Author: the editorial team airassvet.ru. This article was prepared based on a transcript of a podcast by a hands-on context advertising specialist (provided to the editorial team) and cross-checked against Yandex documentation on September 27, 2026. The editorial team did not test the tools described here itself, is not commercially affiliated with the services mentioned, and there are no affiliate links in the article.
Short answer. An AI model combined with the Direct API is already collecting keywords, masks, and search suggestions, grouping negative keywords, creating drafts of Unified Performance Campaigns (UPC), and writing ads, proposals, and reports. It still cannot fully replace a Yandex Direct specialist for two reasons: the Direct API does not cover all campaign types and settings, and someone still has to catch the AI’s mistakes — either a human or a separate verification agent. What changes the job is not a “magic prompt,” but a strong knowledge base, a change log, and checks at every step.
Below is a breakdown of a real demo from the podcast: what the AI handled on its own, where it made mistakes, which API limitations are confirmed by the documentation, and the best order for implementing this kind of automation.
Contents
- Will AI replace a Yandex Direct specialist?
- What AI can already do in Direct
- Where the Yandex Direct API hits limits
- What mistakes the AI made in the demo
- Why you need a knowledge base, not just a prompt
- Local or cloud-based
- How to introduce AI into Direct workflows
- How the Direct specialist role is changing
- Frequently asked questions
Will AI replace a Yandex Direct specialist?
The podcast host owns a context advertising agency and built his own service for working with the Direct and Metrics APIs. According to him, some colleagues are already losing projects, and clients are not going to another contractor — they are going “to AI”: an in-house marketing team decides to set up advertising itself with the help of AI.
His answer to the question “Will it replace the specialist?” is philosophical, but he backs it with a practical point. A large company that outsourced advertising to a contractor has someone to hold accountable for mistakes. If the ad team cut costs, implemented AI, and something broke, responsibility falls on the people who made the decision. So demand for specialists remains, but what clients pay them for is changing.
For a Direct specialist, AI is a routine-work executor connected to the API and knowledge base — not a replacement for the person who sets the tasks, checks the results, and is responsible for the client’s budget. This point is an editorial conclusion from the demo, not a market study.
What AI can already do in Direct
In the podcast, the author works in the Cursor editor: the AI calls tools in his service through the API and decides on its own which tool is needed. The demo project is the website of an official fitness equipment dealer; the only manual work is typing tasks into the chat.
| Task | What it looked like in the demo | Where a human is needed |
|---|---|---|
| Keywords | Parsing keywords from masks, Wordstat frequency by group | Check group relevance |
| Search suggestions | Running masks through suggestion collection with a tree depth | Filter out long-tail and irrelevant phrases |
| Negative keywords | Collecting from all keywords and suggestions, then grouping | Remove words that block qualified demand |
| UPC for search | Two campaign drafts: ads, sitelinks, images, UTM tags | Check group structure, region, and bidding strategy |
| Proposals and reports | Built from call transcripts, a case study database, and competitor analysis in the search results | Align conclusions with the client |
Keywords, masks, and search suggestions
The author asks the AI to collect keywords for the masks “brand + dealer” and “gym equipment,” and to run the masks separately through search suggestion collection — similar to Key Collector. The AI then automatically groups the result and shows Wordstat frequency, even though that was not requested: this behavior is built into the service settings.
Negative keywords by group
Next, the AI collects negative keywords from all keywords and suggestions and groups them: “official site,” “reviews,” “overview,” “instructions,” “specifications,” “assembly,” and “for home.” The author immediately notes that such lists need to be checked manually. For example, he removed the word “site” from the negatives and asked the AI to remember that rule.
Campaign build
Next, the AI creates two UPCs for search — “brand + dealer” and “gym equipment.” Before creation, it confirms the settings: time targeting (weekdays selected), image format (1:1, generate new ones), and aut-targeting categories (only qualified queries kept). The campaigns are saved as drafts; a human launches them. The author asks for one ad per group with one headline: in his view, when a combinatorial ad has too many headlines, it becomes hard to analyze the result, so he moves new hypotheses into new ads.
Proposals, reports, and the change log
The service’s proposals are assembled from the transcript of the first call with the client, follow-up meetings, competitor website analysis in the search results, and the agency’s case study database. For reporting, the author keeps his own project change log: which keywords, negative keywords, and images were added, by whom, and when. He explains this by saying that the change history in Direct itself is inconvenient for building a work report.
