I let AI do my shopping. Here’s what happened.

12 min readFeb 12, 2025

This piece originally appeared in Curious Commerce, a weekly newsletter on how technology is shaping how and why we buy. Subscribe for free!

Melina Flabiano is an ecommerce leader who loves the intersection of technology and shopping. By day, she leads product marketing at Estee Lauder. By night, she writes Curious Commerce, a weekly newsletter breaking down the future of shopping. Subscribe for free here.

AI shopping assistants aren’t just recommending what to buy anymore-they’re actually buying things for us. I tested Perplexity Pro and OpenAI Operator to see if AI is ready to take over my shopping… and let’s just say, things got weird.

You know how in sci-fi movies, there’s always a computer who gains sentience and starts acting on its own? That’s basically agentic AI.

Many of us have interacted with primarily non-agentic AI tools, for use cases such as:

Asking ChatGPT questions, essentially treating it as a more sophisticated search engine
Interacting with AI-enabled assistants like Siri or Alexa
Observing the output of AI-enabled experiences, like product recommendations online, targeted advertising, or tailored recommendations on Netflix or Spotify — a lot of advanced personalization that we enjoy as consumers is powered by AI
Office tools like Copilot summarize meetings and help draft emails

All these use cases are static AI — they produce a predefined range of outputs every time an input is provided, and do not adapt, make decisions, or take actions. They make recommendations, and the models powering the experiences improve over time. But they can’t take your credit card and go on a shopping spree at Bloomingdale’s. At least not yet (cue ominous music).

AI agents, on the other hand, have the ability to act on our behalf to fulfill tasks. For this week, I am focusing on the emerging capability of AI agents to perform tasks, like purchasing products, on our behalf. These tools are being called “AI Shopping Assistants.” But like a lot of technology applications, a lot of the money to be made is behind-the-scenes, in unsexy parts of the business like supply chain and inventory optimization. There are also exciting developments in creative production (developing and testing assets) and customer service. Those use cases and more will be the subject of future newsletters — at this rate, the AI series is shaping up to be at least 20 parts. (Here’s Part I and Part II if you’d like to catch up).

Let’s dive in!

Because buying things online and having them delivered to my door in 2–3 days is still too taxing, I was very excited to test out the shopping capabilities of agentic AI. In the (very near) future, this capability could enable consumers to automate tasks like booking travel and making restaurant reservations. It could allow us to outsource weekly grocery shopping and buying wedding gifts from a friends’ registry entirely to an AI assistant.

As of today, Perplexity Pro is one of the only tools available to offer a functional AI Shopping Assistant who can complete transactions. I had to try it.

First, I asked Perplexity for help buying shampoo. I specified I wanted OGX. It surfaced a few options from CVS, Ulta, and other retailers. The “Buy with Pro” button triggers the agent-enabled purchase flow.

Then it prompts me for some details to complete the transaction.

I was surprised this step wasn’t more seamless. I had just provided my credit card information to sign up for the free trial of Perplexity Pro. So far, it would have been faster to buy this on Amazon, but I’m trusting vision!

There is also a section within cart to specify size and color for SKUs that have multiple color/size combinations. This hints at some limitations with Perplexity’s functionality.

As Maxwell Zeff at TechCrunch points out when he tested this in December, it appears that Perplexity is scraping the web for available product details, vs. listing items for sale themselves in their own marketplace. In other words, I just gave $10 to Perplexity, and Perplexity is going to turn around and buy the shampoo from CVS with that $10. This raises a lot of challenges, for instance — if CVS happened to turn on a promotion between when I paid Perplexity and when they executed the transaction, would I get a refund? Conversely, how much above what I paid is Perplexity authorized to spend if the price changes in the other direction?

Ultimately, I received an email from CVS confirming my purchase 4 hours after I completed the transaction on Perplexity. I immediately noticed an issue — I was charged a $5 shipping fee by CVS that I was not made aware of when I bought via Perplexity. In fact, both screenshots above highlight “Free Delivery” as a benefit of purchasing via CVS. This essentially increased my purchase total by 50% — from $10 to $15 — and shipping cost is a key criteria in purchasing a low-cost item like shampoo.

