Nadezhda Shiroglazova, an AI/ML engineer who has built AI solutions for international consulting firms, talks about how AI agents are changing marketing and sales: they help businesses find customers and help shoppers pick products and buy them.
In mid-October, Meta’s AI agent Muse reached 2.5 million daily active users. People use it for everyday tasks such as finding cheap flights, sorting out subscription problems, and shopping online. For now the agent can ask the user for permission before it buys something, but experts predict that these systems will gradually become more independent.
Businesses, meanwhile, already use AI widely to automate marketing and sales. Agents look for prospects, talk to them, gauge their interest, and help structure the conversation. But what happens when there is AI on both sides of the deal? How will a seller’s agent deal with a buyer’s agent, and how will marketing change if companies have to win over algorithms as well as people?
We talked to Nadezhda Shiroglazova, an AI/ML engineer who builds and deploys AI in business processes. She has worked on technology projects for the consulting firms BCG and Bain, and has built AI solutions for companies as an independent engineer. Her projects include a voice agent for outbound sales that matched the conversion rate of human operators at roughly 60% lower cost per contact, and a quality control system that handles more than 10,000 customer calls a day.
In the interview, she describes how AI agents persuade people to buy, how they find business partners, and how they could eventually become personal shopping assistants.
– AI agents can already find customers, call them, and help companies sell. Now there are tools on the buyer’s side that help people choose products. You work with AI in sales, so what do you think happens when a seller’s agent meets a buyer’s agent? And how close are we to that?
I think we’re already heading that way, but we need to separate AI helping you choose a product from a fully autonomous purchase. The second one takes more than better models. It also needs secure payments, a way to confirm that the agent is authorized to act, and protection for the buyer. So in the near term I expect gradual progress, not a situation where agents buy things in bulk with no human involved.
Trust and security are a separate issue. A personal agent could buy things on its own in certain categories, within a set budget, and ask for confirmation otherwise. It also shouldn’t have to hand the seller your whole purchase history or your personal data. It only needs to share what’s required to find the right product.
For sellers, these intermediaries mean rethinking marketing itself. You used to be able to win a person over with an emotional ad, a brand image, or a good slogan. An AI agent will evaluate an offer on different grounds: how closely the product fits the request, its specifications and price, whether it’s in stock, and how good the delivery and return terms are. Emotional arguments alone are unlikely to settle it.
– Does that mean emotional marketing will stop mattering as AI agents become common?
No, I don’t think so. Companies will just have to address two audiences: the AI agent, which compares offers against specific criteria, and the human, who is still swayed by emotions, values, and the image of a brand.
An agent can compare ten companies in a moment and check prices, specifications, and purchase terms. But even then, a person may want to know more about the brand, see whether its values are close to their own, and only after that decide. And nobody has canceled impulse buying, which is often driven by emotion. So I wouldn’t say emotional marketing is disappearing. Its role is changing. One striking image may not be enough if the offer doesn’t hold up against the competition.
We used a similar principle when we developed the AI agent for outbound sales. It could take a customer through the main stages of a call: finding out what they need, clarifying details, presenting the product, giving the price, answering objections, and confirming the agreement. It remembered the customer’s earlier answers and built the conversation step by step instead of simply reading items off a script.
We didn’t replace human operators completely, though. The agent worked together with the team. If the customer was willing to continue with a person and a free operator was available, the system passed the call over along with the context of the conversation. If the operators were busy, the agent kept going on its own.
– Your voice agent cost about 60% less than an operator, and the share of customers who took the target action stayed the same. How did you structure the dialogue so that the agent could lead the conversation toward a sale?
The result came from steady work on the dialogue architecture. I split the conversation into stages and analyzed where the agent lost the thread or misunderstood what the customer wanted. I added examples of how operators pick up on the other person’s words, and I let the agent put answers in its own words instead of repeating prepared phrases.
Working on empathy made an especially noticeable difference. Conversations got longer, but more of them ended in a purchase. Sometimes a person goes off topic, tells you about their situation, or explains why they’re unsure about buying. If you mechanically drag them back to the price and the offer at that moment, you can lose their trust. The agent has to recognize what lies behind the customer’s reply, respond to it, and only then ease back toward the goal of the call.
