What Makes a Successful Commerce Agent and Where It Delivers the Most Value
An in-depth look at the architecture of Anthropic's commercial AI agents and our practical experience. We explore what these agents can truly achieve and identify the niches where implementation yields the most significant returns.
Anthropic published an analysis of how commercial agents that make it to production are built: The anatomy of effective commerce agents. It's worth reading the full article. Here, we'll cover what it means for a shop owner who simply wants customers to get an answer at eleven in the evening.
We built Commerce Agents using roughly the same principles. The coincidence is pleasant: it means the path is well-trodden.
One Agent, Not a Swarm of Bots#
The article's first conclusion is surprising for those following the trend of multi-agent schemes. A conversation about a purchase isn't a set of isolated tasks that can be distributed among different bots. A person asks about size, then about delivery, then returns to size and says, "let's make it two." This is all one continuous flow, where each reply builds on the previous one.
Therefore, an effective scheme involves one agent that maintains context, plus pluggable skills for specific tasks. The division is simple: what's needed in almost every conversation resides in the main instructions; rare skills are loaded on demand. Anthropic's threshold is roughly this: if a skill is needed in more than a third of conversations, it belongs in the permanent instructions.
For us, the permanent part includes catalog search, store terms, and business profile. Loaded as needed are product comparison, order assembly, cart management, and parsing attachments sent by the customer.
Tools Are Your Systems, Not Their Copies#
The second principle: the agent should not reinvent store logic. Search, inventory, and order placement already function somewhere — the agent's task is to access them, not to build its own version of truth alongside.
In our case, "your systems" are the catalog compiled from posts and your website, the terms and conditions text, the cart, and the payment provider. The agent doesn't invent delivery times; it takes them from your text. It doesn't invent prices; it takes them from the product card. If data is missing, it honestly states it will clarify and asks for contact information — this is also part of the architecture, not an excuse.
A separate point from the article: product cards, buttons, and forms are also tools, with types and validations. This is why the agent displays the product card rather than retelling it in text: the customer sees the photo, price, and button, and you get a predictable interface instead of model improvisation.
Speed and Money#
Here, the article adopts a sober tone, rarely found in AI texts. Chasing tenths of a second is pointless if the answer is poor: quality, not milliseconds, affects returns and average check. But there's no need to overpay either — context caching with a 90–99 % hit rate helps, as does choosing the right model for the task, verified by measurements, not just by feeling.
We divided the models precisely according to this logic: a fast one converses with the customer, while a more thoughtful one handles reports and catalog edits. The shop owner sees this as two separate fields in “Settings”, and usually doesn't touch them — the default values are chosen based on measurements.
Security Without Paranoia#
Three rules from the article that we literally support:
- Dangerous changes are not applied silently. Catalog edits from the owner's agent go into a queue and await your “Apply” button.
- Invented identifiers are rejected. If a product doesn't exist, the agent won't refer to a non-existent SKU.
- Memory resides in the database, not in the prompt. Customer preferences are written separately and pulled in for the next conversation, rather than bloating each request.
Where the Return on Investment Is Greatest#
This is the main practical question, and the answer isn't "it's the same everywhere." Across our shops, the variation is noticeable.
Services and Appointments. The highest return. Nail technicians, dentistry, device repair, training, rentals — where a person asks five to ten questions before purchasing, almost always outside business hours. Every unanswered question in the evening is an appointment lost to whoever responded faster. The agent closes this gap entirely: it knows prices, durations, what to bring, and when slots are available.
Products with Parameters. Clothing, furniture, electronics, cosmetics — anything where size, composition, compatibility, and lead times are asked. Such a conversation takes ten minutes for a salesperson but seconds for an agent, and it doesn't tire of explaining the same thing for the hundredth time.
Sales Without a Website. Instagram, TikTok, YouTube, and Telegram, where the catalog exists only as posts. Here, the agent provides what was previously impossible: a storefront, a cart, and payment via a link in the profile bio, without needing website development.
Nights, Weekends, and Other Time Zones. Half of all inquiries come in when no one is available to answer. This is the cheapest revenue to recover: it already exists, there's simply no one to service it.
Customers in Different Languages. A shop from Warsaw receives questions in Polish, Ukrainian, English, and Russian. The agent responds in the customer's language without needing an in-house translator.
Where the return is more modest: products with a single characteristic and obvious price, sales through a marketplace where communication happens within the platform, and wholesale deals that are closed through negotiation anyway. Even there, the agent handles routine tasks, but it doesn't create a revolution.
What to Do With This#
- Compile your catalog from what you already have: posts, videos, or your website.
- Describe your terms honestly and in detail — this is the only source from which the agent answers questions about delivery and returns.
- Place a link in your profile bio and review the first twenty dialogues. They reveal what your customers are actually asking, better than any analytics.
How to do this step-by-step is described in the article First Launch. Anthropic's architecture breakdown is here.