How does a restaurant train the intusell AI? A step-by-step guide

When the evening service heats up, the phone does not stop: "Do you have a table for 4 tonight?", "Are there gluten-free options?", "How much is takeaway, and how long does it take?", "Do you have parking?". With the kitchen slammed, someone has to watch the phone, Instagram DMs, and WhatsApp all at once. Most of the time, they cannot; the message goes unanswered, and the customer books with the competitor down the street.

intusell takes over this traffic. But to do so, it first needs to learn about your restaurant, your menu, and your way of talking. This article explains how a restaurant or cafe trains the intusell AI from scratch, step by step. This is the restaurant chapter of our sector-based "how to train your AI" series, and it is the cornerstone article of the series.

Quick answer

Why does training matter?

intusell is not a canned chatbot; it is a fully autonomous AI sales assistant that behaves like a senior salesperson. Good front-of-house staff do not start on day one without knowing the menu by heart, either. Training enables the AI to do two things correctly: give the right information (your menu, real prices, allergens) and give it the right way (your restaurant's tone, your way of welcoming guests).

An untrained assistant either gives overly generic answers or guesses about something it does not know. In a restaurant, guessing is expensive: wrong allergen information is a health risk, and a made-up price creates distrust in the customer who sits down at the table. A well-trained assistant stops when it does not know and connects to the team. This is the strongest trust message you can build.

Who is it for?

This guide is for restaurants and cafes whose incoming message traffic is getting busy and who answer the same questions over and over again:

• À la carte restaurants, cafes, bistros, brunch spots, pizzerias, and takeaway businesses
• Teams receiving dozens of reservation and order questions a day via Instagram and WhatsApp
• Businesses in tourist areas that communicate in multiple languages with foreign guests
• Operators who do not want to pull someone away from the kitchen to answer the phone during the rush

You do not need a technical team. All of the steps below are completed from sections in the panel, without writing any code.

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

Dividing the training into two groups makes things easier. The first 5 core steps get the assistant operational; these steps are the part of the setup that is done once. The following 2 continuous steps (label review and opening the channels) are the ongoing part, where you sharpen the assistant as you use it. You will find all of them in order below.

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 restaurant's textual information. Every time the AI generates a response, it automatically searches these sources and uses only the information written there.

Supported source types:

Every file you upload is automatically chunked and made searchable via pgvector. When you upload your menu, and a customer asks "Is there a vegan main course?", the AI finds the right section and responds.

What must be uploaded in a restaurant: the menu and prices, allergen and ingredient information, opening hours and address, reservation and takeaway conditions, venue details such as parking/children/pets, and the most frequently asked questions. Once you upload these six headings, a large part of your incoming messages already becomes answered automatically.

2. Connect the menu and prices

In the first step you uploaded the general texts to the knowledge base; in this step you focus on keeping your actual product backbone — the current menu and prices — accurate. In a restaurant, there is no external catalog integration like in tourism; your menu is the knowledge base itself. That is why keeping a single "current menu" source is critical.

The practical setup looks like this:

• Upload the current menu as a single PDF or Excel file; remove old/conflicting files.
• Collect seasonal or daily-changing prices in a separate short file so they are easy to update.
• Write conditions such as the takeaway surcharge, minimum basket amount, and delivery area in the same place.

The critical point here is this: the AI never invents a price. It only answers from the menu you upload. When asked about an item not on the menu or a price it is unsure of, it does not estimate; it says it does not know and offers to connect to the team. Likewise, stock is not real-time: you tell the AI about the items that have sold out for the day (as a short note or a response rule), and it then avoids recommending that item and offers an alternative.

When the menu changes, the only thing you need to do is add the current file to the knowledge base and remove the old one; the AI uses the new price in its next response. You can see all the capabilities of the knowledge base on the Solutions page.

3. Set the persona and tone

You taught the AI what to say in the first two steps; now you will determine how it says it. There are two settings in the panel:

• Assistant name (ai_persona_name): The name it will introduce itself to the customer with. Most businesses use a warm name aligned with the brand or a real host's name.
• Tone (ai_tone): Whether it speaks warmly and casually, in a sophisticated and calm manner, or quickly and practically.

The tone of a family restaurant and a fine-dining venue will not be the same. The former might be warm, witty, and inviting; the latter measured, polite, and selective. A neighborhood cafe uses a relaxed, everyday language. This setting is reflected in the AI's messaging language on every channel.

