Custom WhatsApp Assistant for WooCommerce

You outgrow ManyChat when your WhatsApp conversations become valuable enough to deserve a bot that actually knows your store, rather than a flow that only pretends to. The upgrade is a custom WhatsApp assistant: a small backend on your own Vercel project connects straight to WhatsApp's official Business Platform, reads live from your WooCommerce data, and answers with Claude as the brain, calling tools that fetch real orders, stock, and prices. You keep WhatsApp itself, because that channel was never the middleman; what you remove is the bot platform metering your contacts and holding your logic. The result answers "did my order ship" from the actual order, bills by real usage instead of contact-list size, and runs entirely under your own account and keys.

The short version

ManyChat is a great start and a poor finish. Its flow builder gets a young store live fast, but a flow is a script that cannot see your catalog, your stock, or a returning customer's history.
The channel stays; the platform goes. WhatsApp messages always travel over Meta's official Business Platform. What a custom build removes is the vendor sitting between you, your data, and the conversation.
The assistant is grounded in your real data. Claude is handed tools that read live WooCommerce records, so "is this in stock" and "where is my order" get answered from the truth, not a menu.
Cost stops tracking your list size. A per-contact plan charges more as your audience grows whether or not those people message you; a custom build charges by WhatsApp's per-conversation fee, a little hosting, and the model usage you choose.
You own the whole thing. Your WhatsApp account, your Vercel project, your keys, your WooCommerce data, with no lock-in and no third party in the chat.
The move is staged, not a leap. You migrate one or two high-value jobs first, prove them on real questions, then widen, so nothing customer-facing breaks along the way.

What does ManyChat get right, and where does it stop?

It is worth being fair, because ManyChat earns its place at the start. The visual flow builder is approachable, you can stand up a basic experience without a developer, and the marketing side, broadcasts and simple opt-in campaigns, is exactly what wins a young store its first repeat sales. As a way to prove that customers will engage with you on WhatsApp at all, it does the job. The trouble is not that the tool is bad; it is that the tool is built to be generic, and your store has stopped being generic.

The limits that eventually arrive tend to come together, and they are structural rather than anything you can configure your way out of.

It does not actually know your store. A flow is a script. It cannot look up a real order, read live stock, quote the right price for a variation, or recall what a returning customer bought last month, because it has no live connection to the data that would make those answers possible.
Your customers' conversations run through someone else. Every message, name, and order detail passes through a third party's systems, which is a layer of trust you hand away on exactly the conversations that touch purchases and personal information.
The bill climbs as you succeed. Per-contact pricing means the bigger and more engaged your audience gets, the more you pay, which is the worst possible moment for a cost to rise.
The intelligence is shallow and bolted on. Whatever AI sits on top of a flow engine was never designed for genuine understanding, so anything nuanced, multilingual, or specific to your products hits a ceiling fast.
You cannot build past the platform. The day you want behavior the builder does not offer, you are negotiating with a tool instead of simply building what your business needs.

Each of these is something a customer feels, in a slow or wrong answer, and something your margin feels, in a sale that did not close or a bill that keeps rising. Recognizing that moment is the same judgment we lay out in build versus buy for an AI chatbot: keep renting while the tool fits, and build when it stops.

What does "custom, without a third party" actually remove?

Let us be precise, because the honesty matters here. The one piece you cannot remove, and would not want to, is WhatsApp itself. Messages travel over Meta's official WhatsApp Business Platform, the same way email travels over a mail provider. That is your channel, not a middleman.

What you remove is the bot platform sitting between you and WhatsApp, and between you and your own data. Instead of ManyChat relaying messages and holding the logic, your own small backend connects straight to WhatsApp's official API, reads directly from your WooCommerce database, and answers with Claude as the brain. It runs on your Vercel project, under your account, with your keys. Nobody else is in the conversation, and nobody else is metering your contact list. That ownership is the whole point of leaving a rented platform, and it is the same instinct we describe in why connecting your stack beats copy and paste.

How does the architecture work, in plain terms?

The whole thing is simpler than it sounds, and it is the same serverless pattern we use for production payments and compliance software. A message comes in, your backend gathers the facts, the model decides what to say, and the reply goes out.