– Fewer than 1% of customers recognized that they were talking to an AI. What made the conversation so convincing?
Several things add up to a natural conversation. For each language I selected a voice and worked on pauses and sound, but a voice alone doesn’t solve it. If the agent answers fast and smoothly but ignores what the person said, or jumps back to the script right after a personal story, it sounds artificial immediately. So I added examples of using context, more flexible wording, and the option to stay for a moment on whatever matters to the customer.
Most of the time, people ended the call for ordinary reasons: they changed their mind, lost interest in the offer, or realized the product didn’t suit them. Talking to an agent wasn’t what stood in the way.
– Not every salesperson can handle objections without putting the customer off. How do your AI agents bring people back to the purchase without turning the conversation into a hard sell?
I studied real calls and built those techniques into the agent’s logic. I taught it to take the context of the conversation into account, to tell different kinds of objections apart, and to word its answers flexibly. If a customer says “that’s too expensive,” the agent doesn’t pull out a ready-made argument. It relies on what the customer said earlier and tries to find the reason for the doubt. Maybe the person doesn’t see the value of the product, or is comparing offers, or isn’t sure yet that they need it. The conversation should go differently depending on which it is.
But I’d frame the question another way: how do people handle objections themselves? Why does one operator hear “too expensive” and wrap up the call, while another understands what the doubt is really about and helps the person decide? It isn’t always down to a good script. An experienced specialist often uses techniques that are hard to formalize. They sense the context, notice small details in the answers, and know when to keep talking and when not to push. Those are the situations that are hardest to teach an AI until we understand ourselves what exactly it needs to learn.
I see that as my job: taking apart real conversations, finding the solutions that so far exist only between the lines, and turning them into examples and principles for the agent. What limits us is our own ability to describe human experience, not the capabilities of AI models.
– In one of your projects, AI helped find influencers and contact potential partners. About 6% of cold messages led to a collaboration, which is a good result for cold outreach to people who had never heard of the company. How was this communication organized?
Imagine a brand team looking for partners on Instagram and TikTok. They need to keep finding new creators in their niche, check who suits the product, write to each of them, and remember who replied, who was promised a product, and who is due another message. It all becomes a big flow of profiles, messages, and small agreements. We had this problem too. Working manually at that volume, we lost potential partners simply because it was hard to keep track of every thread.
I built a platform where the marketer sets the search direction and an AI agent helps move each contact through the first stages of the funnel. It helps find and select creators, prepares personalized first messages, tracks replies, reminds the team about the next action, and suggests a draft response based on the history of the conversation. Each creator’s profile shows their status, tasks, and agreements, and you can draft a contract there too.
– The more communication a company hands to AI agents, the more it matters to know how well they perform. You developed a system that analyzed more than 10,000 calls a day. Can that approach be used to monitor the AI agents themselves?
Yes, I think quality control for AI agents is the next step for these systems. In our project, we automated the evaluation of operators’ work. The system assessed the quality of communication on ten or more metrics, and that raised monitoring coverage from 20% to 100%.
The next stage is to apply the same approach to AI agents: analyze their conversations with customers, spot errors, and check how consistently they lead a conversation and reach their goal. This will matter most once agents start interacting with each other, for example when a seller’s agent talks to a buyer’s agent.
Then the issue is not only how effective the communication is but also security. We have to make sure that a seller’s agent doesn’t ask for unnecessary personal data, breach confidentiality, or try to get around the limits the user has set. That requires dedicated control mechanisms: supervisory agents that watch for violations, assess risks, and cut off a conversation when needed.
– Nadezhda, let’s look at AI agents from the point of view of an ordinary consumer rather than a developer or researcher. Which tasks would you be ready to trust them with yourself?
I like this angle. Today, when I search for products with AI, I often have to explain again what suits me. In the future, a personal agent will be able to factor in my previous purchases, my preferences, and even things I’ve already turned down.
I’d trust it to buy things on its own in certain categories and within a set budget. If a purchase goes beyond that, the agent should ask for my confirmation first. Control over personal data matters just as much. I’m willing to give the agent my purchase history and my preferences so it can choose better. But that doesn’t mean sellers should be able to see my work correspondence, my income, my banking details, or my personal contacts. The agent should pass on only the information that’s actually needed for the particular deal.
