All insights
EN · Agentic Commerce

How to choose an AI company for ecommerce and agentic commerce

Compare AI commerce partners through six proofs: offer, authority, transaction, operations, economics and ownership—before funding a pilot.

Visual system connects catalog and inventory to agent authority, secure payment, order operations, economics and ownership.
Commerce Partner Proof-6 evaluates the complete commercial chain, including failure and transfer. · Generated with OpenAI

Direct answer: hire an AI company for ecommerce only when it can prove six connected capabilities: represent the offer with current data; preserve customer identity, consent and limits; protect price, payment and authorization; integrate orders, inventory, fulfillment, returns and exceptions; measure incremental economics; and transfer knowledge, data and operations. A polished conversational demo does not prove that a provider can move money and orders safely.

Agentic commerce is moving beyond an interface hypothesis. OpenAI’s Agentic Commerce Protocol documentation already treats structured catalogs, inventory and discovery as infrastructure. Visa and Mastercard document mechanisms for authenticated instructions, transaction controls and verifiable intent. Vendor selection therefore has to cover the commercial system, not only a model, prompt or chatbot.

The MAKINAI Commerce Partner Proof-6

Evaluate six proofs: Offer, Authority, Transaction, Operations, Economics and Ownership. Score each from zero to four: zero means absent; one, a promise; two, a described method; three, partial evidence; four, reproducible execution. Require at least three points in Offer, Authority, Transaction and Operations. A high total cannot offset a disqualifying failure in those areas.

1. Offer Proof: does the agent see commercial truth?

Ask the provider to show how titles, variants, images, price, promotions, availability, shipping, restrictions and return policies move from authoritative systems into AI channels. OpenAI recommends structured feeds, validated fields, regular snapshots and intra-day updates. The proposal should name the source of truth, latency, reconciliation, coverage and owner for every attribute.

  • Required evidence: data contract; catalog and variant mapping; pricing and promotion rules; freshness controls; record validation; prohibited-product handling; discrepancy monitoring; procedure for withdrawing incorrect offers. Red flag: a catalog assembled manually for the demo.

2. Authority Proof: who authorized the agent to act?

The partner should distinguish recommendation, cart preparation, confirmation and autonomous purchase. Require user and agent identity, delegation scope, spending limit, duration, allowed merchants, blocked categories and a re-confirmation rule. Visa documents authenticated payment instructions and controls that test whether credential, merchant and amount match what the customer approved.

Request a readable trail linking intent, context, decision and action. Mastercard’s Verifiable Intent proposal reinforces that authorization should be provable and portable among participants. Do not accept blanket consent hidden in terms. High-value items, regulated products, price changes and low-confidence decisions need explicit human escalation.

3. Transaction Proof: do money and orders remain intact?

Have the vendor demonstrate idempotency, duplicate-order prevention, expired-price blocking, tokenization, authentication, fraud controls, tax, reconciliation, refunds and disputes. The vendor need not replace your acquirer or payment provider; it must integrate existing controls and preserve evidence as an agent crosses channels and systems.

The decisive test is not the perfect purchase. Interrupt the flow after authorization, change inventory or price, simulate a timeout and replay the call. The system should determine whether to continue, ask again, compensate or cancel without charging twice. Require observable states, idempotency keys, owners and recovery times.

4. Operations Proof: can the provider close the physical loop?

Commerce ends in picking, delivery, service, exchange and return. The vendor should map catalog, search, cart, OMS, inventory, seller, logistics, CRM, service and finance. For a marketplace, test an unavailable seller, conflicting SLA, substitution, split order and multi-party dispute. An integration that merely creates orders transfers risk to operations.

  • Ask for runbooks, end-to-end observability, alerts, exception queues, safe replay, severity-based support, recovery objectives, audit and root-cause reporting. The provider should explain what its team operates, what remains with yours and how responsibility changes after the pilot.

5. Economics Proof: is there incremental value after all costs?

Compare providers on incremental contribution, not gross conversation volume. Define a baseline, comparison group and measures such as margin after incentives, completion rate, offer errors, cancellations, returns, fraud, cost per order, avoided service and repeat behavior. Include model, data, integration, payment, observability, human review and support costs.

The pilot should test one economic decision, such as low-complexity replenishment against a stable catalog. Do not attribute every assisted sale to the agent. Predefine incrementality, observation window and customer-quality rules. Stop if gains require unsustainable discounts or shift cost into returns and service.

6. Ownership Proof: does the business control what it is building?

The contract should clarify intellectual property, access to data and logs, portability of prompts and evaluations, licenses, subprocessors, retention, supported models, exit and transfer. Require documentation, automated tests and internal training. The provider should show how components can be replaced without rebuilding the whole channel.

How to run a 90-day pilot

In weeks 1–3, select one journey, record the baseline and design controls. In weeks 4–8, integrate offer, authority, transaction and exceptions in a restricted environment. In weeks 9–12, release to a small segment, track economics and operate real incidents. Scale only when all six proofs have evidence and the disqualifying gates score at least three.

  • Minimum outputs: architecture and data contract; authorization matrix; order state model; threat model; evaluation plan; runbooks; economics dashboard; incident register; transfer plan; scale backlog. Red flags: demo without real systems, fixed price without assumptions, “zero errors,” payments deferred to the end, no returns flow, conversion-only measurement and irremovable dependence.

Next step

Use Proof-6 with the RFP guide at https://makinai.co/insights/en/how-to-write-rfp-ai-services and the general vendor scorecard at https://makinai.co/insights/en/how-to-choose-ai-implementation-company-brazil-scorecard. For the underlying model and integrations, read https://makinai.co/insights/en/agentic-commerce-definition-models-stack-pilot and https://makinai.co/insights/en/technical-architecture-integrations-agentic-commerce. MAKINAI connects AI, CRM, commerce and implementation at https://makinai.co/services/en/crm-ecommerce-commerce-transformation.

Sources and references

  1. Agentic Commerce Protocol · OpenAI

    Documents ACP as an open connective layer that lets ChatGPT ingest structured catalog data, understand inventory and surface products.

    2026-08-21
  2. Get started with Agentic Commerce · OpenAI

    Documents structured product feeds, freshness, validation, delivery methods and prohibited-product responsibilities.

    2026-08-21
  3. Visa Intelligent Commerce · Visa Developer

    Documents agent-specific tokens, authenticated payment instructions, transaction controls and outcome signals for disputes.

    2026-08-21
  4. Verifiable Intent · Mastercard

    Explains an open, protocol-agnostic approach linking identity, intent and action with a tamper-resistant authorization record.

    2026-08-21
  5. Generative AI Profile · NIST

    Provides a cross-sector companion to the AI RMF for managing generative-AI risks throughout design, deployment and operation.

    2026-08-21
Making connections

Continue exploring

Agentic Commerce

Operational Playbook for Agentic Commerce: Seller Standards, Machine-Readable Offers, APIs, Anti-Gaming, and SLAs

Read insight
Agentic Commerce

Agentic Commerce: Definition, How It Differs and How to Pilot It

Read insight
Agentic Commerce

Agentic Commerce: A Practical Business Case Framework and 12-Month Pilot Checklist

Read insight
Related capability

CRM, commerce & new models

A modern CRM and commerce strategy connects identity, data, communication, experience, offer and operations. MAKINAI helps companies organize that architecture and implement journeys that learn from customer behavior, coordinate channels and prepare the business for agents and emerging buying interfaces.

Explore this capability