How does an e-commerce brand train the inTusell AI? A step-by-step guide

An e-commerce brand's inbox is always full of the same questions: "When does shipping arrive?", "Will this size fit me?", "How do returns work?", "Is it in stock?", "Where is my order?". Dozens of comments pile up under a single product, dozens of size questions land in DMs, dozens of tracking requests hit email. The team drops its real job — closing a new customer — to churn out copy-paste answers all day. A customer writing from abroad waits for hours. The "I'll buy right now, is there a discount?" message that arrives at 11 PM waits until morning, and by then the customer has bought elsewhere.

inTusell takes over this load. To do that, it first has to learn your brand, your products, and your sales style. This article explains step by step how an e-commerce brand trains the inTusell AI from scratch. It is the e-commerce chapter of our sector-based "how to train your AI" series, and the main article of that series.

Quick answer

Why does training matter?

inTusell is not an off-the-shelf chatbot. It is a fully autonomous AI sales assistant that behaves like an experienced online sales advisor. An experienced salesperson cannot start on day one without knowing your store either. They need to know which products you sell, your shipping and return routine, and which questions to leave to the team. Training makes the AI do two things correctly: give the right information (product details, shipping times, return terms) and give it the right way (your brand voice, tone, and sales discipline).

An untrained assistant either gives overly generic answers or guesses on a topic it doesn't know. In e-commerce, guessing is costly. A wrong shipping time brings complaints, made-up stock means cancellations and returns, and a wrong price loses the customer. A well-trained assistant stops when it doesn't know, says so honestly, and connects the customer to the live team — talking in the right places and staying quiet in the right places.

Who is it for?

This guide is for e-commerce teams whose inbox is dominated by product and order questions and who answer the same things over and over:

• Brands selling through their own site, Shopify, WooCommerce, or a marketplace
• Fashion, cosmetics, accessories, and home-goods sellers who get dozens of DMs and comments a day from Instagram and WhatsApp
• Everyone overwhelmed by message volume, from a one-person team to a 10-person support department
• Brands selling abroad that need to communicate in several languages
• Anyone who wants to re-engage customers who stall at checkout or say "let me think" and disappear

You don't need a technical team. You complete every step below from sections in the panel, without writing code.

The structure of training: 5 core steps + 2 continuous steps

Splitting the training into two groups makes your job easier. The first 5 core steps get the assistant up and running; you do them once. The 2 continuous steps that follow — label review and opening channels — are the ongoing part you keep sharpening as you use the assistant. Here they all are, in order.

1. Upload the knowledge base

The foundation of training is the knowledge base (RAG). In the panel's Knowledge Base section, you upload all of your brand's written information. Each time the AI generates a response, it searches these sources and uses only the information written there. This is the technical answer to the "don't let the AI make things up" worry: the AI talks about whatever you hand it, and nothing else.

Supported source types:

Every file you upload is chunked automatically and made searchable with pgvector. Upload a size-chart CSV, and when a customer asks "Which size for 1.70 m and 65 kg?", the AI finds the right row and responds.

The must-haves for an e-commerce brand: product descriptions and specifications, shipping times and fees, return and exchange terms, size charts, payment methods (cash on delivery, installments), campaign rules, and the most frequently asked questions. Once these are in, a large part of your inbox answers itself.

One important note: if your catalog changes often — new products, prices, stock — keeping the knowledge base current is up to you. Upload the updated file, and the AI uses the new information in its next response. It won't present a price from an old file as current: it knows whatever you give it, and it says when it doesn't know.

2. Set the persona and tone

You taught the AI what to say in the first step. Now you define how it says it. There are two settings in the panel:

• Assistant name (ai_persona_name): The name it uses to introduce itself to the customer. Most brands use a real advisor's name from the team, or a name that fits the brand.
• Tone (ai_tone): Whether it speaks warm and friendly, corporate and crisp, or fast and energetic.

A young streetwear brand and a boutique jewelry brand won't share a tone. The first might speak casually, with humor; the second, elegantly and formally. This setting shapes the AI's message style on every channel, so the tone in an Instagram DM matches the tone in an email.

