The AI Shopping Agents Are Coming (Part 1 in 2 Part Series)

Agentic AI is poised to reshape online advertising strategies significantly by introducing autonomous digital agents that can make decisions and execute tasks independently. These agents can browse, shop, analyze data, and make decisions on behalf of users, acting as virtual assistants or intermediaries. This shift from human consumers to AI agents may require advertisers to adapt their strategies.

First, we need to understand how agentic AI’s capabilities differ from those of Large Language Models (LLMs) and copilots. Agentic AI differs from copilots and LLMs in its ability to act autonomously and perform tasks without direct human intervention. Here’s a breakdown of the distinctions:

• Large Language Models (LLMs): LLMs are the foundational technology for both AI agents and copilots. They are primarily designed to generate text, translate languages, provide information, and brainstorm topics. LLMs lack the ability to independently take action or interact with other software systems on their own.
• Copilots: Copilots can collaborate with users and provide assistance, but they do not make decisions or take actions independently. They are tools that help users complete tasks, but they still require human direction to execute actions. Copilots might generate text, code, or other content, but humans are still needed to take action to make use of these outputs.
• AI Agents: AI agents, on the other hand, are designed to act autonomously on behalf of users. They can complete tasks, interact with other software systems, make decisions, and act independently. AI agents can understand a goal, develop a plan to achieve that goal, and execute actions with minimal human direction. This includes browsing the internet, making purchases, and interacting with other software and workflows. I’ll repeat that second one for emphasis: making purchases.

How AI-Powered Shopping Assistants May Change the Ecommerce Landscape

Now, let’s look at shopping assistants as one example of agentic AI that is clearly on the horizon. AI agents can act as shopping assistants by browsing the internet, comparing prices, finding the best deals, and making purchases for users. For example, Google’s Project Mariner can help research and buy items online, though it currently requires human approval before completing a purchase. Similarly, Perplexity has launched an AI shopping assistant. OpenAI is also beta testing its Operator AI Agent. These agents can streamline the entire purchase process and offer personalized recommendations.

7-Step Process of AI Shopping

Okay, so let’s walk through an example of how an agentic AI might buy a pair of shoes for a person:

1. User Input and Goal Definition: The process begins with a user setting a high-level goal for the AI agent. For example, the user might say something like: “I need a new pair of running shoes, good for road running, and I want them to be comfortable and durable.” The user may also specify a budget, color preferences, or brand preferences.
2. Agent Planning and Research: The AI agent understands the goal and its defined role as a shopping assistant. It then devises a plan to achieve the goal. This includes identifying appropriate websites and gathering information about running shoes. It starts by exploring data sets and destinations to find information, acting like a search agent. The agent might look at shoe reviews, brand websites, and articles about the best running shoes. The AI agent evaluates different types of data and compares approaches based on the user’s goal, acting like a goal-based agent. It considers factors like the user’s running style, the type of terrain they run on (road), and their stated preferences for comfort and durability.
3. Data Analysis and Comparison: The agent analyzes the data it has collected, parsing information like pricing, product reviews, and technical specifications. It prioritizes practical information over emotional appeals since it is an AI agent interacting with the product information. The agent then evaluates actions based on potential options and outcomes, like a utility-based agent, determining which shoes best match the user’s preferences and requirements. It may also check other sources for discounts or promotions in real time. If the agent has multimodal capabilities, it might analyze visuals like shoe images to assess the aesthetics and features, helping to refine its choices.
4. Decision Making and Selection: Based on its analysis, the agent narrows down the options to a few suitable pairs of shoes. It might prioritize shoes with good reviews for comfort and durability. It makes a decision based on the data and goal. The AI agent may also leverage past results and feedback from previous searches or user preferences to refine its choices, acting like a learning agent. The agent may understand the user’s data privacy preferences and would adhere to those.
5. Purchase and Transaction: The agent proceeds to initiate the transaction. It may interact with ecommerce platforms and payment systems. The agent may fill out forms, even negotiate prices if possible, and complete transactions, acting as an intermediary. However, the AI agent will not make a purchase without getting explicit user approval. The user could review the agent’s choice and make any final changes or approvals or ask the agent to refine the search and selection. The agent might use its ability to interact with different software tools to check real-time inventory, and to complete the purchase in the most efficient way.
6. Adaptation and Learning: Throughout the process, the agent is adaptable, handling any trial-and-error. If a particular website is blocked or if there is a problem, the agent can navigate around it and modify its strategy. The agent learns based on various inputs, feedback, and past results. Over time, it may become better at predicting the user’s preferences based on their prior purchases.
7. Post-Purchase: After the purchase is complete, the agent may inform the user the transaction is complete, track the shipping status, and even later follow up with a review request.

This example illustrates how an agentic AI can perform a complex task, going beyond simple conversation to making decisions and taking action on a user’s behalf. The capabilities include planning, data analysis, decision-making, transaction execution, and adaptation based on feedback and new information.

So, that leads me to think of the following questions: 

• What kind of ads might draw more agentic AI bot interest than human interest? 
• Will agentic agents actually click on ads? 
• Will they purposefully seek out sponsored ads, knowing those are the businesses truly engaged in finding new customers for their products? 

All of this is really an intellectual exercise at this point, but it could very well be something internet marketers need to pay attention to in the future. Stay tuned for part 2 all about the three types of ads that may attract AI shopping agents.