// Vercel serverless webhook: WhatsApp in, Claude (with Woo tools) out
export async function POST(req) {
  const update = await req.json();
  const msg = extractMessage(update);              // the customer's actual words
  const reply = await claudeWithTools(msg, {       // Claude is the brain
    lookupOrder, findProducts, checkStock, policy  // tools that read your live Woo data
  });
  await sendWhatsApp(msg.from, reply);             // straight back to the customer
}

The important detail lives in that third argument. Claude is not guessing. Following the model's tool use, it is handed a set of functions that read your real WooCommerce data, order status, product details, stock levels, your shipping and returns policy, and it calls one whenever a question needs a fact. So when a customer asks "did my order ship," the assistant looks up the actual order and tells them, instead of offering a menu. The webhook is verified so only WhatsApp can trigger it, every secret lives in a server-side environment variable rather than in the browser, and the whole backend deploys the moment you push, which is the same webhook discipline we apply to Revolut on Vercel. The deeper mechanics of running a WhatsApp integration on the Cloud API, templates, opt-in, and delivery, are collected in our WhatsApp automation on the Cloud API guide.

What can a store-aware assistant finally do?

This is where the move pays for itself, because the assistant stops performing helpfulness and starts being helpful.

Presale questions answered with real knowledge. Which size fits, whether two products are compatible, what is in stock, what shipping costs to a given country, what the price is for a specific variation. These are the questions that decide whether a hesitant visitor becomes a buyer, and they get answered from your catalog, correctly, in the moment the customer is ready.
Support that resolves rather than stalls. Order status and tracking, returns and exchanges, address corrections, all grounded in the live order in your database, so the customer gets a real answer instead of a holding pattern.
A voice that fits your brand and your customers. On-brand tone, multiple languages, and the judgment to say "let me get a person for this" and hand off cleanly when a question is sensitive or out of scope, which is the line between an assistant that builds trust and one that quietly erodes it.

The thinking behind grounding an assistant in your own records, rather than letting it improvise, is the same case we make in giving your site an AI assistant that knows your business.

What does it cost compared to a per-contact plan?

A fake percentage would not help you, so here is the real shape of it. With a per-contact platform, your cost is tied to the size of your audience, and it climbs as your list grows whether or not those people are messaging you. A custom assistant changes what you pay for. You pay WhatsApp's own per-conversation fee to Meta, a small amount to host the backend on Vercel, and the model usage you actually choose to run through Claude. None of those scale with the raw size of your contact list the way a per-contact plan does.

So the saving is not a single number, it is a trajectory. The larger your audience and the more you lean on messaging, the wider the gap grows between a bill that follows your list and one that follows real usage you control. For a small store sending a handful of messages, the difference is modest. For a store where WhatsApp is becoming a genuine sales and support channel, which is precisely the store that outgrows ManyChat, it compounds.

How do you migrate without breaking what works?

You do not flip a switch and hope. The move is a sequence of small, safe steps.

Map what you have. We review your current ManyChat flows, list what each one does and which ones actually earn their keep, and identify the questions your customers ask most.
Set up the official channel. Your WhatsApp Business account and number on Meta's platform, owned by you, so the conversation belongs to your store from day one.
Build the backend behind a fixed-scope first phase. That first phase is usually one or two high-value jobs done properly, for example order-status answers and your top presale questions, grounded in live Woo data, with a clean handoff to a human for anything beyond it.
Prove it, then widen it. You watch it handle real questions before it ever faces a customer, and once it is solid we layer on more: returns, broadcasts, deeper product help, each shown to you before it ships.
Keep marketing clean. Outbound campaigns still work, through WhatsApp's approved message templates and proper opt-in, so you stay compliant and your number stays trusted.

A readiness checklist before you switch

Is WhatsApp already carrying real presale or support volume, not just the occasional message?
Can you name the three or four questions customers ask most often on it?
Is your ManyChat bill rising with your contact list rather than with actual conversations?
Do you have a WooCommerce store whose orders, stock, and prices are the source of truth?
Are you comfortable owning a WhatsApp Business account and a Vercel project under your own name?
Do you have a clear rule for when a conversation should hand off to a human?
Is there one high-value job (order status, top presale questions) worth proving first?

Common pitfalls

Treating the assistant as a place to dump every flow at once. The instinct after leaving ManyChat is to rebuild everything the old tool did on day one. That turns a clean first phase into a sprawling one and delays the proof that the core actually works. Ship the one or two jobs that carry the most weight, watch them behave, then widen.

Forgetting that WhatsApp has its own rules. Outbound messages outside a live conversation window must use approved templates and proper opt-in, and a number that ignores that gets throttled or flagged. A custom build does not exempt you from Meta's policy; it just gives you a clean place to respect it.

Letting the assistant guess instead of look up. A model with no tools will happily invent a shipping date. The entire value here is that it does not: every factual answer comes from a tool that read your live data, and a returned "I need to check that with a person" is better than a confident wrong answer.