A note on multiple languages: The AI responds in whatever language the customer writes in. If you are in a tourist area, when a foreign guest writes in English, you do not do anything separate; the assistant automatically detects the language and replies in that language while preserving the tone.

4. Define the response rules

The persona defines "who it is," and the response rules define "how it behaves." This is where you transfer your guest-welcoming discipline to the AI. With the AI Manager Chat, you add rules by writing in plain language, just as you would instruct a teammate.

Typical rules for restaurants:

• "On a reservation request, first ask the party size, date, and time; then check availability."
• "When asked about an allergy, read the allergen information from the menu clearly; never guess 'it does not contain it'; if unsure, say the kitchen will be consulted."
• "On a takeaway request, state the minimum basket amount and the delivery area."
• "If asked about an item not on the menu or a price you are unsure of, do not make it up; connect to the team."

Guessing on an allergen question is the most dangerous mistake you can make in a restaurant. The "consult the kitchen if unsure" rule turns the AI from a machine that makes something up for every question into a genuinely reliable host.

If you want to compare two different approaches, you can use the A/B testing feature: for example, you can place a shorter, faster greeting next to a warmer, more conversational one at 50%-50% traffic and measure which one brings in more reservations.

5. Teach it from past calls

This is the step that takes the training from "good" to "specific to your restaurant." In the panel, you upload the voice recordings of your past phone calls (MP3, MP4, WAV, M4A) and mark each one as Won or Lost. In a restaurant, "won" means a call that turned into a reservation or an order.

The system uses these recordings in two ways:

In other words, the AI learns how your best host turns a "we are full tonight" situation into an alternative time slot, how they recommend to an undecided customer, and applies the same approach in similar situations. The KVKK side is protected: PII masking and explicit consent are applied to uploaded recordings.

This step is not mandatory, but do not skip it. The knowledge base teaches the AI "what" it knows; the call recordings teach it "how you welcome guests and sell." When the two are combined, the assistant truly resembles a senior team member.

6. Refine via label review (continuous step)

The first five steps get the AI operational. From there on, it is the continuous part that perfects it over time. Every AI response drops into a label review queue. Here you can do three things: approve, reject, or correct.

The AI learns from these corrections. Say it quoted a dessert at an old amount even though it is on the current list; you correct it, and from then on it uses the right price in similar situations. Or it relayed a promotion incorrectly; you correct it, and it learns your standard wording. Over time, patterns specific to your brand (the phrases you use, campaign names, standard greeting lines) accumulate.

For the first two weeks, we recommend looking at this queue for 10-15 minutes a day. During this time, the correction rate drops quickly, because the AI learns the frequently made mistakes. Review is the "live" part of training: the system gets smarter as it is used.

7. Open the channels (continuous step)

Once the training is ready, you put the assistant in front of customers. intusell brings all channels together in a single inbox: WhatsApp, Instagram DM, Instagram comments, Facebook Messenger, Telegram, web chat, and email.

Channel activation methods:

• Meta channels (Instagram, Messenger): connected 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.

We recommend starting with WhatsApp and Instagram DM first; these are Meta-approved and ready to use immediately. Instagram comment automation, on the other hand, opens gradually, subject to Meta approval. Thanks to channel gates, you control each channel separately: the Instagram DM gate (instagram_dm_enabled) and the Instagram comment gate (instagram_comments_enabled) are separate; you can first automate only DMs and keep comments manual, then activate comments when that gate opens.

For human handoff, there are lock modes: ai_only (the AI answers everything), human_only (everything goes to the team), hybrid (the AI normally answers and escalates to the team when needed). Most restaurants start with hybrid. For details on Instagram and WhatsApp automation: Instagram and WhatsApp automation.

There is a protection layer, but it is not a "blocker"

In restaurant content, allergen and health-related questions are sensitive. In intusell, the real control is the restaurant-specific system prompt that constrains the AI so it does not produce wrong information in the first place: the assistant does not claim an ingredient is "absent" if it is not written on the menu. On top of that, a protection layer scans responses and flags risky phrasing. The default behavior (shadow mode) is to detect and flag; it does not impose a hard block. So the correct framing is this: the AI is constrained, and the protection layer scans and flags — but the real weight of safety is carried by your response rules and the accuracy of your menu information.

How long does training take?

