There's a real tension in AI customer service deployment right now: transparency is increasingly required by law and expected by customers, but the most-cited research on the subject shows disclosure can collapse conversion rates by nearly 80%. Here's a careful look at what the data actually shows — and why the headline finding is less alarming than it appears once you understand the mechanism.

The field experiment everyone cites

The sharpest evidence comes from a 2019 Marketing Science study ("Machines vs. Humans" by Luo, Tong, Fang, and Qu). Researchers ran a field experiment with 6,255 customers of a financial-services firm making outbound calls about loan renewals, randomizing when — or whether — the chatbot disclosed it was AI.

The results by condition:

• Undisclosed: 23.7% conversion rate — statistically on par with proficient human agents (25.1%)
• Disclosed before the conversation: 4.8% — a 79.7% drop. Call length fell from ~64 seconds to ~10 seconds. Customers heard "AI" and hung up.
• Disclosed after the conversation: 11.0%
• Disclosed after the customer had already decided: 23.2% — no meaningful gap from human agents

Same bot. Same knowledge. Same measured empathy. One variable: when the disclosure happened.

Why the penalty occurs — and why it matters

The mechanism is the most important part of this finding. A voice-mining analysis in the study found the disclosed and undisclosed bots were objectively equivalent in knowledge and empathy. What changed was customer perception: once told they were dealing with a machine, people rated the same agent as less knowledgeable and less empathetic — even though nothing about the actual performance changed.

This is algorithm aversion — a well-documented psychological bias against machine decision-making that persists even when the machine objectively performs as well as or better than humans. It's not rational (in this case), but it's real and it affects behavior.

The important nuance: algorithm aversion isn't fixed. The study found customers with prior AI experience showed a significantly smaller disclosure penalty. As agentic AI becomes more familiar in everyday interactions, the aversion weakens — and the cost of honest disclosure falls with it. The 79.7% penalty figure from 2019 is almost certainly an overestimate of the penalty teams would face today, and will continue to shrink.

What this means for resolution rate

For support operations (as opposed to sales conversion), the relevant metric is resolution rate, and disclosure affects it indirectly through abandonment.

A meaningful share of customers disengage when they discover they're talking to AI. Survey data puts this in the range of a third of customers in some contexts, pushing abandonment on disclosed-AI interactions toward 25–30% versus 3–5% for human-fronted ones. An abandoned contact is an unresolved one, which quietly drags first-contact resolution down.

But — and this is the critical point — resolution rate is fundamentally a capability problem, not a labeling problem. An AI that resolves 80%+ of cases end-to-end does so regardless of what badge is on it. The abandonment effect is real but bounded: customers who stay experience the capability. The goal is to minimize abandonment through good disclosure design while maximizing resolution through actual capability.

These aren't in conflict. They reinforce each other: an AI that demonstrably resolves issues earns tolerance for the disclosure. An AI that stalls and deflects earns resentment of it.

What disclosure does to CSAT

The picture here actually flips in disclosure's favor.

CSAT tracks whether the issue got solved, not who solved it. Around 74% of users report higher satisfaction when a chatbot fully resolves their problem without a human handoff, and 87% report positive experiences with AI chatbots overall. Transparency helps CSAT because customers who know they're talking to an AI calibrate their expectations appropriately and judge the interaction more fairly — rather than measuring it against an implicit human standard.

What tanks CSAT isn't disclosure. It's an AI that lacks the context to resolve the issue, can't escalate cleanly, or forces the customer to repeat themselves when they do reach a human. Those are capability and handoff problems, not disclosure problems.

The practical implication: if your AI is genuinely good, disclosure protects your CSAT by setting appropriate expectations. If your AI isn't good, disclosure reveals the problem — which is useful information even if it's uncomfortable.

What disclosure does to NPS and long-term trust

This is where hiding AI creates the most serious risk.

Around 75–85% of consumers say they want to know when they're interacting with AI. 81% consider AI passing as human to be an ethical problem. These aren't fringe positions — they reflect a broad baseline expectation of transparency.

Concealing AI doesn't protect loyalty; it defers the damage. When customers discover — through a slip, a capability boundary, or external reporting — that they were interacting with AI they weren't told about, they experience it as deception. The discovery-after-the-fact effect on NPS and trust is substantially worse than honest upfront disclosure would have been.

Salesforce research adds a specific finding: 44% of consumers are more likely to use an AI agent when its logic is explained, and 45% when there's a clear escalation path to a human. Transparency and control aren't just ethical requirements — they're conversion drivers in the right context.

The legal dimension

This has moved from optional to required in significant markets. The EU AI Act's Article 50 transparency obligations — requiring that people be informed when they're interacting with an AI system — took effect on 2 August 2026. Other jurisdictions are moving in the same direction.

For any team serving EU customers, undisclosed AI isn't a strategy choice — it's a compliance risk. And beyond the legal requirement, the broader trajectory is clear: disclosure is becoming a baseline expectation globally, and building it in now is less costly than retrofitting it later.

The practical playbook for disclosing without paying the penalty

The research points toward a specific approach that preserves outcomes:

Let competence lead, not the disclaimer. Front-loading "Hi, I'm an AI" before the customer has seen any value primes algorithm aversion before the interaction has a chance to earn trust. Open with substance — address the customer's situation — and identify the AI clearly but without making the disclosure the first thing they process.

Make human escalation obvious and instant. The most important trust signal in disclosed AI interactions is that the customer can reach a human easily and quickly. "I'm in control" — I can escalate if I want to — converts the disclosure from "I'm stuck with a bot" to "I'm choosing to continue with the AI." That reframe changes how customers experience the interaction.

Invest in actual resolution. Every satisfaction and trust gain in the research data is downstream of the problem actually getting solved. Disclosure is most costly when the AI isn't capable. It's cheapest — potentially costless — when the AI resolves the issue competently. The leverage is in capability, not in disclosure timing.

Test rather than guess. The optimal disclosure wording, placement, and timing vary by channel, customer segment, and use case. Treating disclosure as something to optimize against real resolution and CSAT data — rather than a fixed script — lets you find the approach that works for your specific context.

The core reframe

The 79.7% conversion penalty from the 2019 study represents a specific condition: early, unearned disclosure of AI in a sales context before any value demonstration, among customers with limited prior AI experience, in 2019. That condition is increasingly rare as agentic AI becomes familiar and disclosure design improves.

The broader data supports a different conclusion: honesty about AI, paired with genuine capability and easy escalation, is an NPS and trust asset — not a liability. The penalty belongs to badly-designed disclosure, not to transparency itself.

What's been your experience with disclosure in practice — have you tested timing or framing variations, and did it move the metrics the way the research predicts?