Artificial intelligence in marketing: tasks, limitations and control

What AI in marketing is

AI in marketing usually refers to two different classes of tools. It’s useful to tell them apart: they fail in different ways.

Generative AI creates new material on request: text, an image, headline variants. Text-based services are built on a language model, a program trained on large volumes of text that, given a task, composes the most probable continuation. The task itself is called a prompt. The output of generative AI is a draft: it can sound confident and still contain mistakes.

Predictive and optimizing algorithms don’t write anything. They make lots of small decisions from data: who to show an ad to, what bid to set. An example is automated bid strategies in ad platforms, where the system itself chooses the bids. In Yandex Direct, strategies aimed at average target values — for example, an average cost per conversion — use machine learning to forecast conversion likelihood and set bids in the auctions. Conversion likelihood here is the probability that a person will complete a target action after a click. In VK Ads, under the «Minimum price» strategy, the system aims to get the best results within the chosen budget.

The main point of using neural networks in marketing is to speed up repeatable preparation and processing of data: drafts, variants, summaries, routine setup tasks. They don’t replace decisions. Who to sell to, what to promise, and how much to spend is decided by whoever knows the business and is accountable for the result.

Where AI is used in marketing

The most useful way to look at AI isn’t through a list of services but through tasks. Each one has a part that can be sped up and a part that stays with a person.

Task	What AI can do	What a person checks and decides
Ideas and post drafts	Suggest topics, a structure and a draft text	Facts, brand tone, value for the reader, the decision to publish
Adapting material to formats	Condense an article into a post, make versions for different platforms	Whether the meaning was lost, whether the versions suit the platform
Ad copy variants	Suggest headlines and text based on the page or product description	Match to the offer, whether the claims are permissible, advertising requirements
Bids and audiences in ad platforms	Set bids and targeting under the chosen strategy	Campaign goal, budget, constraints, evaluating the result
Analyzing reviews and inquiries	Group messages by topic, highlight frequent questions	Spot-checking the groups, drawing conclusions about what to change in the product or communications
Data summaries	Describe an export, flag changes, suggest hypotheses	Whether the source data is correct, the real reasons for the changes, decisions
Replies to customers	Draft a reply, answer a routine question in a chatbot	Complex and disputed cases, personal data, promises made to the customer
SEO	Help with drafts and data analysis	Strategy and changes to the site are a separate topic

Let’s go through these areas in more detail.

Content and social media

Content is usually where people start. A neural network helps come up with recurring features, sketch out a post draft, suggest several headlines, and turn one article into a short post and newsletter copy.

It works best with source material: your own article, a transcript of an expert’s talk, answers to customer questions. It works worse when it’s asked to «write a post about the company» with no input. In that case, the model fills the gaps with plausible but made-up details.

A separate article walks through how to prepare a social media post with a neural network and check it, step by step. If there’s a lot of content to produce and you need a steady flow, the breakdown of how to set up a regular output of materials with neural networks will help.

Advertising

Advertising with the help of artificial intelligence mainly means features built into the ad platforms themselves. Automated bid strategies were covered above. Beyond those, the platforms’ help pages describe features like these.

• Yandex Direct. The ad’s headline and text can be generated with a neural network: Direct suggests copy based on the promoted page and business information, and you can edit it. Autotargeting is a technology that shows ads without keyphrases. It analyzes the ad and the landing page and determines whether the ad fits the query, the user’s interests, or the placement’s topic.
• VK Ads. Text generation analyzes the object being promoted, its key details and how it differs from competitors, and suggests text variants. Image generation creates pictures based on a product or service description; the resulting creatives can be edited. With automatic targeting selection, the algorithm analyzes similar campaigns in similar categories and suggests suitable audience settings.

The names and scope of these features change over time, so check the platform’s current help page before launching.

You’re still responsible for the campaign’s goal and what counts as a result, the budget and constraints, and what claims are acceptable in the ad. Generated text and images are reviewed before launch the same way as anything written by hand. Automation optimizes for whatever goal it was given: pick the wrong goal, and the system will diligently chase the wrong thing. There’s a separate article with more detail on how campaigns and billing in Direct are structured.

Analytics and research

Data-driven marketing means decisions based on data, not gut feeling. AI helps here by processing data: grouping hundreds of reviews by topic, finding recurring questions in customer inquiries, summarizing an export, and suggesting hypotheses for why a metric changed.

There’s one limitation, but an important one: the model only sees what you fed it. If the export is incomplete or contains an error, the output will be too. Hypotheses get checked against the source data, and a person makes the decisions.

Communications with customers

Chatbots answer routine questions, and a neural network drafts replies to emails and messages. For this to work, the boundaries need to be set in advance: what the bot doesn’t answer, at what point it hands the conversation to a staff member, and what it isn’t allowed to promise. Customer data is a separate question, covered in the risks section.

SEO

Neural networks are used in site work too, but that has its own tasks and search engine requirements. A separate piece covers how neural networks are used in search engine optimization.

AI marketer and AI agents: what they are and what they don’t replace

AI marketer is a phrase with two meanings. The first is a marketer who uses neural networks in their work: preparing drafts with them, going through data, testing hypotheses. The second is a service or assistant that carries out part of the marketing tasks on request. In both cases, a person sets the task and is accountable for the result.

