One parcel is already on its way. The other has never left the warehouse. The customer asks to cancel the second item, and the assistant replies, “Your order has been canceled.”

That sentence can hide three different events. A cancellation request might have been received. A line item might have been canceled. A refund might still be pending. The customer cares which one happened, because it determines whether they need to call again.

This is where the economics of an e-commerce agent become visible. Look up the order first, and the agent can take work out of the queue. Let it guess at the status, and the next contact arrives with an angry customer and a harder case.

Three signals for e-commerce investmentThe sources measure order influence, human support productivity and enterprise AI value. Each has a different denominator.
  • 20%Global Cyber Week orders influenced by AI and agentsVendor attribution includes recommendations and customer service, 25 Nov–1 Dec 2025.Salesforce · 2025
  • +15%Issues resolved per hour with AI assistanceAverage effect in a staggered rollout to 5,172 support workers at one software firm.Brynjolfsson, Li and Raymond · QJE 2025
  • 5%Companies achieving AI value at scaleBCG classification in a 2025 study of more than 1,250 firms worldwide.BCG · 2025

Faster replies are only the first result

Demand and influence. Salesforce attributed AI or agent influence to 20% of global orders in its Cyber Week 2025 analysis (25 November–1 December). Its definition includes personalized recommendations and conversational service. The vendor's methodology combines data from its ecosystem with other inputs and extrapolation. The figure does not identify the fraction of orders autonomously placed, or the sales that would have been lost without AI.

Productivity with a human operator. Brynjolfsson, Li and Raymond studied a staggered rollout of conversational assistance to 5,172 customer-support workers at a business software firm. Access increased issues resolved per hour by 15% on average. Effects differed by worker experience and skill. This is evidence about assistance to people in one setting. Refund accuracy, financial authority and retail margins were outside its measurement.

Value at scale. BCG classified 5% of more than 1,250 companies in its 2025 cross-industry study as achieving AI value at scale. Its enterprise classification does not predict the results of an order pilot. It helps explain why a faster reply is a poor investment metric when order records, permissions and escalation queues still require manual reconciliation.

What the agent needs to know about each parcel

A returns associate scans one parcel while another waits on a shelf
Split shipments need line-item status. A single order-level label can obscure what the buyer can still cancel. Editorial illustration

A customer buys two items. One ships, the second is backordered, and the customer wants to cancel the backorder. The agent must identify the correct line and current eligibility, determine the refund and any shipping change, then obtain approval if needed. It should reconcile the merchant's recorded outcome before promising completion. UCP's Order capability describes line items, fulfillment events and adjustments. Implementations must verify which actions and permissions are available and reconcile local order and payment states.

State or eventWhat the customer needs to knowControl
Checkout times outWhether the original order existsLook up the original request before repeating it
One item shipsWhich item shipped and when the other is expectedRead fulfillment at line level
A cancellation is acceptedWhich line was canceled and the new totalConfirm order state before communicating success
A refund is initiatedHow much is pending and what will confirm settlementReconcile order adjustment with payment state
A policy conflict is unresolvedWho can resolve it and what evidence they haveTransfer the case with order history and reason

A small increase in rework can erase the savings

Suppose 20,000 monthly contacts are eligible for an assistant. At $4.50 per correctly resolved baseline contact, a reduction of $0.75 across those cases would save $15,000 before new costs. If 2% of cases create incremental rework at $20 each, rework costs $8,000. A $5,000 monthly operating cost leaves $2,000. Break-even incremental rework in this example is 2.5% of eligible cases. Measure handling and repeat-contact costs independently to avoid counting the same savings twice. The numbers are a sensitivity example, not a result from the Salesforce or QJE studies.

Follow the clean cases

Over the next few years, expect the boundary of automation to follow the quality of the order record. “Where is my parcel?” is a good candidate when carrier events and line items agree. A partial return is harder when the storefront shows one balance, the warehouse another and the processor a third.

A merchant can track how often the three systems agree before assigning write permission. If the rate improves, expand to a reversible action such as canceling an unfulfilled line. If it stalls, the next improvement belongs in the order workflow and its records.

How to run the pilot without hiding repeat contacts

  1. Baseline. Sample historical contacts by reason, channel and order complexity. Establish cost, resolution and seven-day repeat contact. Identify which system owns each status.
  2. Shadow test. Replay split shipments, lost packages, partial refunds and duplicate retries. Use blinded human review to assess factual correctness, eligible action and customer explanation. Include cases with missing state and require escalation.
  3. Assisted rollout. Randomize eligible cases to current service or agent assistance where practical. Compare correct resolutions per paid hour, cost per resolved case, repeat contacts and customer outcomes. Report the effect by case type and operator experience.
  4. Bounded writes. Permit one action with deterministic eligibility, transaction reconciliation and a rollback or escalation path. Stop the pilot on an unexplained financial mutation, duplicate order or inability to reconcile state.

Evidence limits. Vendor influence estimates, enterprise surveys and a customer-support deployment answer different questions. The local comparison tests whether the same workflow earns a net benefit with the merchant's own systems and policies.

Sources and related research