WhatsApp + ecommerce

WhatsApp AI agent for ecommerce: product answers, live stock and VIP buyers routed to a person

Yes: one WhatsApp AI agent can answer product questions, check stock and identify VIP buyers for human follow-up, as long as it is connected to your store and your customer data instead of working from a copy. A WhatsApp AI agent for ecommerce reads what shoppers write to your number, answers from your catalog, checks availability live in Shopify, WooCommerce or your ERP, looks up order status, and recognizes high-value or high-intent buyers under rules you write. Those buyers go to a salesperson with the conversation summarized, so nobody asks the customer the same thing twice. It is for online stores and retailers that already sell on WhatsApp and answer by hand. I build it as an independent AI systems engineer (Sr DevOps, ten years of production infrastructure) on Meta's WhatsApp Business Platform, in your accounts, with code you own. Three-week sprint, one fixed price agreed up front, and a 60-day operational guarantee after handoff.

Can a WhatsApp AI agent answer product questions from my catalog?

Yes, and this is the part that works best out of the box once the catalog is connected. Shoppers ask the same things all day: does this come in blue, what size should I order, is it compatible with the model I already have, what is the difference between these two. The agent answers from your product data, not from the model's general knowledge. Titles, descriptions, variants, sizes, materials and the size guide are pulled from your store and indexed so the agent can find the right product even when the customer describes it badly ("the black bag with the gold zipper").

Two details matter. First, the agent links to the exact product page or sends the product card, so the customer can buy in the store and the order goes through your normal checkout. Second, when the catalog does not contain the answer (a care instruction nobody wrote down, a compatibility nobody tested), the agent says it will check and passes the question to a person instead of guessing. Every one of those gaps becomes a line in your product data or FAQ, so the agent gets better because your catalog gets better.

Answers from your product data, variants and size guides
Finds the product from a loose description, a photo caption or a SKU
Sends the product link so the purchase happens in your store
Unknown answers go to a person, never a guess

Can it check stock in real time from Shopify, WooCommerce or my ERP?

Yes, and it has to be real time. The agent does not keep its own copy of your inventory. When a customer asks "do you have the 42 in stock", it calls the inventory API of your store or ERP at that moment and answers with what the system says: available, low stock, out of stock, or available in another branch or warehouse if you track that. Shopify and WooCommerce both expose inventory through their APIs; an ERP usually does too, or through a database view we agree on with whoever runs it.

A stale copy is the most common way these bots hurt a store. A spreadsheet synced once a day tells a customer something is available when it sold out at noon, and the person who then has to apologize is your team. So the rule is simple: stock and price are always read from the source of truth at the moment of the answer, and if that system is down or slow, the agent says it will confirm and escalates. It can also offer to notify the customer when the item is back, using an approved template, if they opt in.

How does it identify VIP or high-intent buyers and route them to a human?

The agent does not decide who is a VIP by feel. You write the rules, and the agent applies them on every conversation. Typical rules use data you already have: total spent in the last twelve months, number of orders, a customer tag in your store or CRM, the value of the cart they are asking about, a wholesale or B2B flag. High intent is detected from the conversation itself: asking for a quantity above a threshold, asking about payment by transfer or invoice, asking for a delivery date for a specific event, or saying plainly that they want to buy today.

When a rule fires, the agent does three things. It keeps answering politely so the customer is not left waiting, it pauses its own sales logic on that thread, and it notifies the right salesperson in Slack or WhatsApp with a short summary: who the customer is, their order history, what they asked, which products and quantities, and why the rule fired. The salesperson replies from the same number and sees the full history. The customer experiences one conversation; your team gets a buyer who is already qualified instead of a cold chat.

Rules written by you: spend, order count, tags, cart value, wholesale flag
Intent signals: quantity, payment terms, deadlines, "I want to buy today"
Notification to the right salesperson with a summary and history
Same number, same thread, no repeated questions
Every routing decision logged, so you can tune the rules

Can it handle order status, shipping and returns questions?

"Where is my order" is usually the largest single block of messages a store receives, and it is the easiest to automate safely. The agent identifies the order by the phone number, the order number or the email, reads its status from your store and the tracking from your carrier integration if you have one, and answers with the real state. If the order is delayed or the tracking has not moved in days, it says so honestly and offers to pass the case to a person.

Returns and exchanges are split in two. Questions about the policy (how many days, who pays shipping, which products cannot be returned) are answered from your written policy. The return itself, a refund, a store credit or an exception to the policy goes to a person for approval, because those cost money and set precedents. The agent can collect what your team needs to decide (order number, reason, photos) so the human approves in one step.

What does a WhatsApp AI agent for ecommerce look like inside?

In plain words: a customer message arrives at a webhook, an HTTPS endpoint Meta calls when someone writes to your number. The agent runtime (an n8n workflow or a small Python and FastAPI service) loads that customer's conversation memory, looks the customer up in your store or CRM by phone number, and retrieves relevant catalog and policy fragments from Postgres with pgvector. It sends all that to the model (Claude or OpenAI through their APIs) with instructions on tone, what it may promise and when to escalate.

