---
title: "How to set up AI: The definitive guide"
description: "Fuely AI is Chatfuel's AI agent designed to help service businesses respond instantly, book appointments, qualify leads, and route conversations to the right…"
canonical_url: https://chatfuel.com/docs/ai-features/how-to-set-up-ai-the-definitive-guide
markdown_url: https://chatfuel.com/docs/ai-features/how-to-set-up-ai-the-definitive-guide.md
last_updated: 2026-09-08
lang: en
site: https://chatfuel.com/docs
llms_txt: https://chatfuel.com/llms.txt
---

# How to set up AI: The definitive guide

Fuely AI is Chatfuel's AI agent designed to help service businesses respond instantly, book appointments, qualify leads, and route conversations to the right…

## 1. What is Fuely AI (and why it matters) [#1-what-is-fuely-ai-and-why-it-matters]

Fuely AI is Chatfuel's AI agent designed to help service businesses respond instantly, [book appointments](https://chatfuel.com/docs/calendar/calendar), [qualify leads](https://chatfuel.com/docs/leads/leads), and route conversations to the right place. Instead of missing customer messages or relying on slow manual replies, Fuely AI acts as a smart assistant that engages prospects the moment they reach out.

The AI works from structured business data — company details, catalog of services or products, FAQ entries — plus the instructions you write that define how it should behave. For small and medium service providers, this means fewer missed opportunities, faster bookings, and higher-quality leads that arrive pre-qualified before a human ever joins the conversation.

Imagine a local beauty salon that normally struggles to keep up with WhatsApp inquiries. With Fuely AI, every new customer who writes in instantly receives a friendly greeting, can browse the salon's price list, and select an appointment time from available slots. If the client has a unique request — like needing a stylist with specific expertise — the AI gathers the details and seamlessly [passes the conversation to a team member](https://chatfuel.com/docs/ai-features/how-to-stop-ai-during-active-conversations). The result is a smoother customer journey, less workload for staff, and more confirmed bookings without the back-and-forth delays.

## 2. Two equal ways to do everything in this guide [#2-two-equal-ways-to-do-everything-in-this-guide]

You can complete every step below in two ways: in the dashboard, or by asking the [AI Co-Worker](https://chatfuel.com/docs/ai-features/what-the-ai-co-worker-can-do-for-you) in chat. Both do the same thing — use whichever suits you. The Co-Worker can fill your Knowledge Base, build your catalog, write instructions, set handoff rules, and configure any automation. For example:

* *"Set up my AI assistant for my salon from scratch."*
* *"Add my services and FAQs, and hand chats to my team when someone asks for a human."*

Connecting a channel is one of the few steps the AI Co-Worker can't complete for you — authorization happens in your own Meta or TikTok account. The Co-Worker can take you straight to the right page: just ask. Test conversations are also yours to run — the Co-Worker can open the Test chat for you, but you play the customer.

## 3. How Fuely AI works: Channels, Knowledge Base, and Automations [#3-how-fuely-ai-works-channels-knowledge-base-and-automations]

Before diving into setup, it's helpful to understand the core building blocks of Fuely AI. Think of it as an employee who operates based on specific knowledge, instructions, and rules.

### Channels [#channels]

Fuely AI talks to customers on [WhatsApp](https://chatfuel.com/docs/whatsapp/how-to-activate-ai-in-whatsapp), [Instagram](https://chatfuel.com/docs/instagram/how-to-activate-ai-in-instagram), [Facebook](https://chatfuel.com/docs/facebook/how-to-activate-ai-in-facebook), TikTok, and your [website chat widget](https://chatfuel.com/docs/website/set-up-chat-widget-for-website). You can configure everything before a channel is connected — the setup is saved as a draft and goes live the moment the channel connects.

### Knowledge Base [#knowledge-base]

The [Knowledge Base](https://chatfuel.com/docs/ai-features/how-to-set-up-knowledge-base) is where the AI learns about your business. It has three main sections:

* **About Company** — your business profile (name, contact details, hours, address, payment options, website, short description).
* **Catalog** — a focused list of services or products, each with a name, price, and availability toggle.
* **FAQ** — quick question–answer pairs for common customer queries.