Landing pages and SEO content
A separate tool from the author builds a landing page from several competitor pages: it screenshots the above-the-fold section, extracts the text, inserts the client’s data, and generates missing images. Screenshots are needed because text alone does not show what the page is really about. For SEO, a chain of agents is used: content plan, Wordstat frequency, text, tables, images, mobile layout checks, and submission for recrawl through the Yandex Webmaster API. This method exists in Webmaster and has a daily quota. According to the author, on one of the agency’s sites traffic after launching such a chain grew from roughly 500–800 to 2,500 unique visitors per month; he notes that he may be off in the numbers, and analytics data were not shown in the podcast.
Where the Yandex Direct API hits limits
The main bottleneck, according to the author, is the Direct API itself: in his experience, the Metrica API always worked well, while Direct’s did not, and that has not changed. Some of his claims are confirmed by documentation, while others could not be verified from public sources.
Confirmed by Yandex documentation:
- UPCs can be created through the API: support was added November 27, 2023, and they are worked with using the
v501endpoint and the UnifiedCampaignstructure. - Combinatorial ads appeared in the Ads API service March 20, 2026.
- The campaigns.get method accepts four campaign types: TEXT_CAMPAIGN, MOBILE_APP_CAMPAIGN, CPM_BANNER_CAMPAIGN, and UNIFIED_CAMPAIGN. There is no separate type for Campaign Wizard in this list, which aligns with the author’s point that Campaign Wizard cannot be used through the API.
- A keyword can contain no more than 7 words, and a group can contain no more than 200 phrases. The AI ran into that 7-word limit in the demo.
According to the author, this was not checked by the editorial team against the documentation: Through the API, you can’t create AI-generated ads or interest-based targeting, product campaigns require manual cleanup, and balances aren’t pulled from every account—for example, with a shared account or when connected through a manager account. The 100-group-per-campaign limit the AI mentioned in the demo also isn’t referenced in the documentation we found for keyword phrases.
Practical takeaway: before promising a client that “everything can be done with AI,” check the list of campaign types and settings you need against the current API documentation. Anything the API doesn’t support will still need to be handled manually in the interface or in Direct Commander.
What errors the AI made in the demo
The most useful part of the podcast is that the author doesn’t hide the mistakes:
- All keywords in one group. The AI put about 40 keywords into one group. The reason: in a previous project, the author asked it to combine groups, the AI saved that in its log, and then carried the rule over to the new project. It was fixed with the instruction: “group two or three relevant keywords per group and create separate groups with their own ads.”
- Phrases longer than 7 words. Search suggestions were added to the campaign as keywords, and Direct rejected them. The AI filtered out long phrases and duplicates, and left the first attempt as a draft that had to be deleted.
- Region and settings from old campaigns. The AI pulled the region “Russia” from the client’s previous campaigns, where some settings were incorrect. The author noted that they would still have to be fixed manually.
- Extra negative keywords. The word “site” in the negative keywords could have cut off intended demand; the rule had to be specified explicitly.
- Test limits. The final campaigns still had limits the author had once set for testing, and he removed them only after the recording.
The author’s answer to these kinds of mistakes is not to give up on AI, but to use a checking agent. On air, he directly tasks an agent to compare the created campaign against a settings template: campaign type, channel, time-based targeting, auto-targeting, and the number of keywords and ads per group. Using the same logic, a separate agent checks creatives for moderation approval and risky wording.
Why you need a knowledge base, not just a prompt
The author insists: an AI doesn’t “learn” in the everyday sense—it works with the information you give it. That’s why the central object is a repository: a description of what can and can’t be done through the Direct API, campaign templates, the specialist’s preferences, and a log of past decisions. For his training intensive, his team created two guides: one for people and one for the AI, because the model processes the material differently.
That’s where the debate over the idea that “you program it once and then you don’t need AI anymore” comes in. Programming a service consumes a lot of tokens once, and after that the AI calls ready-made tools and processes the finished result—so far fewer tokens are used. But, according to the author, no one will give up AI as an interface: it’s more convenient to assign tasks in chat.
Another point concerns the logic of development. The author doesn’t like the word “vibe coding”: even with AI, you still need stages, task lists, a log, and testing after every change. You can review things manually, have someone else do it, or use a separate agent.