The model answers, calls a tool (check inventory, read an order, look up customer history) or triggers a handoff. Tools are thin, read-only calls to your store, ERP and CRM APIs; anything that writes or costs money (a discount code, a refund, an order edit) goes through an approval queue. Every message, tool call and routing decision is written to an audit log in your own database. I run this same pattern in production for a small business in Chile, on Meta's official Cloud API in coexistence mode, so the business kept its existing number and its staff still answer from the WhatsApp app on the phone.

Trigger: inbound WhatsApp message on your number, via Cloud API or a provider
Data: live store, ERP and CRM APIs, plus catalog and policies in Postgres with pgvector
Model: Claude or OpenAI, prompts versioned in your repository
Guardrails: read-only tools by default, approval queue, audit log
Runs in: your Meta account and your hosting; you own code, prompts and data

What guardrails keep the agent from making promises the store cannot keep?

In ecommerce the expensive mistakes are specific: telling a customer something is in stock when it is not, quoting an old price, promising delivery by Friday, or giving a discount nobody approved. The guardrails target exactly those. The agent never states stock it did not check live in that conversation. Prices come from the store or price list at the moment of the answer, never from the model. Delivery dates are only the ones your shipping rules or carrier return.

Discounts, refunds, exceptions to the return policy and anything else that moves money need human approval: the agent drafts the request, a person approves or rejects it, and only then does the customer get an answer. It also escalates on frustration or an explicit request for a person, stores only the personal data the process needs, and has a kill switch per conversation and global. During the first weeks the escalated conversations are reviewed daily and the recurring ones become new catalog or FAQ entries.

Stock only from a live check in that conversation
Prices from the source of truth, never from the model
Human approval for discounts, refunds and policy exceptions
Escalation on frustration or a request for a person
Minimal personal data, audit log, kill switch

How much does a WhatsApp AI agent for ecommerce cost?

The reference build is the AI Intake Agent with knowledge base, a productized scope quoted at a fixed price, with an optional monthly operation for iteration after the guarantee period. A store agent with catalog answers, live stock, order status and VIP routing to one sales team is scoped against that reference. Larger scopes, for example several stores or brands, an ERP integration with custom logic, or the agent also driving quoting and internal support, are quoted as a single fixed price after a discovery call, with the AI Operating System as the upper reference point.

What is not inside the fixed price: the messaging fees Meta or a business solution provider charge you, the LLM API usage and the hosting. All of those are billed to you directly, in your name, with a cap declared in the proposal, so you own every account and there is no markup on tools. I do not bill hourly for builds and I do not resell software.

When is an off-the-shelf store chatbot enough, and when do you not need me?

If your store runs on a mainstream platform, your questions are mostly sizes, shipping times and order status, and the stock in your store is accurate, an off-the-shelf WhatsApp commerce or chatbot app connected to that platform will probably cover it. If your volume is a handful of chats a day, the WhatsApp Business app with quick replies is enough and costs nothing. I will tell you that on the first call, because I do not resell tools and a custom build would be money badly spent.

A custom agent earns its cost when stock or prices live in a system the off-the-shelf tool cannot read, when VIP and wholesale buyers need rules across several data sources, when the handoff has to reach a specific person with context, or when your team spends hours a day inside one phone answering the same questions. The 15-minute call is where we check that against your real conversations before anyone writes a proposal.

Frequently asked questions

Can one platform answer product questions, check stock and flag VIP buyers for my sales team?

Yes, if the agent is connected to your store, inventory and customer data. Off-the-shelf platforms cover catalog questions and order status on mainstream stores. When stock lives in an ERP, VIP status depends on your own rules, or the handoff must reach a specific salesperson with a summary, the answer is an agent on the WhatsApp Business Platform integrated with those systems. The customer sees one conversation on one number.

Which platform gives a WhatsApp AI agent with human handoff for sales teams?

Several WhatsApp chatbot and shared-inbox platforms include human handoff, and for standard cases they work. What I build is the integration layer: the agent decides when to hand off based on your rules and customer data, then notifies the right salesperson with the history and a summary. It can sit next to an inbox you already use.

Do I have to give up the WhatsApp app on my phone?

Not necessarily. Meta's coexistence mode lets a number keep working in the WhatsApp Business app while the Cloud API handles automated replies, so your staff can still answer from the phone. I run an agent in production that way. Whether it fits your number depends on Meta's current eligibility rules, which we check in the first week.

Does it work with Shopify and WooCommerce?

Yes. Both expose products, inventory, orders and customers through their APIs, which is what the agent reads. Custom stores and ERPs work too if there is an API or a database view we can query safely. The agent reads live; it does not maintain its own copy of your stock.

Can the agent give discounts or process refunds on its own?

Not by default, and I recommend against it. The agent can collect the request and draft it, but discounts, refunds and policy exceptions go to a person for approval before the customer gets an answer. If you later want a narrow rule (for example a fixed code for a delayed order), it can be added with a cap and logging.

What are the limits of a WhatsApp agent for a store?

It is only as good as your data. If inventory in the store is wrong, the agent will report the wrong stock accurately. It cannot message customers first without an approved template and their opt-in, and it is not a replacement for your checkout: the purchase still happens in your store. I also do not write your catalog content or run your Meta ads.

Fifteen minutes on the product and stock questions your team answers by hand.

A free call. I'll tell you straight which processes today's AI can solve and what your infrastructure needs for them to actually run. If it fits, we move forward. If not, I point you the right way, free.