How Much Does an eCommerce AI Agent Cost in 2026?

On 25 September 2026 there is still no honest single price for “an eCommerce AI agent.” The number that keeps getting reused is a cost floor, not a project.

In July 2026 a first-party FactualMinds benchmark put one Bedrock AgentCore support silhouette at about $791 per month at 50,000 sessions, platform plus a small (Haiku-class) model. That comparison lives in the AgentCore vs Amazon Q decision guide. It is not a client bill. It does not include the people who wire Shopify, the eval set, or the associate who still approves a refund.

Who this is for. A CEO, COO, or eCommerce leader who has been handed a vendor slide with one blended “per conversation” price.

Our take: budget seven lines. Refuse any quote that collapses them.

What you are actually buying

An agent that looks up an order is a different bill from an agent that drafts a B2B quote. Both are different from a shopping protocol (ACP or UCP) that lets someone else’s model buy from your catalog. This post is the outbound agent you run. Inbound selling is agentic commerce readiness.

AgentCore itself reached general availability on 13 October 2025. Harness (CreateHarness / InvokeHarness) reached GA on 17 June 2026. Cheaper hosting did not remove integration cost. Amazon Bedrock Agents Classic is the wrong net-new path after 30 July 2026.

The seven cost lines

Maintenance is an eighth line in real life. Put it on the worksheet even though the heading says seven. The point is the split, not the count.

How to calculate, without inventing a total

Use your volume. Do not borrow someone else’s.

1. Count sessions you expect in a steady month. A session is a conversation or a job, not a page view. Write the definition down.
2. Count tool calls per session for the one workflow. A where-is-my-order turn might be two reads (order, shipment). A quote draft is more.
3. Price the platform and the model in the AgentCore pricing calculator. The 12-component breakdown is in AgentCore pricing. An example shape in that post — labeled as not a client — is 250,000 sessions with about 3.2 seconds active. Use it to see which line dominates. Do not paste it into a board deck as your forecast.
4. Add integration as a fixed scope. Named APIs, a sandbox, and a person who can explain a failed order id. If you cannot name the APIs, you are funding discovery, not a build.
5. Add evaluation as a fixed scope. Ten real lookups and three cases that must escalate. The method is ten tickets before the agent talks to a customer.
6. Add human review as hours, not as “the AI saves the team.” Week one is lookups. Refunds, price changes, and purchase orders stay with a person. If leadership’s only win is an unattended write, stop — that is the readiness veto below 16 out of 30.
7. Add a monthly maintenance band for API upgrades and a regression run. A model price cut does not zero this line.

Then stop. A summed “typical project” would be a fiction. ROI ranks which workflow to attempt. It is not a savings guarantee.

Where the bill surprises people

What broke — A support-shaped pilot left AgentCore Browser enabled on every turn. Lookups that only needed the order still opened a browser session. The field guide records support-shaped pilots running about 3× the platform spend when that browser stayed on. Detection: the platform line moved and the answer quality did not. Fix: Browser off; carrier lookup as a named tool. Lesson: optional tools are not free because the prompt rarely calls them. The runtime still can.

Two other misses, both structural:

• Blended vendor pricing. You cannot tell model cost from margin. Ask for the worksheet in the RFP template.
• Retry storms. A failed write that the model “helpfully” retries is a duplicate order risk and a token spike. Idempotency keys belong in the integration line, not in a postmortem.

What this does not decide

Cost does not pick build versus buy. That comparison is build vs buy vs partner-build. Partner-build still leaves you the monthly run cost. Buying a SaaS widget can be cheaper when you have no APIs — and useless when you do.

If you only do one thing

Open the AgentCore pricing calculator with your session count. Write down what the $791 figure excludes. If the excluded lines are blank, you do not have a budget yet.

What to do this week

1. Pick one workflow. Not a fleet.
2. Write session definition and a monthly count.
3. Run the calculator. Keep the export.
4. List APIs the agent must call. If the list is “the admin token,” you are not ready to price integration.
5. Read which agent first before you add a second workflow.
6. Talk to FactualMinds about the cost worksheet, not about a guaranteed saving.

What this post doesn’t cover

• A FactualMinds rate card. We do not publish one here.
• Client invoices. There is no published AI-agent case study on this site.
• Inbound ACP or UCP merchant fees. Those are protocol and platform commercial terms, not this run-cost model.
• Quick Suite seat math, except as the contrast in the July 2026 benchmark (about $3,580 per month for a 500-employee, 25,000-document workforce assistant — a different product).