### Automations [#automations]

[Automations](https://chatfuel.com/docs/ai-features/automations) control where the AI talks and how it behaves in each place. At **Fuely AI** → **Automations** you'll find **Main bot** at the top — its settings are the defaults for everything, and its toggle is the master switch for all AI — and below it an automation for every traffic source, grouped by channel: Instagram DMs, post & reel comments, story replies, Facebook Messenger and post comments, WhatsApp ads, TikTok comments, the website widget, and so on.

Every source's automation starts off (only Main bot is on by default), so activating AI on a channel means enabling that source's automation. On comment, ad, story, and link sources you can also add [custom automations](https://chatfuel.com/docs/ai-features/how-to-create-a-custom-automation) — named rules that react only to [specific posts, ads, or keywords](https://chatfuel.com/docs/ai-features/how-to-set-up-automation-for-specific-posts-or-ads). When several rules could answer the same message, [a priority order decides which one replies](https://chatfuel.com/docs/ai-features/which-automation-replies-to-a-message-or-comment).

Each automation carries the same set of settings — message delays, [follow-ups](https://chatfuel.com/docs/ai-features/how-to-set-up-follow-ups), booking rules, catalog images, and more — and every setting is either inherited from Main bot or overridden locally. Set things once on Main bot, then override only where a source needs different behavior.

A simple rule of thumb: &#x2A;*the Knowledge Base is what the AI knows; Automations are where and how it behaves.**

### Additional instructions (the AI's "brain") [#additional-instructions-the-ais-brain]

Every automation has an **Incoming messages** setting: whether the AI replies there, and the **Additional instructions** — the prompt that shapes how it talks. Write instructions on **Main bot** to apply them everywhere, or on a specific automation to shape just that source. This is where you define the AI's role, the steps it should follow, the output format, and example outputs.

> **Note:** If you used Fuely AI before: the separate Profile fields (agent name, chat language, custom greeting, fallback message, message length, emoji frequency) no longer exist. All of that is now written as plain instructions in **Additional instructions*&#x2A; — e.g. &#x2A;"Your name is Sofia. Speak Spanish. Keep messages under 2 lines and use at most one emoji per message."*

### Handoff to humans [#handoff-to-humans]

If a customer explicitly asks for a real person, or if a request falls outside the catalog, Fuely AI routes the conversation to your team. This is controlled by the **Switch to human** setting on each automation. Separately, the **When AI replies** setting decides whether the AI answers at all — always, or only outside your working hours.

In short, Fuely AI works by combining structured business data, clear instructions, and per-source automations so your assistant knows when to answer, when to book, and when to hand over.

## 4. Quick start checklist [#4-quick-start-checklist]

Before you configure Fuely AI in detail, here's a simple checklist to get your first assistant live:

* [ ] Connect your channel — WhatsApp, Instagram, Facebook, TikTok, or the website widget.
* [ ] Fill in "About Company" — add business name, email, phone, hours, address, payment methods, and website.
* [ ] Build your Catalog — add your main services or products (name + price + availability). Keep it under \~50 items to keep the AI focused.
* [ ] Write 10–25 FAQ entries — capture the most common customer questions. Keep answers short (1–3 lines each).
* [ ] Write **Additional instructions** on Main bot — define the assistant's task steps, output format, and example outputs. Use friendly style (short messages, numbered options, clear confirmations).
* [ ] Enable the automations for the sources you want the AI to cover — every source starts off.
* [ ] Configure **Switch to human** rules — decide when the AI should escalate: human requested, sensitive or complex cases.
* [ ] Test in the Test chat — run 3–5 typical scenarios: booking a service, asking a FAQ, requesting a human.
* [ ] Go live and monitor — watch early conversations. Add new FAQs and refine instructions as you go.

## 5. Step-by-step setup [#5-step-by-step-setup]

Now that you have the big picture, let's go through the setup in detail. Each part of the Knowledge Base plays a specific role, and keeping them clean ensures Fuely AI works reliably.

### 5.1 About Company [#51-about-company]

Your official business profile. This is where the AI gets essential company info.

**What to include:**

* Company name
* Email and phone number
* Opening hours
* Address
* Website
* Payment methods
* Short "About" description

**Best practices:**

* Keep this section factual and concise.
* Don't overload it with services, FAQs, or sales copy — that belongs in the Catalog or FAQ.
* Remember: About Company = who you are, not everything you do.

