The Short Answer

There is no universal CSAT percentage showing how customers react when they learn support was AI. Results depend on disclosure timing, task complexity, resolution, handoff quality, survey wording, and sample. The safest practice is upfront disclosure, an easy human escalation, and a controlled comparison using your own customer feedback and resolution data.

Why Exact Universal Claims Are Misleading

This article previously quoted CSAT ranges and disclosure effects without direct primary sources. Those figures have been removed. A general CX report cannot support a precise percentage unless the linked study contains the exact result, population, and methodology.

What to Test

Compare the same intent under the same conditions and track:

• survey response rate and CSAT using identical wording;
• first-contact resolution;
• escalation and recontact rates;
• abandonment;
• corrections or policy failures; and
• whether AI was disclosed before the interaction.

Do not compare routine AI requests with complex human complaints and call it an AI-versus-human result.

A Trustworthy Disclosure Pattern

Introduce the genie clearly as AI, state what it can help with, and provide a straightforward route to a person. Avoid designing a voice that implies a human identity or revealing automation only after a customer asks.

For routine questions, resolution and speed may matter most. For disputes, sensitive issues, or unusual exceptions, a human should take over with the conversation context attached.

How to Report Results

State the test period, eligible intents, survey wording, response count, exclusions, and escalation rules. Describe small samples as operational observations, not research benchmarks. Report failures and handoffs alongside the average score.

The Salesforce State of Service and Zendesk Customer Experience Trends provide broad customer-service context, but neither should be attached to an exact claim it does not explicitly make.

Practical Next Step

Start with one narrow workflow using a clearly defined voice genie, disclose AI upfront, keep a human fallback, and compare resolution and CSAT with the previous process. Expand only when your own evidence shows the experience is reliable.