A note on languages: the AI responds in whatever language the customer writes in. When a customer from abroad writes in English or German, you do nothing extra — the assistant detects the language and responds in it while keeping your tone.

3. Define the response rules

The persona defines who the assistant is; the response rules define how it behaves. This is where you pass your sales discipline on to the AI. With the AI Manager Chat, you add rules by writing in plain language, just as you would brief a teammate.

Typical rules for e-commerce brands:

• "If you're unsure about price, stock, or shipping information, don't make it up; say you don't know and hand the conversation off to the team."
• "If the customer is undecided, suggest a size/product, but don't apply pressure."
• "When asked about returns/exchanges, read the policy from the knowledge base clearly; add no commentary."
• "Share a discount code only for defined campaigns; don't invent a discount on your own."
• "If the customer gives an order number, direct them to the team for tracking." (Order status is kept in a live system.)

The most common mistake in e-commerce is an AI guessing just to say something, and that guess turning into a cancellation, a return, or a bad review. In inTusell, this boundary has two layers. The primary control is the system prompt given to the AI: from the start, it is constrained so that it won't step outside the knowledge base or invent price and stock. On top of that, a safety layer scans responses and flags risky phrasing. This layer runs in shadow mode by default: it detects and flags, but it does not hard-block the response. The system's reliability, in other words, comes from constraining the AI correctly from the start — not from a wall that blocks everything.

To compare two approaches, use the A/B test feature. Put a softer suggestion next to a more direct closing line at 50%-50% traffic and measure which one converts to more sales.

4. Teach from past sales conversations

This is the step that moves training from good to specific to your brand. In the panel, you upload audio recordings of your past sales conversations (MP3, MP4, WAV, M4A) and mark each one Won or Lost — did the customer complete the order, or leave it in the cart?

The system uses these recordings in two ways:

From these, the AI learns how your best salesperson explained value to someone who said "it's too expensive" and moved a "let me think about it" toward an order. The KVKK side is protected: uploaded recordings get PII (personal data) masking and explicit consent.

This step is not mandatory, but don't skip it. The knowledge base teaches the AI what it knows; past conversations teach it how you sell. Put the two together and the assistant really does resemble an experienced online sales advisor.

5. Correct with label review (continuous step)

The first four steps get the AI up and running. What follows is the continuous part that perfects it over time. Every AI response lands in a label review queue, where you can do three things: approve, reject, or correct.

The AI learns from these corrections. Say it explained a shipping time incompletely, or called a product by the wrong name. You correct it, and from then on it uses the right wording in similar cases. Over time, patterns specific to your brand build up: the product names you use, your standard campaign sentences, the way you steer people through a return.

For the first two weeks, we recommend spending 10-15 minutes a day in this queue. The correction rate drops quickly, because the AI learns the mistakes it makes most often. Review is the live part of training: the system gets smarter as you use it.

6. Open the channels (continuous step)

Once training is ready, you put the assistant in front of customers. inTusell gathers every channel in a single inbox: WhatsApp, Instagram DM, Instagram comments, Facebook Messenger, Telegram, web chat, and email. Whichever channel a customer writes from, the same trained assistant responds with the same information.

How you open each channel:

• Meta channels (Instagram, Messenger): connect with one-click OAuth.
• WhatsApp: connects in about 1 minute by scanning a QR code from your phone — no Meta Business approval required. The official Cloud API option is also available.

As a first step, we recommend opening WhatsApp and Instagram DM, which are approved and usable right away. Instagram comment automation — the "price?", "in stock?" comments under product posts — opens gradually, subject to Meta approval. The channel gates are separate: the DM gate (instagram_dm_enabled) is independent of the comments gate (instagram_comments_enabled), so you can automate only DMs first and open comments once approval comes through.

Handing off to a human has three working modes: ai_only (the AI answers everything), human_only (everything goes to the live team) and hybrid (the AI normally answers and escalates when needed). Most brands start with hybrid. The AI handles recurring product and order traffic, and sensitive cases go to the team. One note: a file, image, audio, or video attachment arriving is not on its own a reason to hand off. The AI tries to make sense of the attachment and hands off only when it can't complete the task.