### 5.2 Catalog [#52-catalog]

A structured list of your services or products. Each item has a name, price, [images](https://chatfuel.com/docs/catalog/how-to-attach-photos-to-services-or-products), and availability.

**What to include:**

* Core services (e.g., "Haircut – $30")
* Packages or promotions (optional)
* Only what you actually want customers to book or ask about

**Best practices:**

* Stay focused: under \~50 items for clarity.
* One service = one item (avoid duplicates).
* Never paste your catalog into Additional instructions or About Company — it wastes token space and breaks AI context.

### 5.3 FAQ [#53-faq]

The "ready-to-go answers" section for your most common questions.

**What to include:**

* Shipping/delivery details
* Prices or deposits (if not already in Catalog)
* Booking and cancellation rules
* Location and parking details

**Best practices:**

* Keep answers short (1–3 lines).
* Phrase questions exactly as customers ask them.
* Don't add long explanations; let the AI expand if needed.
* Never copy FAQ into Additional instructions or About Company — duplication makes the assistant inconsistent.

### 5.4 Additional instructions [#54-additional-instructions]

The master instruction set that tells the AI how to behave. This is where you define its role, tasks, output style, and example responses.

**Where to write them:**

1. Go to **Fuely AI** → **Automations**.
2. Open **Main bot** (to shape the AI everywhere) or a specific automation (to shape one source).
3. Open the **Incoming messages** setting and fill in **Additional instructions**.

**What to include:**

* Role: "You are a WhatsApp assistant for a \[business type]."
* Task: Step-by-step flow (greet → collect info → offer options → confirm).
* Output format: WhatsApp-friendly (short, numbered options, emojis optional).
* Example output: A real sample chat message so the AI mirrors the style.

**Best practices:**

* Keep instructions focused on behavior and format, not raw data.
* Don't paste Catalog or FAQ here (they already live in the Knowledge Base).
* Use the Role → Task → Output Format → Example Output skeleton for clarity.
* Write on Main bot first; override on a specific automation only when that source really needs different behavior.

**Together, About Company, Catalog, FAQ, and Additional instructions form the AI's complete context:**

* About Company = your identity
* Catalog = what you offer
* FAQ = quick answers
* Additional instructions = how the AI talks and acts

## 6. Writing prompts that work [#6-writing-prompts-that-work]

Additional instructions are the assistant's brain. A sloppy prompt makes your AI vague, repetitive, or too wordy. A clear prompt makes it sharp, predictable, and aligned with your business.

### 6.1 Bad vs. good prompts [#61-bad-vs-good-prompts]

**Bad prompt:**

```
Answer customer questions and be nice.
```

Problems:

* Too vague → AI may over-explain.
* No structure → inconsistent replies.
* No style → doesn't sound like WhatsApp.

**Good prompt:**

```
### Task
1. Greet the customer warmly.
2. Ask what service they want.
3. Ask when customer wants to come.
4. Confirm the booking.

### Output Format
- Short WhatsApp-style messages (max 2 lines each).
- Always offer numbered options.
- Always confirm before finishing.

#### Example Output
- Welcome to our salon!
- Which service would you like to book today? 💇 Haircut, 💅 Manicure
  or 💆 Spa?
- When do you want to come?
- Got it — you'd like [service] on [date and time]. Do I understand correctly
  that everything is fine, and I can invite the manager to confirm the meeting?
```

Why it works: clear steps, WhatsApp style, sample message.

### 6.2 Multi-step flows & one-shot examples [#62-multi-step-flows--one-shot-examples]

For flows that need reasoning (like lead qualification), break them into steps. Always show the expected answer style with an example output.

