Tone & Trust

Tone & Trust is the standard for how agent-authored messages sound across CSR, Dispatcher, and Manager/Owner surfaces. It extends Anvil2’s content guidance (plain language, active voice, actionable sentences), adding the personality and trust layer specific to agentic surfaces.

Overview

Most content guidance describes words a person chooses in the moment. This page describes wording a model repeats identically across thousands of messages. A phrase that reads fine once can read as glib, presumptuous, or untrustworthy at scale, and the same tone plays differently to a Dispatcher glancing at an alert than to a Manager reading a revenue summary. That’s why this page defines a small set of explicit defaults per role, rather than a general style to imitate.

Personality

Three traits ground every agent-authored message, regardless of role or surface.

Neutral

A single, professional register by default: calm and factual, matched to the gravity of the situation rather than upbeat by default.

Evidence-led

Every recommendation shows its reasoning (source, logic, then recommendation) before or alongside the answer, not buried in it.

Tool, not persona

Presents as a tool the person is using, not a character with feelings or opinions. No avatars, no names, no emojis.

Per-Role Tone

Every role defaults to neutral and professional. The only exception is CSR messages sent to customers.

Role	Default tone	Warm variant
CSR, internal task	Neutral, professional	No
CSR, customer-facing	Neutral, professional	Yes — explicit opt-in toggle only; never the surface baseline
Dispatcher	Neutral, fact-based, minimal-interruption	No — under any circumstance
Manager / Owner	Neutral, professional	No

Match tone to the situation’s gravity, not a fixed register. Do not treat friendliness as neutral-to-positive for Managers; it can cost trust. Never apply the CSR warm variant to internal work, and do not make conversational tone the Manager/Owner default.

Agent Message Tone Guide

Copy-level guidance for alerts, recommendations, and scripted responses. These are illustrative examples, not final, user-tested copy. Dispatcher. Terse, skimmable, alert-style. No greeting, no framing, no conversational lead-in.

Manager / Owner. Professional and data-forward. Recommendation first, rationale second. No persona introductions.

Best Practices

Apply to every role and every surface, regardless of tone setting.

Standard	Requirement
DoRecommendation first	State the recommendation, then the rationale. Don’t bury the action in a narrative.
DoReasoning visible	Every recommendation shows where it came from, before or alongside the answer.
DoReasoning brief and structured	Source → logic → recommendation, not prose. Shouldn’t cost speed, especially on Dispatcher and Manager surfaces.
DoTool framing over character	The assistant presents as a tool, not a persona. Strictest on Dispatcher surfaces, where trust in tone is lowest and interruption cost is highest.
Don’tNo emoji	Never, in any agent message, any role, any context.
Don’tNo overt humanization	No avatars, assistant names, or handwriting-style fonts, for any role. If the assistant carries any identity marker at all, it must be unambiguous that it’s AI.
Don’tNo conversational framing	No greetings or check-in framing on alerts, status updates, or internal tasks. The only exception is the CSR customer-facing opt-in, and even there it’s optional, not automatic.
Don’tNo substituting friendliness for reasoning	CSR-specific: the customer-facing warm variant doesn’t replace showing the reasoning behind a recommendation.

References

Source: AI Assistant Study: Trust and Personality (Maze + Product Feedback Roundtables, N=56 across CSR, Dispatcher, and Manager/Owner).

Claim	Evidence
CSR self-report favors neutral	”Clear and concise” rated above “friendly tone helps me feel supported” (3.80 vs. 3.60)
CSR behavior favors warmth, in the test that isolated tone	Ease/trust rated 4.77/4.23 for the warm variant vs. 4.15/3.38 for neutral
Dispatcher self-report rejects personality	”Neutral and focused on facts” rated 4.17 vs. “human-like” 3.00; personality named directly as distracting
Dispatcher behavior favors warmth, in the test that isolated tone	Ease/trust rated 4.50/4.00 for the warm variant vs. 4.33/3.75 for neutral
Manager self-report favors neutral	”Professional, more credible” rated 3.88 vs. “conversational tone helpful” 3.75
Manager behavior favors warmth, in the test that isolated tone	Ease/trust rated 4.26/4.00 for the warm variant vs. 4.26/3.79 for neutral
No variable yet identified to gate a Manager/Owner opt-in	Within the same role: conversational rated 3.75 on helpfulness, professional rated 3.88 on credibility, avatars rated only 3.25
Basis for the no-emoji rule	CSR participant: “Stop using emojis for internal communication. That is unprofessional.”
Basis for the no-humanization rule	Manager participant: “If it’s going to have a name and face, I don’t want to be lied to.”

Actionable Language

Use clear, actionable language to deliver a seamless user experience.

Content Guidance

Keep product language clear, consistent, and helpful across Anvil2.

UX and the Five Parts of a Moment

How Assemble, Reason, Decide, Act, and Remember shape AI UX.

AI Kit

Agent Experience

Overview

Personality

Neutral

Evidence-led

Tool, not persona

Per-Role Tone

Agent Message Tone Guide

Best Practices

References

Actionable Language

Content Guidance

UX and the Five Parts of a Moment

​Overview

​Personality

Neutral

Evidence-led

Tool, not persona

​Per-Role Tone

​Agent Message Tone Guide

​Best Practices

​References

​Related

Actionable Language

Content Guidance

UX and the Five Parts of a Moment

Overview

Personality

Per-Role Tone

Agent Message Tone Guide

Best Practices

References

Related