---
title: "Setting up CAI"
slug: "setting-up-cai"
updated: 2025-12-12T18:41:59Z
published: 2025-10-03T09:09:52Z
canonical: "docs.natterbox.com/setting-up-cai"
---

> ## Documentation Index
> Fetch the complete documentation index at: https://docs.natterbox.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Setting up CAI

*Please note: this requires version 1.330 or greater for the Natterbox app.*

### **Introduction to Conversational AI** ✨

Now your customers can experience the power of AI directly, as our Assistants and Agents help manage your business calls.

These Assistants act as intelligent routers—no more listening through endless Switchboard options!

They can also answer customer queries directly, saving your agents time and doing much more.

Whereas our Agents can then attempt to fully serve your customer’s needs.

![Icon representing artificial intelligence with a circuit-like design and the letters 'AI'.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.50.png)

### Assistants Overview

A Natterbox AI Assistant is a smart helper for managing inbound call traffic. While they don’t directly manage customer data, they’re designed to assist and route customers efficiently. Here’s an overview of each available Assistant:

![An open book icon representing knowledge and learning in a digital format.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.21.png)

#### [Knowledge Assistant](/v1/docs/256-knowledge-knowledge-bases)

A powerful assistant that is able to handle queries from your customers, simply connect up a knowledge base such as a website or document and a knowledge Assistant is then immediately an expert in that field, quickly serving up answers.

![Icon representing a decision point with arrows indicating multiple directions.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.58.50.png)

#### [Routing Assistant](/v1/docs/253-router)

A powerful assistant that can handle customer queries by connecting to a knowledge base such as a website or document. Instantly, this Assistant becomes an expert in that domain, delivering fast, accurate answers.

![A question mark inside a speech bubble symbolizing inquiry or curiosity.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.58.44.png)

#### [Get Info Assistant](/v1/docs/254-get-info)

Need customer names, security checks, or ticket numbers? This Assistant gathers the essential information and sends it directly to your agents’ screens when they pick up the call—so they’re ready to go.

![A gear symbol with a human head silhouette, representing cognitive processes and technology.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.14.png)

#### [Skills Assistant](/v1/docs/257-skills)

Working hand-in-hand with the Routing Assistant, the Skills Assistant identifies the capabilities required to resolve an issue and ensures it’s routed to the most qualified team.

![A calculator icon representing mathematical calculations and financial tasks.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.28.png)

#### [Instruction Assistant](/v1/docs/259-instructions)

A versatile Assistant that can run live analysis during calls. For example, it can summarize the interaction so far and relay that summary to your agents upon pickup—giving them immediate context on the customer journey.

![Icon representing audio features with a speaker and sound waves illustration.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.06.png)

#### [Voicemail Assistant](/v1/docs/258-voicemail)

When lines are busy, this Assistant ensures voicemails are quickly summarized and routed to the appropriate team—keeping nothing missed or delayed.

![A speech bubble icon representing communication or messaging features.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.36.png)

#### [Response Assistant](/v1/docs/2510-response)

A straightforward Assistant used to deliver message-board-style updates and announcements to your customers—simple, clear, and reliable.

![An icon depicting a person wearing headphones, representing customer support or assistance.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.59.00.png)

### [Agents](/v1/docs/24-agents)

Agents are the next step in AI powered automation, where Assistants help fill a read-only type role, Agents are far more hands on and are able to manipulate data.

### How to configure CAI

With the exception of the Knowledge Assistant, CAI is predominantly be managed within the routing policies.

To begin using CAI simply create an new **Natterbox AI** component from the routing policies :

![Natterbox AI interface showing options to link or create containers for actions.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/image-1744115363644.png)

From here you can select the ‘App’ tab and choose which Assistant you like to use :

![Natterbox AI interface displaying various app options for configuration and management.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/image-1744115506535.png)

To configure an assistant checkout the following articles to view their respective setup :

- [Knowledge Assistant](/v1/docs/256-knowledge-knowledge-bases)
- [Routing Assistant](/v1/docs/253-router)
- [Get Info Assistant](/v1/docs/254-get-info)
- [Skills Assistant](/v1/docs/257-skills)
- [Instruction Assistant](/v1/docs/259-instructions)
- [Voicemail Assistant](/v1/docs/258-voicemail)
- [Response Assistant](/v1/docs/2510-response)

Alternatively you can follow this [QUICK START GUIDE](/v1/docs/252-cai-quick-guide) to put together some fast solutions for quick value use cases.

![Icon representing a professional individual with a tie and formal attire.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.58.37.png)

### Persona

The Persona App is not an Assistant, instead it’s a pre-requisite for all other Assistants.

The Persona allows you to define how your Assistants will talk and behave, it’s recommended that every policy that contains an Assistant has a Persona App defined before, like so :

![Natterbox AI diagram illustrating connections between Persona and Knowledge elements.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/image-1744118068594.png)

![A hand symbol representing support or assistance in a digital context.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-09 at 10.56.57.png)

### Human Escalation

As your Assistants will try their best to service your customers, there will always be scenarios where your human agents will be needed to tap in and take over.