**Example: Real Estate Lead Capture**

```
### Role
You are a WhatsApp assistant for a real estate agency.

### Task
1. Greet the lead.
2. Ask which property they're interested in.
3. Ask for budget range.
4. Ask when they'd like to schedule a viewing.
5. Confirm and offer to connect with an agent.

### Output Format
- Each step = one short WhatsApp message.
- Use numbered options when possible.
- Confirm details in the end.

#### Example Output
- 👋 Hi, thanks for reaching out to City Realty!
  Which property are you interested in?
  1) Apartment on Main Street
  2) House in Green Valley
  3) New downtown condos

- (After reply) Great! What's your budget range?
```

### 6.3 Prompt skeleton [#63-prompt-skeleton]

Here's a reusable template for any business:

```
### Task
1. Greet the customer.
2. Collect needed details (service/product, time, preferences).
3. Ask when customer wants to come.
4. Confirm and offer to connect with a manager.

### Output Format
- WhatsApp-style messages (max 2 lines).
- Numbered options for easy replies.
- End every message with the next step.

#### Example Output
- 👋 Hi! Thanks for contacting <Business Name>.
  How can I help today?
  1) 📅 Book a service
  2) 💳 Prices
  3) 👩‍💼 Talk to a team member
```

The key takeaway:

* **Bad prompts = vague and wordy.**
* **Good prompts = task steps, output rules, example output.**

This ensures Fuely AI behaves consistently across every chat.

## 7. Handover to humans: Switch to human [#7-handover-to-humans-switch-to-human]

AI is powerful, but it should never try to handle everything. Fuely AI gives you fine control over when to escalate a conversation and who it should go to. This is managed in the **Switch to human** setting on each automation — set it on Main bot for one shared behavior everywhere, or override it per source.

### 7.1 Rules (when to escalate) [#71-rules-when-to-escalate]

You can define up to 20 rules — the conditions that trigger a handover. Common ones include:

* Customer explicitly requests a real person ("I need to speak with someone," "Can I talk to a manager?").
* Customer's request falls outside AI's scope (refunds, legal/medical/financial issues).
* Customer provides confusing input several times in a row.

### 7.2 Where handed-over chats land [#72-where-handed-over-chats-land]

By default, handed-over chats go to your shared team inbox, where anyone on the [team](https://chatfuel.com/docs/account/how-to-add-teammates) can pick them up. Each rule can instead route the chat to a specific person — for example, sales questions to Seller A, complaints to Support. See [how to set up chat assignment rules](https://chatfuel.com/docs/ai-features/how-to-set-up-chat-assignment-rules).

**Best practice:** For time-sensitive topics, assign a responsible teammate in the rule so the chat has a clear owner instead of waiting in the shared inbox.

### 7.3 Deciding when the AI replies at all [#73-deciding-when-the-ai-replies-at-all]

Separate from handoff rules, the **When AI replies** setting controls whether the AI answers in the first place. It has two options:

* **Always** — the AI replies to every message (default).
* **Outside working hours** — the AI replies only when your business is closed, [based on your business hours](https://chatfuel.com/docs/ai-features/how-to-configure-working-hours); during the workday your team handles chats.

There is no "never reply" option. If you want a channel handled entirely by your team, turn that channel's automation off in **Fuely AI** → **Automations** — a disabled automation stays silent, and its settings are kept for when you switch it back on.

### 7.4 Where prompt writing fits [#74-where-prompt-writing-fits]

* **Don't write prompts inside Switch to human rules.** They are purely for escalation logic. Prompts (Role, Task, Output Format, Example Output) belong in Additional instructions.
* **Why:** Mixing prompts into Switch to human creates confusion. The AI will treat them as escalation conditions rather than behavior guidelines.

Think of it this way:

* **Additional instructions** = how the AI should talk.
* **Switch to human** = when the AI should stop talking.

### 7.5 Best practices recap [#75-best-practices-recap]

* Keep escalation triggers specific ("I want to speak to a person").
* Assign time-sensitive rules to the right teammate; let the shared inbox catch the rest.
* Write prompts only in Additional instructions, not here.

With rules and assignees set correctly, your AI knows exactly when to handle a customer and when to gracefully hand off.

## 8. Testing & iteration: making your assistant smarter [#8-testing--iteration-making-your-assistant-smarter]

Launching your assistant isn't the end — it's the start of a cycle of testing, learning, and improving. Fuely AI makes this simple with the **Test chat** and real conversations.

