Tuesday night, 9:40 PM. A practice manager is reviewing last week's after-hours recordings. Forty seconds into one of them, the caller - a patient trying to reschedule - stops mid-sentence and asks, "Wait. Am I talking to a real person?" The agent, configured by some vendor to stay on script, answers: "I'm here to help with your appointment. What day works best?" Two seconds of silence. Click.
Play that recording to anyone honest and they'll tell you what lost the call. It wasn't the AI. The caller was rescheduling happily until she asked a direct question and got a dodge. The moment your agent evades "are you a robot?", every claim it makes afterward - the balance, the appointment time, the confirmation number - inherits the evasion. So should your AI agent announce itself? In 2026 the answer has three layers: what the law says, what the research says, and what your own call log says. They point the same direction.
The rules are converging on one answer
Start with the phone. In February 2024 the FCC ruled that calls using AI-generated voices count as "artificial or prerecorded voice" calls under the TCPA - not a proposal, an interpretation of existing law, in effect now. For outbound calling, that pulls AI voice into the consent regime that already governs robocalls. The FCC then went further and proposed a rule that would require AI calls to announce themselves: a clear, plain-language disclosure that the caller is hearing an AI-generated voice, delivered at the opening of the call, plus specific consent for AI calls. Comments closed in late 2024, and as of this writing the rule hasn't been finalized - but nothing in the pending version got weaker between draft and docket. We covered the collections angle in detail in our TCPA piece; if your agents call debtors or patients, that rule is the one to build for, and the compliance bar for collections is already higher than for anyone else.
The states aren't waiting. Utah's AI Policy Act has required since May 2024 that a business using generative AI with consumers disclose it - at minimum when the consumer asks, and proactively in sensitive interactions. That makes the dodge in the opening scene not just bad manners but, in at least one state, a legal problem. California has prohibited using a bot to mislead people about its artificial identity for commercial purposes since 2019, and its newer chatbot laws keep extending the principle. More states add disclosure bills every session.
Every regulator that has examined the question has landed on the same side: the customer's right to know outranks the vendor's fear of the answer. You can build for the strictest version now - disclosure in the first sentence, everywhere - or retrofit it later under deadline. Nobody has ever been fined for disclosing too clearly.
What disclosure actually does to a call
The honest answer: the research splits, and the split is instructive.
On one side, COPC's 2025 customer-experience research found that customers who knew they were talking to AI reported satisfaction dramatically higher - 34 percentage points - than customers who weren't told. Academic work published in 2025 under the name "the transparency dilemma" found the opposite valence: people who learn a counterpart used AI trust them less than people never told. And one industry call report found abandonment jumping from roughly 4% on human-answered calls to nearly 25% when callers realize they've reached an AI.
Read those side by side and a pattern emerges. The satisfaction gains come from disclosure done upfront, on purpose, followed by a competent interaction. The trust losses and hang-ups cluster around discovery - the caller figuring it out mid-call, or finding out afterward. Being told is fine; catching you is not. The variable isn't whether the caller knows. It's whether you told them, and what happened in the sixty seconds after.
Callers don't punish you for using AI - they punish you for hiding it, and for wasting their time after admitting it. A disclosed agent that resolves the call in three minutes beats an undisclosed one that gets caught, every time, in every study worth reading.
The playbook: disclose in one sentence, then earn the next thirty
Here's what the strict-and-honest version looks like in practice:
- Disclose in the first sentence, in plain words. "You've reached Ridgeline Plumbing - I'm their AI assistant, and I can book you in or get you help right now." Not "virtual concierge," not "automated experience." Callers know what AI means; hedging reads as hiding.
- Prove competence immediately. The disclosure buys you about thirty seconds of suspended judgment. Spend them doing something a phone tree can't: pull the account, quote the actual balance, offer tonight's real availability. Usefulness converts skeptics; small talk does not.
- Keep the exit visible - and instant. Callers tolerate AI when they know the door to a human is unlocked. "You can ask for a person anytime" only works if the transfer is immediate and arrives with context attached, not a fresh queue and a second interrogation.
- Never fake humanity. No simulated typing pauses, no scripted "um"s engineered to pass, and never - under any configuration - denying being AI when asked. In Utah that denial may be unlawful. Everywhere, it converts a routine call into the recording that ends up in a complaint.
- Same identity on every channel. If your website chat discloses and your phone line doesn't, the customer who uses both now knows you disclose only when forced. One policy, every channel.
Notice what's not in the playbook: apologizing. An agent that opens with "I'm just the AI, sorry" teaches the caller to distrust it. The disclosure is a statement of fact, not a confession.
Run your own numbers
Vendor studies argue in both directions, so don't adjudicate this with someone else's data. Run the test on your own line for one month. Turn on first-sentence disclosure and pull the recordings weekly. Count four things: hang-ups in the first thirty seconds (the alleged cost of honesty), resolution rate on disclosed calls versus your human or undisclosed baseline, mid-call "am I talking to a robot?" interruptions (with first-sentence disclosure, this number should drop to zero - if it doesn't, your disclosure isn't landing), and whether any complaint or review that month mentions the AI at all.
If disclosure costs you calls, the recordings will show whether it was the sentence that lost them - or the sixty seconds after it. In our experience it's almost always the sixty seconds: the agent that discloses and then fumbles the account lookup was never going to keep that caller anyway. Fix the competence and the disclosure takes care of itself.
Start small, measure it
Verlingo agents disclose by default - first sentence, plain language, on voice and chat alike - and hand off to your team the moment a caller asks, with the transcript and context attached. Every plan includes recordings and transcripts, so the one-month audit above is something you can run on us from day one, not something you take on faith.
Pick one line - after-hours, overflow, the billing queue - and run it disclosed for a month. Score the four numbers. If honesty is costing you calls, you'll see exactly where; if it isn't, you've just future-proofed yourself against the rule the FCC has already drafted. Setup takes minutes, the trial is free, and the rates are published - no bundle to decode. The disclosure question is going to be answered for your industry eventually. Better to have a month of your own recordings when it is.
