AI in customer support for commerce: automating without losing the human touch

Customer support in eCommerce is at an inflection point. Customers expect immediate answers (82% want a response in under 10 minutes, according to Salesforce, 2025), yet support operations remain labor-intensive and costly to scale. Every peak season — Hot Sale, Buen Fin, Christmas — brings the same tension: tickets spike, response times degrade and satisfaction drops.

AI doesn't solve this problem by replacing humans with bots. It solves it by building a hybrid model where AI conversational agents handle repetitive queries with precision and human agents focus on the cases that require empathy, judgment and negotiation.

At Edgebound Labs we build conversational agents that access the catalog, order history and brand policies in real time. They are not generic bots — they are systems specialized for commerce that understand your business.

The problem with traditional commerce support

Before talking about AI, you have to understand the problem. Support operations in enterprise commerce share these patterns:

• Volume concentrated in repetitive queries: 60-70% of tickets are questions about order status, return policies, availability and payment issues. They have a defined answer and require no human judgment.
• Unpredictable seasonal peaks: during Hot Sale or Christmas, volume can increase 3-5×. Hiring and training temporary agents takes weeks and quality suffers.
• Rising costs: the average cost per ticket with a human agent is US$6-12 (Zendesk). For 10,000 monthly tickets, that's US$60K-120K a year on L1 support alone.
• Fragmented channels: email, web chat, WhatsApp, social media and phone — each channel with a different team or tool, without a unified view of the customer.

Conversational agents with RAG: the new generation

Rule-based chatbots (decision trees with predefined options) were useful in their time, but they have clear limits: they don't understand natural language, they don't handle variations and they frustrate the customer who can't find their case in the menu. Agents based on RAG (Retrieval-Augmented Generation) are a completely different architecture.

How RAG works for commerce support

• Step 1 — Ingestion: knowledge sources are indexed in a vector database: catalog, return policies, FAQ, sizing guides, T&C, change history.
• Step 2 — Retrieval: given a question, the system searches for the most relevant fragments in the vectorized database — more precise than keywords because it understands semantic meaning.
• Step 3 — Generation: the LLM (GPT-4, Claude) generates the answer using those fragments as context. It doesn't make things up — it responds with real, verifiable information.
• Step 4 — Action: for queries that require action (check an order, start a return, apply a coupon), the agent calls the commerce APIs in real time.

At Edgebound we use the OpenAI Assistants API and Anthropic Claude as base models, with vector databases in Pinecone or pgvector on AWS. The agent connects to the APIs of commerce engines (Shopify, Commercetools, BigCommerce) to execute actions in real time.

The hybrid model: AI + humans

The best support is neither 100% automated nor 100% human. It's a hybrid model where each layer handles what it does best.

Tier 1: AI agent (automated) — 60-70% of volume

• Order status and real-time tracking.
• Return and exchange policies (with automatic eligibility verification).
• Product availability and alternatives; common payment issues.
• Questions about shipping, hours, stores; sizing guidance and recommendations.

Tier 2: AI-assisted human agent — 25-30%

Cases that the AI agent escalates with all the context preloaded: damaged-product claims, chargeback disputes, orders with multiple issues, VIP customers. The human agent receives a summary of the conversation, the history, recent orders and the applicable policy. They don't start from scratch.

Tier 3: Specialist (exceptions) — 5-10%

Legal cases, critical escalations and commercial negotiations that require special authorization.

Metrics that matter

Agentic commerce: the future that already started

In 2026, conversational agents are evolving toward agentic commerce: AI systems that don't just answer questions but execute complex flows autonomously. Examples:

• "I want to return the sneakers I bought last week" → the agent verifies the purchase, validates the policy, generates the shipping label and schedules the pickup. No human intervention.
• "I need something similar to what I bought, but in size 9" → it checks the history, finds in-stock products, applies a loyalty discount and generates a direct purchase link.
• "I was charged twice" → it verifies the transactions, detects the duplicate charge, initiates the refund and confirms. If it can't resolve it, it escalates to L2 with all the context.

How to automate without losing the human touch

The biggest risk of automating support isn't technical — it's experiential. A bot that frustrates is worse than a long wait. The rules we follow at Edgebound:

• Transparency: the customer always knows they're talking to an AI agent. We don't pretend to be human.
• Easy escalation: at any moment they can ask for a human. One click, no menus, no repeating the story.
• Brand personality: the agent speaks in your brand's tone, not like a generic bot.
• Clear limits: the agent knows what it doesn't know. If it doesn't have enough information, it says so and escalates. It never makes things up.
• Continuous improvement: every conversation generates data to improve the system.

Frequently asked questions (FAQ)

← Back to Lab Report