A call will become escalated if the call meets one of the two following criteria :

1. The Assistant detects that it is unable to serve the customer’s request
  1. i.e. the customer’s query is not present within the available knowledge base
2. The customer makes an explicit request to talk to a person

Human escalation provides the ability for an Assistant to detect when it’s time to forward the call on, it’s currently only relevant to the following Assistants :

- [Knowledge Assistant](/v1/docs/256-knowledge-knowledge-bases)
- [Routing Assistant](/v1/docs/253-router)
- [Get Info Assistant](/v1/docs/254-get-info)
- [Skills Assistant](/v1/docs/257-skills)

To enable the ability to escalated to a human simply enable the option from the **Escalation** tab :

![Configuration options for enabling human escalation detection in a user interface.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-08 at 14.37.31.png)

Once enabled a new path will open up on the component, this can then be connected to your human agents, such a **PBX group** or **Call Queue**.

![Diagram illustrating Natterbox AI process with steps for information retrieval and escalation.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/image-1744119825603.png)

Then you can **customise your message** to your customers as it hands off by clicking into ‘Human Escalation’ App and populating the field:

![Configuration settings for human escalation in customer support interface.](https://cdn.document360.io/c81c490c-daaa-4ebd-a3c3-1a2dee8b7617/Images/Documentation/Screenshot 2025-04-08 at 14.43.05.png)

### Prompt Engineering

As part of the CAI configuration, some of the steps will involve direct instructions that will be provided to the Assistants, a prompt, and just like with Prompt Studio there is an art to designing effective instructions to get the most out of the AI assistance.

For CAI the key guideline for effective prompt engineering is to keep all prompts **short and iterate** :

- Set clear goals using as few words as possible
- Test, monitor and update your prompts with continual improvements

### Guardrails

AI typically for the most part will be polite and tempered in how it responds and behaves, but there can always be exceptions and AI is not immune to being mislead and behaving in unexpected and unwanted ways.

With Guardrails you can provide some protection to ensure the Assistants and Agents will always act and behave as you intend in a polite and professional manner.

We offer three levels of protection Low, Medium, High; each policy level implements varying thresholds of protection, with the following features:

- **Prompt Attack Protection** Safeguards against malicious attempts to manipulate the AI (also known as jailbreak attacks), ensuring the assistant stays within its intended behaviour.
- **Content Filtering**
  - **Hate speech**
  - **Sexual content**
  - **Violence**
  - **Insults**
  - **Misconduct**
- **Profanity Filtering** Applied in both the **High** and **Medium** strength policies.

#### **How it Works**

- **Prompt Attack Protection** is applied to customer requests only.
- **Content Filters** (Hate, Sexual, Violence, Insults, Misconduct, and Profanity) are applied to both:
  - Customer requests
  - AI Assistant responses

If any GuardRail is triggered, the AI Assistant will respond with: **"Sorry, I cannot answer this question."**

This ensures a safe and controlled environment for customer interactions at all times.

#### **Enabling a Guardrail**

By default guardrails are enabled for all calls at a **Medium** level.

To change and/or disable a guardrail policy, add the following line to a [**Script Engine**](/v1/docs/script-engine-app) **before** your Assistant or Agent Apps.

All AI Assistants and Agents executed in the call flow after this point will have the guardrail policy applied:

**High**

```plaintext
-- Fsx8 
-- EnableEslASync 
-- OPTION:NOESCAPE  
session.modify("AIGuardRailPolicy", "cai.high-strength-content-policy")
```

**Medium**

```plaintext
-- Fsx8 
-- EnableEslASync 
-- OPTION:NOESCAPE  
session.modify("AIGuardRailPolicy", "cai.medium-strength-content-policy")
```

**Low**

```plaintext
-- Fsx8 
-- EnableEslASync 
-- OPTION:NOESCAPE  
session.modify("AIGuardRailPolicy", "cai.low-strength-content-policy")
```

**Disable**

```plaintext
-- Fsx8 
-- EnableEslASync 
-- OPTION:NOESCAPE  
session.modify("AIGuardRailPolicy", "DISABLE")
```

#### **Choosing a Policy Level**

Whilst protecting against prompt attacks and preventing any use explicit language makes it seem like using a High level no-brainer, the trade-off is the risk of false positive triggers that may block the Assistants and Agents from being able to assist in genuine cases.

For example -

**User:** “I need information about how to treat a ‘die-off’ reaction from probiotics.”

**AI Guardrail Trigger:** Flags the word "die-off" as potentially violent or referring to suicide.

**AI response:** “Sorry, I cannot answer this question."

Unless your customers are likely to be of the vulnerable population we recommended using a **Low** **or Medium level** depending on your business and customers.

#### **Strict Guard Rail Requirement**

If you have a requirement to use highest level of protection, in addition to using the High level policy there is a further step you can use which is to set the AI behaviour to ‘**Synchronous Mode**’.

By **default the AI’s behaviour is Asynchronous**, meaning the AI streams it’s response **immediately** to the caller, so there is a risk that some unwanted behaviour makes it a customer before it cuts itself off on discovering a guard rail trigger.

In ‘Synchronous Mode’ all content the AI communicates is **pre-checked** before it makes it to the customer, ensuring nothing unwanted is ever said.

Use the following to enable ‘Synchronous Mode’ alongside a High level policy :

```plaintext
-- Fsx8 
-- EnableEslASync 
-- OPTION:NOESCAPE  
session.modify("AIGuardRailPolicy", "cai.high-strength-content-policy") 
session.modify("AIGuardRailStreamingMode", "SYNCHRONOUS")
```

The drawback is this method **introduces some latency** as it the service reviews each response, so if you wish to ensure every conversation maintains it’s natural flow we recommend **not** using this option by default.

Same as an IVR.

Using different options (i.e. press 1 for sales, 2 for support) an IVR allows your customers to determine their call route