio and visual materials made by large AI models, without an obligation to add a label, and a bill that proposes to oblige the developers of such models to label materials is under consideration as of October 5, 2026. Start with one type of text, time it before and after, and count how many errors the review caught.

I work on AI agents and automation. If you would like to look at my projects or discuss your own task, visit my portfolio.

Sources