### 8.1 Use the Test chat before going live [#81-use-the-test-chat-before-going-live]

Every automation has a **Test chat** — a sidebar that [simulates a conversation with exactly that automation's settings](https://chatfuel.com/docs/ai-features/how-to-test-your-automation). Open the automation you want to check (a channel's DMs, a comment rule for a specific post, or Main bot) and write as if you were the customer. This is the safest way to check if:

* Messages are short and clear (WhatsApp style).
* FAQs and Catalog answers show up correctly.
* Handoff rules trigger when they should.
* The source you're testing behaves the way you set it — including any settings you overrode there.

### 8.2 Core flows to test [#82-core-flows-to-test]

1. **Booking a service/product** — Can the AI guide the user step by step? Does it confirm before finishing?
2. **Answering a FAQ** — Does the answer match what you wrote? Is it short (1–3 lines) or too long?
3. **Human request** — Does "I want to talk to a real person" trigger escalation?
4. **Edge cases** — Typos ("haicut" instead of "haircut"), incomplete input ("Tuesday" with no time), repeated vague answers.

### 8.3 Iteration cycle [#83-iteration-cycle]

1. **Test** with the Test chat and real conversations.
2. **Spot issues:**
   * AI says "I'm not sure" → add to FAQ.
   * AI loops → simplify task steps in Additional instructions.
   * AI gives too-long answers → tighten Output Format.
3. **Update the Knowledge Base** (About, Catalog, FAQ).
4. **Refine instructions** (add clear steps, examples).
5. **Re-test.**

This cycle is continuous: every real customer conversation is free training data for improving your assistant.

### 8.4 Best practices [#84-best-practices]

* **Start small:** Launch with your top 5 services + 10 FAQs. Add more only after the basics work.
* **Keep messages human:** If it feels like email, rewrite it.
* **Review handoffs weekly:** Make sure escalation happens at the right moments.
* **Update monthly:** Hours, prices, services, FAQs.

By running quick tests and iterating, you ensure Fuely AI feels less like a "bot" and more like a real team assistant — fast, reliable, and human-friendly.

## 9. Industry templates [#9-industry-templates]

These ready-to-use prompts give you a head start. Just paste them into **Additional instructions** (in the Incoming messages setting on Main bot, or on the automation you're shaping) and adapt details to your business.

### 9.1 Salon / Beauty Booking Assistant [#91-salon--beauty-booking-assistant]

```
### Task
1. Greet the customer.
2. Ask what service they want (haircut, manicure, spa).
3. Ask when they would like to come.
4. Confirm and offer to connect with a manager.

### Output Format
- Short WhatsApp messages (max 2 lines).
- Use numbered options for service selection.
- Always confirm before escalating.

#### Example Output
👋 Hi Anna! Thanks for reaching out to Glamour Salon.
Which service would you like today?
1) 💇 Haircut
2) 💅 Manicure
3) 💆 Spa

(After reply) Great! When would you like to come?
(After reply) Perfect, I'll connect you with our manager to confirm your booking ✅
```

### 9.2 Clinic / Healthcare Assistant [#92-clinic--healthcare-assistant]

```
### Task
1. Collect the patient's name.
2. Ask which doctor or specialty they want.
3. Ask when they'd like to visit.
4. Confirm and connect with a manager.

### Output Format
- Each step = one short WhatsApp message.
- Numbered options for specialties.
- Always confirm before escalating.

#### Example Output
👋 Welcome to City Health Clinic. May I have your full name, please?
(After reply) Thank you, [Name]. Which service do you need?
1) 🦷 Dentist
2) 👩‍⚕️ General Practitioner
3) 👶 Pediatrician

(After reply) When would you like to come in?
(After reply) Got it ✅ I'll connect you with our team to confirm your appointment.
```

### 9.3 Real Estate Lead Assistant [#93-real-estate-lead-assistant]

```
### Task
1. Greet the customer.
2. Ask which property they're interested in.
3. Ask for their budget.
4. Ask when they'd like to view it.
5. Confirm and connect with a manager.

### Output Format
- Short WhatsApp-style steps.
- Numbered options for properties.
- Confirm before escalating.

#### Example Output
👋 Hi! Thanks for contacting City Realty. Which property would you like to see?
1) Apartment in La Candelaria ($500)
2) House in Altavista ($120,000)
3) Studio in Sabana Grande ($350)

(After reply) Great! What's your budget range?
(After reply) When would you like to schedule a viewing?
(After reply) Perfect ✅ I'll connect you with our manager to confirm.
```