A working setup takes half a day:

1. Uploading the first files to the menu and knowledge base: 1-2 hours (shorter if your menu is digitally ready)
2. Separating out the current price file: a few minutes
3. Persona, tone, and first response rules: 30 minutes
4. Connecting the channels: 1-5 minutes per channel

But there is no moment where "training is done." Over the first two weeks, as you approve and correct responses in the label review queue, the assistant sharpens to fit your restaurant. Uploading past call recordings and running A/B tests are also ongoing improvements. 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 static menu bot. It is not a tool that shows a few pre-written answers in sequence; it is an assistant that understands every question asked and generates a dynamic response from the knowledge base.
• It is not a chatbot tied to a rule tree. There are no "press 1, press 2" menus. The customer writes freely, and the AI understands the intent.
• It is not a real-time stock/POS system. You tell it about sold-out items and price changes; it does not make things up, and it asks when it does not know.
• It does not keep a managed waitlist. It does not run a queue like "I have put you in line and will call when a spot opens." Instead, it brings an opportunistic re-offer to eligible WhatsApp customers in follow-up at the right moment (next section).

In short: not a bot that gives canned answers, but a welcoming and sales assistant that represents your restaurant to the extent that you train it.

After training: reservations, reminders, and proactive follow-up

After you train the assistant, two powerful modules come into play. The first is the appointment engine: you define your opening hours, table capacity, and reservation statuses; the AI asks party size and time on WhatsApp or Instagram and creates a reservation for an available slot. Automatic reminders (for example, 1 day and 2 hours before) reduce the no-show rate; the customer manages their own reservation via the /manage-appointment/{token} link, cancelling or rescheduling if needed. Alongside table reservations, this engine also handles takeaway order routing with the same logic. Details: restaurant reservation automation.

The second is proactive follow-up (CRM): the system automatically re-engages a customer who has not visited for a while or who left a conversation half-finished — for example, with a "special offer for the weekend menu" message. This is not a waitlist; it is an opportunistic re-offer to eligible WhatsApp customers in follow-up. You will find how to run the assistant you trained in daily service in the usage article of the series: how a restaurant uses intusell.

Frequently asked questions

How long does restaurant AI training take?

A working setup is done within half a day: uploading the menu and prices to the knowledge base, setting the persona and tone, and a few response rules. The real refinement accumulates over the first weeks as you approve and correct responses in the label review queue. Training is not a one-time event; it is continuous.

Do I need technical knowledge to train the AI?

No. All training is done from the panel; no code, API key, or developer is required. You upload your menu to the knowledge base as a PDF/Excel file, choose a persona, and write response rules in plain language. You connect channels with one-click OAuth or a QR code for WhatsApp.

Will the AI give a wrong price or mention a dish that does not exist?

No. The AI only answers from the menu and prices you upload to the knowledge base. When asked about an item not on the menu or a price it is unsure of, it does not make it up; it says it does not know and hands off to the team. Not inventing prices or stock is the most important trust rule.

What happens for an item that has run out?

You enter the items that have sold out for the day as a response rule or a short knowledge-base note; the AI will not recommend that item and offers an alternative instead. The AI is not a real-time stock system, so you tell it about sold-out items. When unsure, it does not guess; it asks the team.

Will the AI learn the new menu when the menu changes?

Yes. When you add the current menu to the knowledge base as a new file, the AI uses the new prices and items in its next response. Removing the old file is enough to avoid any conflicting information.

Does it reply to Instagram comments automatically as well?

WhatsApp and Instagram DM automation are opened first (Meta-approved). Instagram comment automation opens gradually, subject to Meta approval; the DM gate and the comment gate are separate settings in the panel. Even before comments are ready, you run fully on DM, WhatsApp, and web chat.

Next step

You have trained your assistant; next comes using it in daily service. The next article in the series explains how to run the intusell you trained during a real evening service: inbox management, the reservation flow, takeaway order routing, and team handoff. Continue straight from there: how a restaurant uses intusell.

You can read this series for other sectors too: you will find a similar setup in automotive AI training for automotive, tour agency AI training for tourism, and clinic AI training for health.

If you would like to see it live before starting the setup, let's open your panel together in a 20-minute session via Request a demo, or write to hello@intusell.com. You can find all the capabilities of the restaurant solution on the restaurant solution page, package and quota details on the pricing page, and other guides in the all articles list.