This guide explains how to configure AI chat for **high resolution, good customer experience, and long-term reliability**.

The most important thing to understand up front:

> AI chat is not one big prompt.
>
> Different settings are used at different moments to make different decisions.

Putting the right information in the right place matters more than adding more instructions.

* * *

## 1\. The Right Mental Model

AI chat configuration answers four different questions:

1. **Should this conversation be escalated to a human?**
2. **Is there a workflow that should handle this request?**
3. **If answering, how should the assistant respond?**
4. **How should the response sound?**

Each section of the product maps to one of these questions.

* * *

## 2\. Style Guide: How the Assistant Should Sound

**Primary purpose:** control _tone and delivery_

**Applies everywhere:** all responses, including workflow responses

### Use the Style Guide for:

- Tone (warm, neutral, formal, concise)
- Verbosity (short vs detailed)
- Empathy framing
- Formatting preferences
- Language patterns to avoid or prefer

**Examples**

- “Be calm, direct, and reassuring.”
- “Avoid over-apologizing.”
- “Prefer short paragraphs and bullets.”

### Do _not_ use the Style Guide for:

- Escalation logic
- Business rules
- Process steps
- Decision logic

**Rule of thumb**

> If two agents gave the same answer, the style guide controls how that answer _feels_.

* * *

## 3\. Critical Instructions: How the Assistant Answers (When Answering)

**Primary purpose:** guide _answering strategy_ when the assistant is responding conversationally

Critical instructions help the AI decide **how to respond**, not **what action to take**.

### Use Critical Instructions for:

- Whether to attempt an answer vs. ask clarifying questions
- How to handle ambiguity
- How confident or cautious to be
- Cross-cutting conversational guardrails

**Good examples**

- “Attempt a best-effort answer when confidence is high.”
- “Ask clarifying questions if multiple interpretations are possible.”
- “Avoid speculation when information is incomplete.”

### Do _not_ use Critical Instructions for:

- Escalation conditions
- Workflow steps or logic
- Detailed policies
- Tone or verbosity (use the style guide)

### How many is reasonable?

- **Recommended:** ~5–15 instructions
- More is not better — unclear or overlapping instructions reduce reliability

> Clear instructions improve performance.
>
> Conflicting or overly detailed instructions can make responses less consistent.

* * *

## 4\. Handoff Rules: When to _always_ Escalate

**Primary purpose:** define situations that should _always_ go to a human

Think of these as **hard stops**.

### Use Handoff Rules for:

- Legal threats
- Safety or emergency scenarios
- Highly regulated requests
- Customer segments that should always speak to an agent

When a handoff rule matches, escalation happens immediately.

### Do _not_ use Handoff Rules for:

- General confusion
- Long conversations
- Common support topics that workflows can handle

> Every handoff rule trades automation for certainty. Use them sparingly.

* * *

## 5\. Company-Specific Handoff Instructions: How to Decide _When_ to Escalate

**Primary purpose:** guide judgment around escalation

These instructions **influence decisions** but do not force escalation.

### Use these instructions for:

- Explaining what “urgent” means in your business
- When escalation should be delayed to allow automation to resolve the issue
- How to interpret emotional language in your domain
- Guidance such as:
  - “Prefer workflows when they can fully resolve the request.”
  - “Escalate only after multiple unsuccessful attempts.”

### Do _not_ use these instructions for:

- Step-by-step procedures
- Exact phrase matching
- Policy text copied from internal docs

Think of this section as:

> “How a strong human agent would decide _when_ to escalate.”

* * *

## 6\. Workflows: When Something Should _Happen_

**Primary purpose:** handle structured, repeatable actions

### Use Workflows for:

- Multi-step processes
- Data collection
- Verification
- Actions like returns, cancellations, updates, or lookups

If you notice the assistant frequently **explaining** instead of **acting**, that’s usually a signal that:

- A workflow is missing, or
- A workflow is too narrowly scoped

### Do _not_ use Workflows for:

- Simple explanations
- One-off questions
- Tone or style guidance

* * *

## 7\. Practical Setup Guidance

When configuring AI chat:

- Start with **fewer, clearer instructions**
- Add handoff rules only where escalation is truly required
- Use workflows for consistency and repeatability
- Revisit and refine instructions rather than continually adding new ones

* * *

## 8\. The High Level Summary to Remember

- **The style guide controls how responses sound.**
- **Critical instructions guide how the assistant answers.**
- **Workflows control what the assistant does.**
- **Handoff rules control when automation stops.**
- **Clear structure beats more configuration every time.**

If you're now ready to set up your AI Chat settings, check out the [How to Configure Your Assembled AI Chat using an LLM](https://support.assembled.com/hc/en-us/articles/43278977852557-AI-Chat-How-to-Configure-your-Assembled-AI-Chat-Using-an-LLM) guide to get started!