### 9.4 Fitness / Gym Assistant [#94-fitness--gym-assistant]

```
### Task
1. Greet the customer.
2. Ask what class or training they're interested in.
3. Ask when they'd like to come.
4. Confirm and connect with a manager.

### Output Format
- Short WhatsApp replies.
- Numbered options for classes.
- Always confirm before escalating.

#### Example Output
💪 Welcome to PowerFit Studio!
Which option interests you today?
1) 📅 Group classes
2) 🏋️ Personal training
3) 🎟 Free trial

(After reply) Great! When would you like to come?
(After reply) Got it ✅ I'll connect you with our team to confirm your spot.
```

### 9.5 Delivery / Dispatch Assistant [#95-delivery--dispatch-assistant]

```
### Task
1. Ask for the pickup address.
2. Ask for the drop-off address.
3. Ask when the delivery is needed.
4. Confirm and connect with a manager.

### Output Format
- Each step = one short WhatsApp message.
- Confirm before escalating.

#### Example Output
🚚 Hi! Thanks for choosing FastGo Delivery.
Please share your pickup address.

(After reply) Got it ✅ Now, what's the drop-off address?
(After reply) When do you need the delivery?
(After reply) Perfect, I'll connect you with our dispatcher to confirm.
```

This way, Fuely AI collects all key details (service/product + time) but never pretends to know schedules or availability. The human manager always does the final confirmation.

## 10. Maintenance: keeping your assistant fresh [#10-maintenance-keeping-your-assistant-fresh]

Launching Fuely AI is just the beginning. To stay effective, the assistant needs regular updates. A business that updates hours, prices, and answers monthly will feel trustworthy. A business that leaves outdated info will quickly frustrate customers.

### 10.1 What to update regularly [#101-what-to-update-regularly]

1. **Catalog** — Add or remove services/products as your business changes. Update prices if they change — AI should never invent them.
2. **FAQ** — Add new questions that come up in chats. Remove irrelevant or outdated answers. Keep answers short (1–3 lines).
3. **Additional instructions** — Review your Role → Task → Output Format → Example Output. Adjust tone, flow, or escalation steps if needed.
4. **Automations** — Pause custom rules for finished campaigns, add rules for new posts or ads, enable new sources as you expand to more channels.

### 10.2 Using real conversations to improve [#102-using-real-conversations-to-improve]

* Review chats weekly.
* Spot where AI says "I'm not sure" → add that question to FAQ.
* Notice long back-and-forth before handoff → simplify the Task in your instructions.
* See repeated typos (e.g., "haicut") → add synonyms or alternate phrasing in FAQ.

### 10.3 Iteration rhythm [#103-iteration-rhythm]

* **Weekly:** Review conversations for gaps. Adjust FAQs, instructions, or escalation rules.
* **Monthly:** Update Catalog. Add new services or remove inactive ones.
* **Quarterly:** Review the entire instruction structure. Check tone of voice consistency. Run test scenarios in the Test chat.

### 10.4 Best practices [#104-best-practices]

* **Start lean:** 10–20 FAQ entries, 5–15 Catalog items. Expand gradually.
* **Avoid duplication:** Never paste FAQ or Catalog into Additional instructions or About Company.
* **Keep it WhatsApp-sized:** Rewrite long explanations into short, friendly messages.
* **Train by example:** If you want a certain style, add an Example Output.

### 10.5 Maintenance mindset [#105-maintenance-mindset]

Think of Fuely AI as a new team member:

* It gets better the more guidance you give it.
* It needs fresh info to stay useful.
* It can't guess — so keep its Knowledge Base clean and updated.