Locally or in the cloud
The author handles some tasks on his own computer: transcription, image generation, and video generation. His setup is 64 GB of RAM and a graphics card with roughly 20–22 GB of VRAM; for high-quality video, he cites the RTX 5080–5090 range. The cloud removes the need to buy hardware, but you pay for the amount of resources used.
For tasks that need to run 24/7—campaign checks, balances, and change monitoring—the author uses rented servers (VDS). A typical server doesn’t have VRAM, so he leaves generation to the computer and monitoring to the server. In Cursor itself, he switches models based on task complexity: for simple operations he uses a faster, cheaper one, and for complex tasks a more thorough one.
How to introduce AI into Direct work
Below is an editorial workflow assembled from the podcast’s practices. It has not been separately tested and does not guarantee results.
- Start with a change log. Record every action by both the person and the AI in the project. Without a log, there will be no client report and no way to analyze mistakes.
- Build a knowledge base. Campaign templates, negative keyword rules, ad requirements, and API limitations—all in a format that is convenient for the AI.
- Limit the scope. Work with only one active project at a time and create campaigns only as drafts. The author separately introduced the rule of an “active project” so the AI wouldn’t create a campaign in the wrong account.
- Add a checking agent. Compare against the settings template after every creation or bulk change.
- Review creatives separately. Texts and images created by AI should be checked against Direct moderation requirements and advertising law before launch. This does not replace a legal review of specific wording.
- Leave the launch to a person. A draft becomes an active campaign only after a specialist reviews it.
What changes in the Direct specialist profession
According to the author, the main shift is from quality to quantity. While the AI assembles two campaigns, a specialist can simultaneously handle several more projects in other chats. He sees the same trend in SEO: what wins is not perfect optimization of one page, but many pages in a well-tuned workflow.
His forecast for agencies is blunt: the share of junior specialists will shrink, and team leads will remain. In the author’s estimate, many agencies spend about 50% of revenue on maintaining the team; that is his estimate, not research data. In his experience, account management costs have matched campaign creation costs—he gives an approximate benchmark of about 50,000 ₽. At the same time, clients expect more: fast custom reports, meeting transcripts, and testing new offers. The campaign-creation “buttons” themselves are getting cheaper, and value is shifting toward hypotheses, analytics, and oversight.
He sees a separate risk in token costs: models are getting more complex, and requests are getting more expensive, so it’s more efficient to keep your own service on your own computer or server instead of solving the task from scratch every time with a cloud AI.
Frequently asked questions
Can you create an EPC in Yandex Direct through the API?
Yes. EPC support was added to the API in November 2023; campaigns are created with the add method of the Campaigns service at v501 using the UnifiedCampaign structure.
Can you work with Campaign Wizard through the API?
Campaign Wizard is not included in the list of campaign types returned by the campaigns.get method. The podcast author also says that such a campaign cannot be created through the API.
Why does Direct reject keyword phrases assembled by AI?
A common reason is length. A keyword phrase in Direct can contain no more than 7 words, and search suggestions are often longer.
Do you need to check negative keywords assembled by AI?
Yes. In the demo, the AI added the word “site” to the negative keywords, which could have cut off target queries. The author checked the list manually and zakreplyal the exclusions in the rules.
How much computer power do local AI models need?
It depends on the task. For transcription and images, the author uses 64 GB of RAM and a graphics card with 20–22 GB of VRAM; for high-quality video, they mention GPUs in the RTX 5080–5090 range.
How the airassvet.ru team helps implement an AI agent in advertising workflows
If you want to replicate this setup for your marketing department or agency, the team airassvet.ru can:
- audit one advertising workflow — for example, keyword research or a monthly report — and document the data sources and constraints;
- build a company knowledge base (RAG) from your SOPs, campaign templates, and negative keyword rules so the agent relies on them instead of the model's general knowledge;
- integrate the AI agent with your systems and set up testing and manual approval before changes go live.
The first step is to choose one process, document how it works now, what data it uses, what limitations it has, and what metric will define a successful pilot. No numerical impact is promised in advance. If the process is selected, discuss the project.
Conclusion
The neural network in Yandex Direct already handles the main routine work of a direct marketing specialist: gathering keywords and suggestions, negative phrases, drafts of EPK, ads, offers, and reports. But in the demo, it grouped keywords differently than needed, carried over someone else's settings, and ran into the 7-word limit, while the Direct API does not cover all campaign types. So the working setup looks like this: a knowledge base, a change log, a checking agent, and a person who launches the campaign. The direct specialist does not disappear — they become the person who builds and controls this system.