ZZ Solutions
Work/AI SystemsAI-native pattern

Cross-Platform Commerce Orchestration

Agentic multi-merchant cart unification, cross-platform product sourcing, and fulfillment coordination — one checkout from any combination of merchants

~3 hrsSaved per complex orderManual multi-merchant shopping
5+Merchant platforms unifiedIn a single checkout
$12–35Modeled fee per orderCoordination value capture
100%Approved before paymentHuman-in-the-loop always
Overview

The problem & context

This case is grounded in the SEN commerce problem — not generic product search, but the genuine complexity of buying from multiple US merchants in a single coordinated transaction. A user wants an iPhone from Apple.com, headphones from Amazon, and a book from a specialty retailer. Without an orchestration layer, that means three browser sessions, three checkout flows, three shipping timelines, three return policies, and hours of coordination. The engineering challenge is creating one reliable agentic workflow that spans arbitrary merchant combinations — and keeps the human in control before any money moves.

Challenge

Multi-merchant purchasing is unsolved not because search is hard, but because the coordination is. Merchant APIs return incompatible product structures. Checkout flows don't compose. Delivery comes from three different sources. Returns require contacting three different support teams. The friction is so high that users either limit themselves to a single platform or abandon the purchase entirely. A commerce assistant that only improves search misses the actual problem.

Solution

I built the backend architecture for an agentic commerce workflow: Saya (the AI) parses the user's intent, runs parallel product discovery across merchant APIs, normalizes incompatible product data into comparable candidates, assembles a multi-merchant cart, and presents a structured approval before executing anything. The execution layer then places orders across merchants, tracks delivery from each source, and manages returns through a single interface. The system is explicitly designed so no payment is ever triggered without a human decision.

Architecture

AI system design

Type colors:● AI● System● Human● Output● Data

Loading diagram…
Business Impact

By the numbers

Operational metrics and modeled business impact. Hover bars for exact values.

Loading chart…
Impact

Measurable outcomes

Each ring animates to its final value on load.

93%

Checkout time saved vs manual

85%

Return friction removed

100%

Human approval coverage

90%

Cross-platform data normalized

Before & After

What changed

BeforeAfter
Time per multi-merchant order93% faster
Before~3 hours (manual)
After~12 minutes (Saya)
Merchant coverage per request5× broader
Before1 platform
After5+ platforms
Checkout flows requiredSimplified
Before3–5 separate forms
After1 approval gate
Returns processUnified
BeforePer-merchant separately
AfterUnified interface
Design North Star

Let a user buy from any combination of US merchants — in one conversation, one approval, one checkout.

Design Rationale

Why this architecture, not another

The core insight is that multi-merchant commerce is a coordination problem, not a search problem. Any single merchant already has good search. The unsolved problem is: how do you combine a cart from five different stores, normalize incompatible product data, and present one coherent decision to the user before any money moves? The agent-per-responsibility model is the answer. Each stage — discovery, normalization, assembly, quality control — owns its own failure mode. The approval gate isn't optional: it's the trust boundary that makes the rest of the automation acceptable to users.

System Design

Architecture principles

The choices that made this system practical in production: clear ownership, controlled failure modes, and business evidence before automation.

01

One agent, one responsibility

Discovery, normalization, assembly, and quality control are separate agents. Failures are isolated — one broken merchant integration doesn't collapse the whole workflow.

02

Human approval at every irreversible boundary

The AI orchestrates everything up to and including the recommendation. Payment and order placement require explicit user approval.

03

Evidence before action

Every recommendation carries source data, price breakdown, shipping estimate, and validation result so the approver understands exactly what they're approving.

Delivery Model

How the implementation was structured

01
Phase 01

Merchant API abstraction

Build a connector layer that normalizes product data, availability, and pricing from each merchant into a common schema.

02
Phase 02

Agent orchestration

Implement parallel discovery, cart assembly, and quality gate agents with structured outputs and per-agent retry logic.

03
Phase 03

Approval + fulfillment

Wire the approval gate to multi-merchant order placement, delivery aggregation, and centralized returns management.

Production Controls

What made it safe to scale

Scalable

Merchant connectors are independent — adding a new platform is a new connector, not a rewrite.

Reliable

Failures are isolated by agent responsibility; one unavailable merchant doesn't abort the whole workflow.

Trustworthy

Payment is always gated by human decision. The system never acts on money without explicit approval.

Agent Stack
Production

AI agent responsibilities

Master Orchestrator
Coordinates the full pipeline — dispatches agents in parallel, synthesizes results, decides whether to present or retry with additional context
Discovery Agent
Searches across multiple merchant APIs simultaneously using semantic matching — surfaces the most relevant results without manual browsing
Normalization Agent
Converts incompatible merchant product structures into a single comparable schema — makes cross-platform comparison accurate and explainable
Cart Assembly Agent
Constructs the multi-merchant cart, calculates consolidated shipping estimates, and flags availability or delivery conflicts before approval
Quality Gate Agent
Validates all results against business rules before presenting to the user — prevents malformed, out-of-stock, or out-of-policy items from reaching checkout
Outcomes

Business impact

  • Users can request products from any combination of US merchants in a single conversation — the complexity of multi-merchant coordination is fully absorbed by the system
  • Checkout time for a complex 6-item, 3-merchant order drops from ~3 hours of manual browsing to ~13 minutes of AI-assisted workflow
  • All irreversible actions (payment, order placement) sit behind a human approval gate — the AI prepares, the user decides
  • Returns and delivery tracking from multiple merchants are managed through one interface instead of separate vendor support queues
  • Service fee model captures value at the coordination layer — the harder the shopping task, the more value the platform delivers
Stack
Multi-Agent AI OrchestrationOpenAI Assistants APINext.js 15NestJSGraphQLGCP Cloud RunPub/SubMongoDBRedis

Need something like this?

Book a free 90-minute audit. We'll look at your workflow and tell you honestly whether the same approach applies.