E-commerce: order status, returns, and WhatsApp recovery that actually converts

The shape of e-commerce support

Pull a representative sample of an e-commerce brand's inbox and you will find the same three questions, over and over:

1. "Where is my order?" The caller has a tracking number somewhere in their email but cannot find it.
2. "How do I return this?" Usually followed by "and will you pay for shipping?"
3. "Does this fit / is this in stock / is this coming back?" Product questions that are already answered on the PDP, but the customer wants a human to confirm.

An AI agent that answers these three well — grounded in the real order record, the real policy document, and the real product catalog — handles enough volume that the human team can focus on the chargebacks and VIP escalations where they add real value.

Grounding is everything

The reason naive support bots fail e-commerce is that they are untethered from the data. They hallucinate return windows. They promise shipping that is not available. They "look up" an order they cannot actually see.

The fix is boring and essential: give the agent tools that read from the real systems.

• Order lookup by email + order number. Not by name. Not by partial match. The input identifiers that will not surface the wrong order.
• Return eligibility by SKU and purchase date. Policy is not uniform. Final-sale categories, holiday extensions, swimwear, intimates. The agent reads the policy, applies it to this order, says the specific answer.
• Inventory and restock expectations. If you have an ETA, surface it. If you do not, say so plainly. "Restock expected" with no date is worse than "we don't know."

What good looks like

For e-commerce, three channels do most of the work:

• Web chat on the PDP and the help center. Catches the pre-purchase questions and the "where is my order" trickle. Low friction, high volume.
• Email autoresponder tied to the same agent. Writes the first reply, including a draft the human can send with one click. Reduces first-response time without putting a human-in-the-middle on every ticket.
• WhatsApp for shipping updates, returns, and recovery. This is where the leverage is, and most brands under-invest in it.

WhatsApp recovery, done without being creepy

Cart recovery is the obvious use case, but it is also the one most brands over-rotate on and burn out their list. A better pattern:

• One nudge, timed to the pattern of your category. Fashion is 4 hours. Electronics is 24. Perishables is 90 minutes. Do not send three.
• Open with the specific product, not "you left something behind." "The Oxford shirt in navy, size M — still thinking it over?" is a different message than a template.
• Give the agent a real tool: offer a discount, but only if it is within the margin band you set. A hard "no discount, ever" is too rigid; an unbounded "offer whatever" is a cost center.
• Hand off to a human if the customer asks a sizing or fit question. This is where humans close.

Common failure modes

• Agent cannot see the order. The customer gives their email; the agent "cannot find it in the system." You just lost the trust of someone who had a tracking number three clicks away. Fix the tool before you ship the agent.
• Policy drift. Marketing changes the return window for a sale and nobody updates the knowledge base. The agent is now wrong, confidently. Treat the policy doc the agent reads from as the source of truth, not a copy.
• Over-apologizing. Agents trained to be deferential say "I'm so sorry" ten times. Real support says "Here is what happened, and here is what I can do." Set the tone explicitly.

How ZemKit fits

• Tools, not scripts. The agent calls your order API, your return endpoint, your inventory system. It is not pattern-matching on text.
• One agent, chat + WhatsApp + email. The same brain answers across channels. The customer who opened a return on the website and asked for a status on WhatsApp gets a consistent answer.
• Escalation with context. Chargebacks, VIPs, "I want to speak to a manager" — these route to a human with the full transcript, the order, and the policy decision the agent made. The human picks up the thread, not the problem.

What to measure

• Resolution rate without human touch. For the three canonical questions, this should be above 60% within 30 days.
• Agent-recovered revenue. Cart nudges and return-saves attributable to the agent. Instrument this before you launch, not after.
• Negative-signal rate. How often does the customer ask for a human? Rising negative signal is the leading indicator that policy drift has happened — before NPS catches it.

---

If you want to see this on your own catalog and policy doc, we will build a working support + recovery agent before the demo so you can push on it with real customer messages.