---
title: Teaching Your AI to Think — HubSpot Customer Agent
---

1 / 10

← → or click to navigate

←

→

HubSpot Solutions Architecture

# Don't teach your AI the *answer*. Teach it to *think*.

 A 10-minute guide to writing guidelines that actually make  
 HubSpot's Customer Agent smarter — not just more verbose.

AUDIENCE

Customer Success + SA Teams

DURATION

~10 minutes

The Problem

## Most Customer Agents fail for the same reason.

What teams do

 ❌  "If asked about refunds, say: *Refunds are processed within 3–5 business days.*"

 ❌  "For billing questions, respond with the pricing page link."

 ❌  Copy-paste the FAQ into Guidelines.

What the Agent needs

 ✅  *"We prioritize speed for first-time issues, but escalate if the same customer contacts us twice in 7 days."*

 ✅  The **reason** behind the answer, not just the answer itself.

 Giving an AI direct answers is like giving a new hire a script — it breaks the moment something unexpected happens.

How the Customer Agent Works

## Three layers. One decision engine.

🗂️

#### Layer 1 — CRM Context

Contacts, Companies, Tickets, Deals, Conversation History — the "who is this person?" layer

Real-time

📚

#### Layer 2 — Knowledge Sources

Knowledge Base articles · Website pages · Internal documents — the "what do I know?" layer

Indexed

🔗

#### Layer 3 — External Integrations

Gmail · Slack · Google Docs · Teams — the "what else can I check?" layer

Optional

**Guidelines sit above all three layers.** They tell the agent *how to use* what it finds — not what to say.

The Core Principle

## Guidelines are the *operating system*. Knowledge is the *database*.

Guidelines

How to think  
& behave

\+

Knowledge Sources

What it  
knows

\=

Output

A genuinely  
smart agent

 ⚠️  Mixing them up is the #1 reason Customer Agents give outdated, robotic, or irrelevant answers.

Layer 1 — CRM Data

## The agent knows *who* they're talking to. Does your Guideline use that?

### Ask your customer:

- 🏷️
  
  **Are VIP / at-risk customers tagged in CRM?** If yes → the agent can treat them differently
- 📋
  
  **What ticket properties matter most?** Priority, type, open/closed status, owner
- 🔁
  
  **Is there a repeat-contact threshold?** "Same issue twice in 7 days → escalate" is a Guideline, not a KB article

Guideline Example

 "If the contact's Lifecycle Stage is **Customer** and they have more than **2 open tickets**, always prioritize speed over completeness and offer to connect them with a human agent."

 💡  CRM data makes Guidelines **personalized**. Without clean CRM properties, the agent treats every customer the same.

Layer 2 — Knowledge Sources

## Great Guidelines depend on *great knowledge architecture*.

KB Articles

 Ask: *"Could a new hire act on this article alone?"*

 If **No** → fix the KB first. The agent will be confused too.

Website Pages

 Outdated pricing pages, old product names, or conflicting info will be surfaced as-is.

**What's indexed matters.**

Internal Docs

 Granularity varies wildly. The agent reads these literally.

 Standardize format before uploading.

 🎯  Your Guideline can say *"prefer KB articles over website content for pricing questions"* — but only if the KB is actually trustworthy. **Source quality = answer quality.**

Layer 3 — External Integrations

## More data ≠ better agent. *Signal-to-noise* is everything.

What to watch for

- ⚠️
  
  Slack channels with **too much internal noise** — the agent can't filter context vs. chatter
- ⚠️
  
  Google Docs with **version history confusion** — the agent reads the current state only
- ⚠️
  
  Email threads that contain **sensitive internal discussion** not meant for customers

The Design Principle

 For external sources, the most important Guideline is:

**"What should the agent *not* look at?"**

 💡  Connect only what's clean, current, and customer-safe. Use Guidelines to set priority order when sources conflict.

SA Playbook

## 3 questions to unlock the right Guidelines.

1

Scenario Mapping

"What are the top 3 scenarios where you absolutely cannot afford for the agent to fail?"

→ Work backwards from failure to write the decision logic

2

KB Quality Test

"If a new hire read only your KB, could they handle 80% of tickets without asking for help?"

→ Yes = good KB. No = fix the source before writing Guidelines

3

Silent Rules

"Are there questions the agent might know the answer to — but should never answer?"

→ Uncovers the "what not to do" logic — the most overlooked part of Guidelines

Before vs. After

## What "thinking guidelines" actually look like.

❌ Answer-based (brittle)

 "If the customer asks about refunds, tell them refunds take 3–5 business days."

 "For pricing, send the link to the pricing page."

 "Always say thank you at the end."

vs

✅ Thinking-based (resilient)

 "We value **speed over perfection** for first-contact issues. If uncertain, offer to connect with a human rather than guessing."

 "Escalate when: same customer, same issue, 2nd contact within 7 days — regardless of severity."

 Answer-based Guidelines expire. Thinking-based Guidelines scale with every new scenario.

Remember This

# "Don't give your agent a script. Give it a *value system*."

### GUIDELINE

The OS — how it thinks, what it prioritizes, when it escalates

### KNOWLEDGE SOURCE

The database — what it knows, verified, current, well-structured

### CRM DATA

The context — who it's talking to, and why that changes the response

HubSpot Solutions Architecture · 2025