A working demo is not yet a commerce agent.

The most useful signal in Accenture’s study is not the trust figure in the headlines. It is the distance between asking an agent for help and handing it the payment.

Consumer Pulse Research 2026 surveyed 25,590 people in 16 countries. It measures stated willingness, not observed purchases. It still shows where early features belong: in tedious and reversible tasks, before the final decision and well before financial autonomy.

Trust is not payment authorization

Source: Accenture Consumer Pulse Research 2026. Percentages show stated openness at each level. They are not usage or conversion rates.

The 7 gates

1. 01

   Commercial facts come from the source of truth

   The model may explain a trade-off. It should not produce an identifier, price or availability. Those values come from merchant catalogue, inventory and policy systems, with country, currency and customer context.

   Exit testA stock change between search and cart is detected, explained and never produces a silent substitution.

2. 02

   Every tool has a narrow permission

   Comparing products does not require order history. Reading performance does not grant permission to change a price. Read, stage and apply capabilities remain separate.

   Exit testAn identifier invented by the model, pasted by a user or planted in a review is rejected before the backend.

3. 03

   A proposal is never an authorization

   Every consequential action becomes a staged change. A person approves it in a real surface, then the server rechecks stock, budget, caps and policy at the exact moment of application.

   Exit testA promotion that becomes invalid after approval is cleanly rejected and remains traceable.

4. 04

   Cart, order and payment are 3 different states

   A built cart is not an order. A created order is not an authorized payment. The system exposes each transition and reconciles a timeout before attempting it again.

   Select→Cart→Checkout→Authorize→Confirm
5. 05

   Safety lives beyond the prompt

   Quantity caps, budgets, protected fields and regulated copy are enforced in code. Concurrent writes are serialized. A prompt may guide behavior, but it cannot guarantee a financial action.

   Exit testA malicious instruction in a product name or review expands no permission.

6. 06

   Evals follow the whole transaction

   A good answer can hide a bad tool call. Tests inspect tools, state, product cards, protected language and the final outcome. Anthropic suggests 50 to 100 cases per flow as a starting point, designed with product, legal, operations and support specialists.

   Exit testStock change, duplicate attempt, cross-account access and data-plane injection cases pass in CI.

7. 07

   Economics are measured per correct task

   A cheaper model can consume more turns, tools and corrections. The useful denominator is a correctly completed task, with a quality floor plus median and tail latency budgets.

   Exit testThe scorecard separates cost per call, cost per completed task, recovery rate and human correction rate.

Two consequences merchant teams should plan for

The first is commercial. In Accenture’s survey, 37% of consumers classified as loyal say they would let an agent switch brands for a better fit. This does not predict an observed switching rate, but it indicates what merchants will need to make comparable: total price, availability, terms, quality and fulfilment evidence. A brand promise that merchant systems cannot substantiate may not enter the agent’s shortlist.

The second is evidentiary. Mastercard’s framework describes intent data covering cart contents, transaction limits and validity windows, plus an audit trail that may help with disputes. For the business, that argues for logs connecting recommendation, consent, tool call, state change and payment result. Mastercard does not say this record eliminates disputes or transfers liability; it provides verifiable context for investigating them.

Sources for this analysis: Accenture Consumer Pulse 2026 · Mastercard Agent Pay Acceptance Framework

A rollout in 3 moves

“Trust is the currency of commerce.”

Sherri Haymond, Mastercard, quoted by Anthropic

The useful verdict

A demo proves that a model can hold a conversation and call a few tools in a planned scenario. A production-ready commerce agent must prove something else: a commercial fact not confirmed by the source of truth cannot reach the cart, payment or merchant systems, and an unauthorized or stale action fails closed.

These 7 gates do not certify a system. They give the merchant concrete failures to provoke before real customers find them.