All insights
Fundamentals6 min read

The Sentence Your Agent Has to Survive

Vishal Sachar

Co-Founder & CEO of CLRT

Ask a business owner what worries them about putting an AI voice in front of buyers and the answer is rarely the technology. It is one sentence: the moment the agent says it is not a person. They have heard that buyers hang up, and suspect that the honest version of the product is the one that does not sell. So the market has spent two years arguing about whether, when and how quietly an agent should admit it. That argument is over, and not because anyone won it. The law now requires the sentence in Europe and, for bots that sell online, in California. American telephone rules make an AI caller get consent first, and the UAE Central Bank lets customers of licensed financial institutions refuse AI agents. What remains is a harder, more useful question: what happens in the conversation after it.

79.7%1
Cut in purchase rates when customers were told before an outbound sales call that the caller was a chatbot (field experiment, 6,200+ customers)
Marketing Science, 2019
2 Aug 20262
The date from which the EU AI Act generally applies, including Article 50's duty to tell people, at the latest at the first interaction, that they are dealing with an AI system
EU AI Act, Art. 113, 2024
EUR 15m3
The maximum fine under the EU AI Act for breaching the Article 50 transparency duties, or 3% of worldwide annual turnover where that is higher
EU AI Act, Art. 99(4), 2024

The fear is not superstition. The strongest evidence comes from a field experiment published in Marketing Science in 2019 by Xueming Luo and colleagues, run with more than 6,200 customers of a financial services firm who received outbound sales calls. When nobody was told, the chatbots sold as well as the firm's proficient staff and four times as well as its inexperienced ones. When customers were told before the conversation that the caller was a machine, purchase rates fell by more than 79.7 per cent. By simple arithmetic on the paper's ratios, a bot that announced itself sold less than the least experienced humans on the floor. The authors also found the damage was smaller when the disclosure came later in the call, and among customers with prior experience of AI. The first of those remedies is now closed, and rightly. The second is the one that matters.

FIG. 01Told before the conversation that the caller was a chatbot, customers cut purchases by more than 79.7 per cent. Index derived from the ratios in Luo, Tong, Fang and Qu, Marketing Science, 2019.
01The law settled it

The law has been closing in on the sentence from several directions. California has made it unlawful since July 2019 to use a bot online to mislead a person about its artificial identity in order to sell, and a business that discloses is not liable. In February 2024 the US Federal Communications Commission ruled that AI-generated voices are artificial voices under its robocall law, so an outbound AI call needs the called party's prior express consent and must identify the business behind it. The UAE Central Bank's telemarketing regulation, in force from 31 March 2026, lets the customers of licensed financial institutions choose between a human, a robocall or an AI agent. And from 2 August 2026, Article 50 of the EU AI Act, the one rule that fixes the timing, requires that people are told they are dealing with an AI system, clearly and at the latest at the first interaction, with fines of up to EUR 15 million or 3 per cent of worldwide turnover. That closes the late door, and Europe's omnibus reforms left it standing.

FIG. 02Four regimes, dated from the day each applied, with bars running to this article's date. Sources: California Business and Professions Code 17941 and 17943, FCC 24-17, the CBUAE regulation as reported by Pinsent Masons, and the EU AI Act.
02Why buyers go curt

So the sentence is fixed. What is not fixed is what it costs, and the 2019 data says where the cost comes from. Disclosed customers became curt, cut the calls short and judged the bot less knowledgeable and less empathetic, even though it performed as well as it did undisclosed. The disclosure did not change the machine. It changed the buyer's expectation of the machine, and the expectation then priced the conversation. Two details sharpen this. The calls were outbound and unasked for, where a buyer's patience is at its thinnest. And the penalty shrank with prior AI experience, which means the cost of the sentence is not a constant of human nature. It is a prior, formed by every poor bot the buyer has met, and a prior can be overturned by evidence inside a single call.

A second field experiment, published in Management Science in January 2026 by Yuqian Xu, Hongyan Dai and Wanfeng Yan, shows the prior moving. A freight platform randomly assigned 11,000 truck drivers to outbound calls from its voice dispatcher. Disclosing the chatbot's identity at the start cut the response probability by around 11 per cent. Making the voice more human in manner, with the interjections and small fillers of ordinary speech, raised response probability, conversation length and order acceptance intention by more than 5.6, 24.9 and 10.1 per cent respectively. Most telling, outcomes still improved when the identity was disclosed and the humanising features were added together. The researchers point to trust, and found drivers shared more information with the more human dispatcher. Freight dispatch is not a purchase, so read it directionally. The direction is that a conversation can win back what the sentence costs.

FIG. 03Disclosure cut replies by around 11 per cent; humanising features raised every measure. Xu, Dai and Yan, Management Science, January 2026; freight dispatch, so read directionally.
03After the sentence

This is where most voice agents are being judged wrongly, by their builders as much as their buyers. The effort goes into the voice, the latency and the warmth of the greeting, and the sentence is treated as a compliance line to get through. But the sentence is the moment the buyer sets the bar, and the next minute either clears it or confirms the prior. What clears it is not sounding human. A more human manner helped in the freight data, by the authors' reading, because it built trust, and trust has to be repaid with substance: knowing the product, answering the actual question, showing the buyer the thing they asked about rather than describing it. An agent that is merely pleasant after admitting what it is has spent the buyer's patience twice. The design question was never the wording of the disclosure. It is whether the conversation deserves it.

The disclosure is fixed by law. What it costs is decided by the conversation that follows it.

A deeper dive

Faced with a measured penalty, the tempting response is to engineer the sentence itself: a human-sounding name with the admission folded into a long greeting, a disclosure delivered quickly in a pleasant register, wording tested for the version buyers least notice. Each of those is a bet against the law's own language, which in Europe asks for information that is clear and distinguishable, and in California for disclosure that is clear, conspicuous and reasonably designed to inform. It is also a bet against the buyer. A disclosure the buyer half hears and later works out does not escape the prior. It confirms the worst version of it, that the business was willing to blur the line to make a sale. The same reasoning governs the humanising cues. In the freight experiment the interjections and fillers worked alongside an explicit disclosure, which is the honest reading of that result: humanity of manner, never ambiguity about identity. A voice designed to make people wonder is the one design that the evidence and the law both reject.

The second trap is measurement. Most teams judge a voice agent by one number, conversion, and compare it with the humans or with last quarter's version. That number hides the shape the field experiments reveal. A disclosed agent's call has two phases: the seconds in which the buyer revises their expectation, and the rest of the call, in which the agent either earns the revision back or does not. A business that cannot see where buyers leave, what they asked just before, and whether the agent answered that question or merely replied to it, will spend its effort on the wrong half, usually the voice. The prior is also moving underneath the measurement. The 2019 study found the penalty smaller among customers with AI experience, and every month more buyers have used capable assistants elsewhere. That cuts both ways. A buyer who has met a good assistant arrives with a higher bar, not a lower one, and an agent that only sounds better than last year's bots will lose to the comparison the buyer is already making.

Work with CLRT

CLRT built Closer for the minute after the sentence. It speaks with the buyer and puts what they asked about on their screen in real time, so the expectation the disclosure sets is overturned by evidence rather than charm. If you are weighing a voice agent and the disclosure is what worries you, the useful conversation is about what your buyers need in that minute, and CLRT will have it with you. Meet Tom, Closer's agent, at clrtstudio.com/closer, and judge the minute after the sentence for yourself.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.