Once the channels are open, CRM and proactive follow-up kick in. For eligible WhatsApp customers in follow-up who said "let me think" and disappeared, or who stalled at checkout, the assistant can bring a well-timed re-offer — the part of e-commerce that leaves the most money on the table and gets the least attention. A separate article covers all the details of daily operations, from incoming message to sale, follow-up, and handoff: how an e-commerce brand uses inTusell.

How long does training take?

A working setup takes half a day:

1. Uploading the first files to the knowledge base: 1-2 hours (shorter if your existing documents are ready)
2. Persona, tone, and first response rules: 30 minutes
3. Connecting channels: 1-5 minutes per channel
4. First test conversations: 30 minutes

But there is no moment when training is done. Over the first two weeks, the responses you approve and correct in the label review queue sharpen the assistant to fit your brand. Setup is fast; mastery is continuous.

What it isn't

Putting inTusell in the right category matters, because the wrong expectation leads to the wrong setup.

• It is not a chatbot that gives canned answers. It is not a decision tree but an autonomous sales assistant that represents your brand as well as you train it.
• It is not a tool that fabricates price or stock. It uses only the information you upload. It says when it is unsure and hands off to the team. This is the most important trust message.
• It is not an e-commerce platform (Shopify/WooCommerce) or a payment system. It does not replace your store. You always close the sale and issue the invoice, and inTusell is not a party to the payment.
• It is not a managed "waitlist" service. It does not send a bulk notification when stock arrives. It is an assistant that brings a well-timed re-offer to eligible customers in follow-up.

In short: it is not a bot that spits out an answer to a single question. It is a sales assistant that represents your brand as well as you train it — and never fabricates what it doesn't know.

Frequently asked questions

How long does e-commerce AI training take?

You can complete a working setup within half a day: uploading the knowledge base, setting the persona and tone, writing a few response rules, and connecting channels. The real refinement builds up over the first weeks as you approve and correct responses in the label review queue. Training is not a one-time job; it is continuous.

Do I need technical knowledge to train the AI?

No. You do all the training from the panel; no code, API key, or developer is required. You upload PDF, Excel or CSV files to the knowledge base, add your web URLs, choose the persona, and write response rules in plain language. You connect channels with one-click OAuth or by scanning a QR code.

Does the AI give wrong or made-up product/stock/price information?

No. The AI uses only the information you upload to the knowledge base, and it never invents a price, stock status, or shipping detail it is unsure about. When it does not know, it says so honestly and hands the conversation off to the team. This is the behavior that earns the most trust in e-commerce.

Which channels can I connect?

WhatsApp, Instagram DM, Instagram comments, Facebook Messenger, Telegram, web chat, and email are gathered in a single inbox. Meta channels connect with one-click OAuth; WhatsApp connects in about 1 minute by scanning a QR code from your phone. Instagram comment automation opens gradually, subject to Meta approval.

Can the AI answer order, shipping, and return questions?

Yes. Once you upload shipping times, return and exchange terms, the size chart, and product details to the knowledge base, the AI answers these recurring questions automatically. When a case falls outside the policy, or it is unsure, it adds no commentary; it hands off to the team. For details, see automating order, shipping, return, and product questions.

Is my data safe under KVKK/GDPR?

Yes. inTusell is end-to-end encrypted and KVKK/GDPR-compliant. Uploaded audio recordings get PII (personal data) masking and explicit consent, and customer data is kept isolated per tenant.

Next step

You've trained your assistant; next comes using it in daily operations. The next article in the series explains how to run the inTusell you trained on a real workday: incoming product questions, cart recovery, shipping tracking, returns, and handing off to the team. Continue straight from there with how an e-commerce brand uses inTusell. For the details of channel setup, the Instagram and WhatsApp sales automation article is ready too.

We wrote this guide for e-commerce, but the same 5-step logic works in every sector. Take a look at the main articles for neighboring sectors: AI training for real estate, AI training for tour agencies, and AI training for clinics. If you'd like to see a setup tailored to your brand live, use Request a demo and we'll open your panel together in a 20-minute session, or write to hello@intusell.com. The e-commerce solution and solutions pages hold the full set of e-commerce capabilities, the pricing page covers packages and quotas, and the all articles list holds the other guides.