## 11. Troubleshooting: common pitfalls & how to fix them [#11-troubleshooting-common-pitfalls--how-to-fix-them]

Even a well-built assistant may run into issues. Here are the most common ones you'll see, why they happen, and how to solve them.

| Problem                                              | Likely Cause                                                | How to Fix                                                                                  |
| ---------------------------------------------------- | ----------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
| AI doesn't reply on one channel or source            | That source's automation is off — every source starts off   | Enable the source's automation in Fuely AI → Automations                                    |
| AI doesn't reply anywhere                            | Main bot is toggled off — it's the master switch for all AI | Turn Main bot back on                                                                       |
| AI is silent during the day but replies at night     | **When AI replies** is set to **Outside working hours**     | Switch it to **Always** (or keep it — that may be exactly what you want)                    |
| Wrong reply under a specific post or ad              | A custom automation outranks the one you expected           | Check the priority order and adjust the rules' targeting and keywords                       |
| AI loops or repeats questions                        | Additional instructions are unclear or too long             | Rewrite the Task into short, numbered steps; add an Example Output                          |
| AI says "I'm not sure" too often                     | Missing FAQ or Catalog item                                 | Add the missing question/answer or service to the Knowledge Base                            |
| AI gives long paragraphs                             | No Output Format in the instructions                        | Add "Short WhatsApp-style messages, max 2 lines" to Output Format                           |
| AI promises unavailable times                        | Instructions incorrectly include availability               | Remove time slots from the instructions → instead ask "When would you like to come?"        |
| AI makes up prices or discounts                      | Price info not in Catalog or FAQ                            | Add correct prices to Catalog; never rely on AI memory                                      |
| Escalation doesn't happen when user asks for a human | No Switch to human rules set                                | Add rules in the **Switch to human** setting with keywords like "human," "agent," "manager" |
| Handoffs sit unclaimed in the shared inbox           | Rules don't assign a specific teammate                      | Assign time-sensitive rules to a person so each chat has an owner                           |
| AI mixes Catalog/FAQ with its instructions           | Duplicate info pasted into Additional instructions or About | Remove Catalog/FAQ from the instructions; keep them only in the Knowledge Base              |
| AI ignores customer typos                            | FAQ only includes one phrasing                              | Add alternate phrasings or synonyms into FAQ questions                                      |
| AI too formal or robotic                             | Tone of voice not defined in the instructions               | Add a style guide to Output Format (e.g. "friendly, warm, short messages")                  |
| AI confirms bookings instead of manager              | Instructions don't specify escalation                       | Rewrite the Task: collect details → ask for preferred time → confirm → escalate             |

### 11.1 General debugging flow [#111-general-debugging-flow]

When something feels off:

1. **Check the automation** — is the source's automation on? Is Main bot on? Is **When AI replies** set the way you expect? Is another rule outranking this one?
2. **Check the Knowledge Base** (Catalog, FAQ, About).
3. **Check Additional instructions** (clear steps + output style?).
4. **Check Switch to human** (rules + assignees).
5. **Test in the Test chat** to replicate the problem, fix, and re-test until the flow feels natural.

With these fixes, most issues can be solved in minutes.

## 12. Checklists [#12-checklists]

### 12.1 Launch checklist [#121-launch-checklist]

Before going live, confirm you've done the following:

* [ ] Connected your channel (WhatsApp, Instagram, Facebook, TikTok, or website)
* [ ] Filled out About Company with correct hours, address, and contacts
* [ ] Built a Catalog (≤50 items, accurate prices)
* [ ] Added 10–25 FAQ entries (clear, short answers)
* [ ] Wrote Additional instructions with Role → Task → Output Format → Example Output
* [ ] Enabled the automations for every source you want the AI to cover
* [ ] Configured Switch to human rules and assigned owners where needed
* [ ] Chose the right **When AI replies** option (Always, or Outside working hours with business hours filled in)
* [ ] Tested flows in the Test chat: booking, FAQ, after-hours, handoff
* [ ] Drafted after-hours and escalation messages

### 12.2 Update checklist [#122-update-checklist]

Run this monthly or quarterly:

* [ ] Update business hours (holidays, seasonal changes)
* [ ] Refresh Catalog (new services, removed items, updated prices)
* [ ] Add new FAQs based on recent chats
* [ ] Remove irrelevant or outdated FAQs
* [ ] Review Additional instructions for clarity and tone consistency
* [ ] Check Switch to human rules: still correct? Assigned to right people?
* [ ] Review which automations are enabled: pause finished campaigns, cover new posts and ads
* [ ] Test the Test chat again with 3–5 flows
* [ ] Review customer feedback and adjust messages accordingly

With these checklists your team can launch quickly, stay consistent, and maintain a professional AI assistant without confusion.