Reflections from my first AI purchase

Perplexity’s agentic shopping capabilities are in the early stages, and I really respect the team’s hustle to bring this to market and be a first mover, even though it’s pretty bare bones. Perplexity is clearly making shopping a differentiator, rolling out features over the holiday geared to recommend and purchase gifts.

Right now, Perplexity Pro acts as a bridge between shoppers and retailers rather than operating its own storefront. That means it’s scraping product data and executing transactions on my behalf-but the process isn’t always seamless.

OpenAI Operator: Ordering a new toy for my son

ChatGPT introduced AI Operator, their agentic assistant, 3 weeks ago.

I will not attempt a detailed technical explanation, but it’s interesting to note that Operator is essentially trained to act like a human and click through to complete actions on a website like we might do, vs. using API integrations. This enables more use cases to ramp up faster, as Operator can essentially do anything a human can do online, if it has access to the right context.

I had to test it.

Under shopping, there are prompts to help get started. I want to buy my baby a new toy.

I used this prompt: I have a baby who is 8 months old. I’m looking for a developmentally-appropriate toy that is educational, costing between $10–15 and available with free shipping. Can you share some options? I’d also like you to buy it for me.

In our Chat, Operator opened up their own browser and started searching for toys…crazyyyy.

It selected a toy that fit my criteria and opened a window in Target.com for me to buy it. Operator didn’t actually buy the toy for me-it surfaced the best option and then walked me through the purchase flow. Unlike Perplexity, it doesn’t complete transactions autonomously (yet).

It took Google 4 years after launching search to launch Google Shopping (in 2002), which then ramped up fairly slowly to the several billion monthly searches it commands today. Meanwhile, these AI companies have waited mere months to move into shopping. I think it’s likely that by this summer, many of us will be using AI shopping assistants to complete tasks.

Practical examples of how this could look in our day-to-day life

You ask your preferred AI Shopping Assistant to reorder all of your consumables, like pet food and toiletries using either voice or typed commands. Because you are sharing your order history with the AI, it knows your preferences and can reorder the same items at the best price.
Your AI assistant has access to your Google calendar and knows you have a wedding coming up in 6 weeks. It prompts you to order a wedding gift and asks if you want to order something from the wedding registry link provided via email. You can share a desired price range and mail a wedding gift without ever opening the registry yourself.
You notice your socks are worn out so you ask your AI to order another pack similar to the last ones you ordered a few months ago. Because you’re sharing order history, it knows your preferences and can surface the relevant item and some better-priced, similar alternatives.
You rented an Airbnb with friends for the weekend. In the group chat, you send some links for recipes you want to cook together. You send these links to your AI assistant, who assembles an Instacart order with all the ingredients you need and schedules delivery for a few hours after your check-in.

Who will own the consumer-AI assistant relationship?

If I am interfacing with an AI shopping assistant for many shopping tasks, I won’t be visiting Amazon, Google, or brand sites as much as before. I’m sure this scares legacy retailers and marketplaces, who want to maintain their position in the value chain.

This is prompting a scramble to develop agentic AI shopping assistants.

An AI shopping assistant from OpenAI, Perplexity, or another AI-first company

These companies are at the cutting edge of building AI capabilities and thus are well-positioned to build consumer-facing products like AI shopping assistants. OpenAI Operator and Perplexity’s recent launches highlight their advances in the space.

I believe a challenge will be how best to add commerce to the equation. As I observed when testing Perplexity’s shopping capabilities, products aren’t yet listed for sale on a true marketplace. With Operator, the AI is searching external sites for me, but I am still owning the transaction (though that may change).

These platforms also lack some of the context on who I am as a customer — what are my interests, order preferences, etc.? Theoretically, I could grant them access to my purchase history in some way — bank statements? emails? — so they could best tailor recommendations to me. Like any personal assistant, there will be an onboarding period.

