Legal Requirement
EU AI Act requires transparency when users interact with AI systems. Specific exceptions exist (law enforcement, etc.), but customer-facing CRM almost always requires disclosure. US state laws expanding similar requirements.
Article 50 of the EU AI Act, in force from August 2026, requires disclosure “in a clear and distinguishable manner at the latest at the time of the first interaction.” Penalties reach EUR 15M or 3% of global turnover. California SB 1001 already requires bot disclosure in commercial contexts; Colorado AI Act and Utah’s Artificial Intelligence Policy Act extend similar duties. Health and finance use cases face overlapping obligations from HIPAA disclosure norms and CFPB UDAAP guidance — disclosure is no longer optional anywhere your customers live.
Wording
Plain language. “You’re chatting with an AI assistant. Ask to speak with a human anytime.” Not “Our digital service experience is here to help” (euphemistic). Clarity beats cleverness.
Test wording against an 8th-grade reading level (Flesch-Kincaid 60+). Use the word “AI” or “automated assistant” — not “virtual agent” or “smart helper,” which user research consistently shows people misread as a remote human. Localize: in German markets “KI-Assistent” tests better than “AI assistant.” Place disclosure as the first message, not in a footer. Persist a small “AI” badge near the input field so users who scroll back still see the status.
Handoff Signals
User says “agent” or “human” or “representative” — escalate immediately. Don’t require the user to say the exact magic word. Sentiment signals (frustration, repeat phrasing) also trigger escalation.
Build a regex + LLM-classifier hybrid: regex catches the obvious (“speak to a person”, “real human”, “transfer me”), the classifier handles indirect signals (“this isn’t working”, repeated rephrases of the same question, profanity, three failed turns on one intent). Set a hard ceiling — no more than 6 turns on a single unresolved issue before offering escalation proactively. Track the escalation rate per intent; spikes mean the bot is failing silently.
Trust Cycle
Clear disclosure builds trust. Users who know they’re talking to AI and have clear escape path report higher satisfaction than those unsure. Hiding AI status erodes trust when discovered.
Pew Research data from late 2025 shows CSAT rises 12 points when AI is disclosed up front versus discovered mid-conversation. Trust collapses sharply when users feel deceived — repeat-contact rates double and social-media complaints spike. Disclosure is a CSAT lever, not a tax.
Common Failure Modes
Disclosure buried in a chat window’s collapsed header. Voice bots that mimic human speech patterns without saying “I am an AI.” Email agents signing as a human name (“Sarah from Support”) with no automated-message footer. Each is a regulatory exposure under Article 50 and an erosion of customer trust.
What to Do This Week
Audit every customer-facing AI surface for first-message disclosure and a one-tap human escape path.
