Dew Design System
AI guidelines

AI writing guidelines

How to write for AI-powered experiences across Freshworks products. Covers Freddy AI personality, transparency patterns, trust, and tone.

Intro

AI is woven into the fabric of Freshworks products through Freddy AI. From suggested replies to intelligent ticket routing, AI touches nearly every workflow. The words we use around AI shape how much users trust it, understand it, and ultimately benefit from it.

These guidelines ensure that every AI interaction feels natural, honest, and helpful — never gimmicky, opaque, or overconfident. They apply to all AI-powered features across the Freshworks product suite.

Core principles

Six principles that guide how we write for AI across all Freshworks products.

Transparent

Always let users know they're interacting with AI. Never disguise automated responses as human-written.

Trustworthy

Be honest about what AI can and cannot do. Set accurate expectations and never overstate confidence.

User-controlled

Keep humans in the loop. Let users edit, override, or dismiss AI suggestions at any point.

Clear

Write AI responses in plain language. Avoid technical jargon, and explain reasoning when it helps the user decide.

Helpful, not performative

AI should solve real problems. Don't add AI interactions that exist only to showcase the technology.

Calibrated

Match the AI's confidence level in its language. Use hedging for uncertain suggestions and definitive language only when warranted.

Freddy AI guidelines

Freddy is Freshworks' AI assistant. It has a distinct personality that should feel consistent across all products — from Freshdesk to Freshservice to Freshsales.

Knowledgeable but not condescending

Freddy explains complex topics simply without talking down to users. It provides context when helpful but doesn't over-explain.

Do

This ticket looks similar to 3 others from this week. Want me to link them?

Don't

Based on my advanced natural language processing capabilities, I have identified semantic similarities between this ticket and 3 other tickets.

Proactive but not intrusive

Freddy offers help when it detects opportunity but never forces itself into the workflow. Suggestions are easy to dismiss.

Do

I noticed this customer wrote in before. Here's their history.

Don't

ALERT: I've found previous interactions! Click here to view my AI-powered analysis!

Conversational but professional

Freddy uses natural language with contractions but avoids slang, humor in serious contexts, or overly casual tone.

Do

I couldn't find a match. Try broadening your search terms.

Don't

Oops! No luck finding that one 😅 Maybe try something else?

Naming conventions

  • Always write “Freddy AI” on first mention (not “Freddy” or “the AI”)
  • After first mention, “Freddy” alone is fine within the same context
  • Use “Suggested by Freddy AI” for attribution labels, not “AI-generated”
  • Don't personify Freddy with gendered pronouns — use “it” or rephrase

Tone and voice for AI

AI copy follows the same Dew voice principles as all UI copy, with additional guidance for AI-specific scenarios.

Confidence language

HighThis is, Will, Matches
MediumLikely, Appears to, Seems like
LowMight be, Could be, Possibly

Action framing

Frame AI actions as suggestions, not decisions.

“Freddy suggests categorizing as Billing”
“Freddy categorized this as Billing”

Transparency patterns

Users need to know when they're interacting with AI, what data informed a suggestion, and how confident the system is. Use these patterns consistently.

PatternPurposeExample
AI attributionLabel AI-generated content clearly so users know the source.Suggested by Freddy AI
Confidence indicatorsShow how certain the AI is about its suggestion so users can calibrate trust.High confidence · Based on 50 similar tickets
ExplanationBriefly explain why the AI made a suggestion to help users evaluate it.Suggested because this customer mentioned 'billing' and 'refund'
Editable outputsPresent AI drafts as starting points, not final answers. Make editing frictionless.Here's a draft reply. Edit it before sending.

Error and uncertainty

AI will get things wrong. How we communicate failures and uncertainty directly impacts user trust.

AI can't generate a response
Do

Freddy couldn't generate a suggestion for this ticket. Try adding more details.

Don't

Error: AI model returned null response. Please try again later.

Low-confidence suggestion
Do

This might be a billing issue, but Freddy isn't sure. Review the ticket details.

Don't

AI confidence: 32%. Category: Billing.

AI feature is unavailable
Do

Freddy AI is temporarily unavailable. You can continue working — we'll let you know when it's back.

Don't

Service disruption: AI inference endpoint is down. ETA unknown.

Trust and safety

Building trust in AI is earned through consistent, honest behavior over time. These rules protect both users and the Freshworks brand.

Never auto-execute irreversible actions

AI can suggest deleting, sending, or modifying — but the user must confirm. Always show a confirmation step for destructive actions.

Respect data boundaries

Be explicit about what data Freddy accesses. Don't reference data the user wouldn't expect AI to see.

Graceful degradation

When AI features fail, the product must remain fully functional. AI is an enhancement, not a dependency.

Feedback loops

Let users flag bad AI suggestions with thumbs up/down or 'not helpful'. Use this to improve trust and accuracy over time.

Do's and don'ts

Quick reference for writing AI copy across Freshworks products.

Do
  • Label all AI-generated content with a clear 'Freddy AI' attribution
  • Use hedge words ('might', 'appears to', 'likely') when confidence is below threshold
  • Provide an easy way to dismiss, edit, or override every AI suggestion
  • Explain the reasoning behind suggestions in one short sentence
  • Use sentence case for all AI-related labels and messages
  • Show what data the AI used to reach its conclusion when possible
  • Write fallback messages for when Freddy can't help
  • Test AI copy with real users to catch trust issues early
Don't
  • Don't pretend AI output is human-written
  • Don't use absolute language ('will', 'definitely') for uncertain predictions
  • Don't add AI features that don't solve a real user problem
  • Don't use AI jargon ('model', 'inference', 'tokens') in user-facing copy
  • Don't apologize excessively when AI gets something wrong
  • Don't auto-execute actions without user confirmation
  • Don't use humor or personality quirks in error states
  • Don't promise capabilities Freddy doesn't reliably deliver

Start writing for AI

Review the writing guidelines, explore component patterns, or check the word list.