A Google AI shopping assistant An Amazon AI shopping assistant

Google is well-positioned here because so much of my day-to-day life is managed in Google tools — Gmail, Calendar, and Search. Google has a rich portrait of my purchase history (via order emails), my schedule, and interests. There is a lot of context for an AI shopping assistant to tap into to make recommendations for products or to anticipate needs. Being truly useful is critical for these applications to go from “fun to test” to truly sticky behaviors that change the way we shop long-term. I wrote recently about Google Shopping’s AI focused makeover. This is setting the stage for agentic AI shopping assistant capabilities.

An Amazon AI shopping assistant

Amazon is investing heavily in AI across their business. In the shopping assistant space, they have rolled out Rufus, a chatbot that helps with shopping queries inside Amazon’s app. It’s not an autonomous shopping assistant (yet), but it could evolve into one.

Amazon has a data advantage. I buy the majority of my everyday purchases — from baby supplies to kitchen stuff — on Amazon. They own my order history data and know a lot about my preferences. If I were Amazon, I wouldn’t want to share that data for free with an unaffiliated AI agent who might use it to recommend products from other retailers — in fact, Amazon was one of the earliest companies to block OpenAI from crawling their site. Any outside AI shopping assistant may be hampered by a lack of access to a lot of my purchase history.

I think Amazon will build their own assistant to work within Amazon — for instance, allowing me to buy groceries via Whole Foods and schedule delivery using conversational text and voice prompts. I’m sure Walmart will work to offer a competing tool. It’s an oft-quoted stat in ecommerce that 50% of product searches start on Amazon. Using an Amazon AI assistant ensures those dollars stay on Amazon.

Will brands push back on AI agents “shopping” on their sites?

When an AI agent completes a transaction on a customer’s behalf, brands lose the ability to “sell” in the traditional sense, meaning they can’t tempt an AI to grow their basket in the same way they can prompt a human to buy a pair of shoes to match the jacket in their cart. These bots could represent low-value, unprofitable customers to some brands.

Some retailers already restrict AI-powered bots from scraping their sites (Amazon being a notable example). If AI shopping assistants gain traction, we could see more brands push back or require stricter API-based integrations.

Putting on my brand hat, I can imagine a nightmare scenario where AI assistants place lots of orders on my site, which affects my inventory levels and operations, only to cancel or modify them when they don’t meet the consumer’s requirements. Especially in the early days, this could create a lot of noise and turn me off to hosting this type of traffic.

On the other hand, restricting AI agent traffic entirely can be shortsighted, since Brands want to drive sales, and these agents may make up an increasing share of traffic in the coming years. Perhaps AI agent traffic becomes like any other affiliate channel — taking a small slice of revenue for recommending your brand in return for the sale.

The death (re-birth?) of SEO

The age of keyword stuffing and SEO hacking is over. As more people start their searches outside of traditional search engines, organic traffic won’t just be determined by rank on Google. It will also be determined by how relevant AI shopping assistants find your products and content.

Just like how brands need to maintain product feed data to surface on Google Shopping and other platforms, they will need to format and maintain product data in a way that AI shopping assistants can parse it. This can drive advantages in discoverability by these tools, increasingly inclusion in AI assistant recommendations, and hence, conversions.

It’s still unknown exactly how brands should “optimize” their sites and content for these tools, as they are so nascent. But brands that move early to make themselves discoverable on these platforms will be big winners.

How do you sell to an AI? With another AI, of course!

So consumers will all have AI shopping assistants. What about brands? Brands will have AI ambassadors.

For example, on a fashion site, the brands’s AI might introduce itself in machine terms: “Here are the latest arrivals and current promotions,” and your AI might respond, “My user is looking for a size M black dress, ideally A-line, under $100, and prefers cotton. What do you have?” The two AIs can rapidly exchange info — availability, prices, detailed specs — far faster than you could clicking filters.

This sounds very futuristic, but you can imagine the implications for ecommerce if a large share of traffic starts coming from AI assistants. The impetus to lay out websites in the traditional sense — to be easy for humans to understand — would become less important vs. other objectives.

There’s so much more to come here, but for now I’ll work on using Operator and Perplexity to automate my life so I have more time to play around with AI tools and write this newsletter.