An AI agent is a system that, given a task, carries out a chain of steps and chooses which tools to use on its own. One automation service’s documentation describes the difference this way: a chain carries out a predefined sequence of calls, while an agent uses a language model to decide what actions to take and tries to pick the right tools for the response. In marketing, an agent might, for example, pull data from several reports, prepare a summary, and draft an email to the team.

Agents usually operate inside an automation setup, where work tools are linked together. A separate article covers how to connect services into an automated workflow.

The more an agent does on its own, the more important it is to define the boundaries in advance:

• which data and accounts it has access to;
• which actions it performs on its own, and which only after confirmation;
• where it’s logged what it did and when.

It’s sensible to leave publishing, sending messages to customers, and changes to ad campaigns for a person to confirm. Strategy, positioning, choosing the audience and the offer, and the budget also stay with a person. What matters most is accountability for everything that goes out in the company’s name.

Limitations and risks

Made-up facts and figures

A hallucination is a neural network’s response that looks plausible but doesn’t match reality. It could be a study that doesn’t exist, a wrong figure, a made-up product feature, or a fabricated review. In marketing, a mistake like this turns into a promise to a customer. So every figure, quote and feature is checked against its source, and anything that can’t be confirmed gets removed.

Generic copy and losing the brand’s voice

Without any input, the model writes «generically» — with the same turns of phrase as thousands of other texts. For materials to sound like your company, you need tone-of-voice guidelines, good examples, and an editor who actually edits the draft, not just checks spelling.

Errors in the input data

A neural network doesn’t know an export is incomplete, that prices are outdated, or that a service’s description changed long ago. It will diligently process whatever it’s given. The output’s quality is never higher than the input’s.

Customers’ personal data

Under the Personal Data Law, personal data is any information relating to a directly or indirectly identified or identifiable individual. Transferring such data also counts as processing it, in the law’s terms. So order exports, correspondence and contact databases shouldn’t be uploaded to external AI services until you’ve checked the service’s terms and your legal basis for such a transfer. Where possible, strip out names, phone numbers and other customer details before uploading. It’s best to agree on the exact procedure with the company’s lawyer.

Rights to generated images and text

The terms of use for generated material are set by each service. Read them before an image makes it into an ad or onto the site. Separately, check that the image doesn’t contain someone else’s logos, recognizable people, or other companies’ products.

Advertising is still advertising

The Advertising Law defines advertising through distribution, the audience, and the purpose — drawing attention to the object being advertised. How the text or image was prepared doesn’t factor into that definition. An ad suggested by a neural network is checked against the same requirements as one written by hand.

How to roll AI into marketing: step by step

We don’t automate chaos. First we build a system. If materials are currently prepared with no clear order or criteria, a neural network will speed up that disorder, not fix it. So the rollout starts with the process.

1. Pick one repeatable task. For example, post drafts from existing material, ad copy variants, or grouping reviews. The task should come up regularly, and the result should be easy to check.
2. Describe the current process and quality criteria. Who prepares the material, who checks it, and what signs mark a piece as done.
3. Prepare the brief and the input. Context, source material, the output format, constraints: what not to write about, which claims are off-limits, what the tone should be.
4. Get a draft from the AI. Not just one, but several variants, if the task allows it.
5. Check and sign off. A person checks the facts, meaning and tone, edits it, and decides whether to publish or launch it.
6. Compare it with the previous process. Time, number of edits, quality of the result.
7. Scale only what has proven itself. Add the next task once the first one is working reliably.

How to tell AI is actually helping

There are no universal benchmarks for the savings here: it all depends on the task, the starting process and the quality of the review. So you assess it with your own data, before and after.

• Time to prepare one piece of material — including review and edits, not just generation time.
• Number of edits and rewrites. If drafts have to be rewritten from scratch, there’s no savings.
• Errors caught during review. If there are many of them and they’re the same type, it’s worth improving the brief and the input.
• Results of posts and campaigns — using the same metrics as before, over a comparable period.

If preparation got faster but the number of errors rose or quality dropped, that’s a reason to go back to steps 2–3, not to expand the automation.

Frequently asked questions

Will AI replace marketers?

No. AI takes on part of the preparation: drafts, variants, summaries. Strategy, decisions about budget and the offer, approving materials and being accountable for them stay with a person. The work itself changes: less time goes into the first draft, more into framing the task and checking the result.

Can you publish text written by a neural network without editing it?

You shouldn’t. A neural network’s output is a draft: it can contain made-up facts, inaccuracies, and phrasing that doesn’t sound like your brand. Only publish what a person has checked and signed off on.

Is it safe to upload customer data to a neural network?

Only after you’ve checked. First find out where the service stores the data and how it uses it, and whether you have a legal basis for that transfer. Where you can, strip out any detail that could identify a person.

Which task should you start rolling AI into marketing with?

With one repeatable task whose result is easy to check: post drafts from existing material, ad copy variants, grouping reviews. The steps are in the «How to roll AI into marketing» section.

Does advertising created with AI need to be labeled?

If the material counts as advertising on the internet, then yes: how it was prepared doesn’t affect the requirements. Under the Advertising Law, advertising on the internet, with the exceptions it lists, must carry an «advertising» label and point to the advertiser, or to a site with information about them. It can be distributed once an ad-data operator has assigned it an identifier. This article of the law has no separate rules for advertising prepared with a neural network. In disputed cases, it’s best to check the labeling procedure with a lawyer.