How Much Does an AI Shopping Agent Cost Per Conversation?

Ask three vendors what an AI shopping agent costs and you will hear three different units: per conversation, per resolution, per month. None of them is your cost. Your cost is whatever one average conversation takes to run, plus everything it takes to keep the agent correct.

What vendors charge today

Vendor roundups in October 2026 list Intercom's Fin AI Agent at $0.99 per outcome, Gorgias at $0.85 to $1.00 per automated interaction depending on plan, and Freshdesk's Freddy AI Agent at $0.10 per session. Alhena, an e-commerce AI vendor, puts the general range at $0.50 to $2.00 per conversation or $0.75 to $2.00 per resolution (Alhena, ClickPost). Those are list prices from companies selling the product, and they change.

The unit matters as much as the number:

Why the pricing page understates the cost

Alhena's own analysis argues that total cost lands at two to five times the pricing page once six layers are added. It sells a competing product, so read its numbers as claims, but the categories are the right questions to put to any vendor or to your own build.

Plate 1 · Where the cost sitsSix layers, vendor-reported

Two items deserve a closer look. The first is integration: Alhena cites $8,000 to $40,000 for API work alone at a mid-market store connecting six systems, and $50,000 or more for custom builds. The second is error correction. It estimates 5 to 15 agent hours a week spent fixing wrong answers, which at $25 an hour comes to $6,500 to $19,500 a year. Both are one vendor's estimates, not measured industry averages.

Build your own number: the cost of one conversation

Whatever you buy or build, the useful unit is the cost of one average conversation. It has more parts than most teams expect.

Plate 2 · Anatomy of one conversationWhat you pay for each time

The model is called on every turn, and each call carries the system prompt, the conversation so far and the products retrieved for that turn. So tokens per turn grow as the conversation goes on. Add retrieval and tool calls, a second model call if you check answers before sending them, and, for the conversations that reach a person, that person's time.

Cost per conversation = turns × (model cost + retrieval and tools + answer check) + handoff rate × cost per handoff.

With the starter inputs, the handoff line is the largest part of the cost, by a wide margin. Change the inputs and watch it move. Double the turns and the model part doubles here. In a real conversation it grows faster, because the history is resent on every turn. Cut the handoff rate and you may save more than any model discount. That is arithmetic on example inputs, not a benchmark, and it is the reason to measure handoffs from day one.

Which pricing model fits which agent

This is our view, not a rule. A shopping assistant has exploratory, multi-turn conversations, so per-conversation pricing grows with engagement. Ask for volume tiers or a monthly cap. A support agent that closes tickets suits per-resolution pricing, but write the definition of "resolved" into the contract. If your traffic spikes in peak season, get the peak-month quote in writing, because that is where hybrid and flat plans diverge.

Six questions to ask before you sign

• What counts as a conversation or a resolution, and who decides?
• What did a peak month cost a store of my size last year?
• Which integrations are in the quote, and which are billed separately?
• What does a handoff cost, and does a conversation that ends in one still bill?
• Can I replay my own shopper questions against the agent before I pay?
• What happens to my data and my test set if I leave?

The cost of a wrong answer belongs in the model

A cost model that stops at the invoice misses the biggest line. A wrong answer costs a refund, a lost customer or, as Air Canada found in 2024, a ruling that the company owns what its chatbot said. That is why the cheapest agent on paper is rarely the cheapest in production, and why we test before launch. Here is how we do it.

Common questions

What is a typical cost per conversation for an AI shopping agent?

Vendor pricing generally falls between $0.50 and $2.00 per conversation or $0.75 to $2.00 per resolution, according to Alhena’s comparison. Your real number depends on turns per conversation, retrieval and tool calls, answer checks and the share of conversations handed to a person, so calculate it with your own inputs.

Is per-resolution pricing cheaper than per-conversation?

Not automatically. It bills only conversations the AI closes, which can look cheaper, but the vendor defines “resolved” and the model rewards closing quickly. Compare both on your own traffic mix.

Why does total cost end up higher than the subscription?

Integration, knowledge-base setup, maintenance, error correction and peak-season overage sit outside the headline fee. One vendor puts the total at two to five times the pricing page.

Does building an agent cost less than buying one?

It depends on what you already have. If your catalog data, search and order APIs are in good shape, a build mostly pays for the agent logic, the checks and the test harness. If they are not, those costs appear either way under a different line. A Diagnostic prices it for your stack.