A concrete case. A specialty retailer we looked at had grown WhatsApp into a real presale channel and was paying a per-contact plan that climbed every month as their list grew, even though most contacts never messaged. Worse, the flow answered "is this in stock" with a generic menu, so customers with cash in hand drifted off to email and waited a day for a human. We moved them in stages: first a backend on their own Vercel project that answered stock and order-status questions from live WooCommerce data with Claude, handing off cleanly on anything sensitive, then broadcasts through approved templates once the core was solid. The wrong-price and dead-menu moments stopped because every answer now came from the actual catalog, the monthly bill decoupled from their contact count, and the whole conversation finally belonged to them rather than to a platform in the middle. The lesson was not "custom beats ManyChat everywhere." It was that a store selling on WhatsApp eventually needs an assistant that can see the store.

How we build it so you can trust the result

A channel that talks to your customers in your name deserves a careful process, so ours keeps you in control the whole way. We start with discovery and a clear roadmap, fix the first phase in scope and price up front, and show you demos before each next step, with feedback that turns into changes quickly because you talk to the developer building it rather than a queue. Above all, you own everything: your WhatsApp account, your Vercel project, your keys, and your WooCommerce data, with no lock-in, which is the entire point of leaving a rented platform. This is the everyday substance of our work, not a theory: we run real production backends on Vercel, including the customs-invoice.com compliance platform, and we ship full storefronts with live payments such as LeO-Optic, with the WooCommerce craft behind it done as WooSmiths. The judgment about when a store should stop stacking plugins and build something purpose-fit is the same one in when a custom WooCommerce build is the right call. If your store has outgrown ManyChat and you want a WhatsApp assistant that truly knows your catalog and your orders, runs on your own Vercel project, and thinks with Claude rather than a script, tell me how you use WhatsApp today and I will outline the cleanest first phase to get there.

FAQ

When should I move off ManyChat?

Move when your WhatsApp conversations start touching things a script cannot see: live stock, a specific order's status, the right price for a variation, or what a returning customer bought before. Those are the moments a flow answers with a menu and a buyer drifts away. A second signal is the bill, if your per-contact plan is climbing with your list size rather than with real conversations, you are paying more precisely as the tool delivers less. When both are true, the channel has become valuable enough to deserve an assistant that actually knows your store.

Do I have to leave WhatsApp to build a custom assistant?

No, and you would not want to. WhatsApp is the channel, not the middleman. Messages always travel over Meta's official WhatsApp Business Platform, exactly as email travels over a mail provider, and that stays. The piece you drop is the intermediary bot service, the vendor that relayed your traffic and stored your automation while charging by headcount. In a custom build, your own backend connects straight to WhatsApp's API and reads directly from WooCommerce, so nobody is metering your contacts or standing in the conversation.

Will a custom WhatsApp assistant be cheaper than ManyChat?

It depends on your scale, and the honest answer is a trajectory rather than a fixed percentage. A per-contact plan charges more as your audience grows, whether or not those contacts message you. A custom assistant charges by WhatsApp's own per-conversation fee, a small hosting cost on Vercel, and the model usage you choose, none of which track your raw list size. A shop with light messaging volume sees only a slim gap. A shop where WhatsApp is turning into a real sales channel sees a list-sized bill and a usage-sized one pull apart steadily.

How does the assistant actually know my orders and stock?

Through tools. The model does the reasoning, but it never queries your database directly. It is given named functions, one for order status, one for product lookup, one for stock, one for policy, and it invokes whichever the question calls for. So a "where is my order" question triggers a real lookup against the actual order rather than a canned reply, and an "is this in stock" question reads current inventory. The model decides what to say; the functions make sure what it says is true, which is the difference between an assistant that resolves problems and one that invents answers.

Can it still do marketing broadcasts?

Yes. Outbound campaigns keep working through WhatsApp's approved message templates and proper opt-in, the same rules any compliant sender follows. The difference is that the broadcast now runs from your own backend rather than a rented platform, so your number and your list stay under your control. We typically layer broadcasts on after the support and presale core is solid, so the assistant proves it can answer real questions correctly before it starts reaching out, which is also how a number keeps the trust that keeps it deliverable.

How long does the migration take, and is it risky?

It is deliberately staged so the risk stays low. We map your existing flows, stand up your own WhatsApp Business account and number, then build a fixed-scope first phase covering one or two high-value jobs, usually order status and your top presale questions. You review it against live questions on a staging setup, and only after it proves reliable do we extend it to returns, deeper product help, and broadcasts, each shown to you first. Because the old setup keeps running until the new one is proven, nothing customer-facing goes dark during the move, and you own every piece of it as it lands.

From ManyChat to a Custom WhatsApp Assistant for WooCommerce, with Claude as the Brain