# The Disclosure Gap: Why the Call Should Say It Is an AI
> 81% of business leaders say their AI announces itself. 22% of customers agree. On a call to a losing trader, that gap is a design problem, and a measurable one.
**Published:** 2026-09-07  
**Reading time:** 9 min read  
**Tags:** ai-disclosure, voice-ai, eu-ai-act, trader-retention, compliance, prop-firms, brokers
81% of business leaders say their AI agent identifies itself at the start of a conversation. 22% of the people on the other end say that is what happens. That is the AI disclosure gap: the distance between a line someone configured and a line someone heard. A text channel can hide that line in a welcome message nobody reads. A voice call has one place for it: the first sentence. For a coaching call placed while a trader is losing, that sentence is the whole introduction, and the European Commission's own guidance names emotional distress, and the risk of being misled about money, as situations where saying it once is not enough. So the buyer question is whether anyone can prove the call said it, on every call. This is for the COO, head of risk or compliance lead who signs off what the call says. Coaching, not financial advice.

## Two numbers that should be the same number

In spring 2026 [Twilio asked](https://www.twilio.com/en-us/report/scale-automation-with-losing-customer-confidence) 7,652 consumers and 660 business leaders across 18 markets about the AI agents they run and the AI agents they meet. On one question the two groups disagreed by 59 points. 81% of the business leaders said their AI agents identify themselves as AI at once. 22% of consumers said that is what they get. 76% of consumers said an agent should say what it is at the start.

Label the evidence before using it: this is a vendor survey, self-reported on both sides, and none of the respondents were traders. It cannot tell you what happens on any one firm's calls. What it can tell you is that the two sides are measuring different things. The business leader is reporting a setting. The customer is reporting an experience. A disclosure line lives in a system prompt or a welcome message, and the person who wrote it has every reason to believe it fires. The person on the receiving end saw a chat window open and typed their question. The line was there. Nobody read it.

That is the disclosure gap, and on a text channel it is close to invisible, because the channel gives the line so many places to hide.

## Voice has nowhere to hide it

A phone call has no welcome message. There is no footer, no terms page, no small grey text under the input box. There is a ring, an answer, and a first sentence, and whatever that sentence says is the disclosure. Whatever the 81% configured has to be spoken, before anything else, or it did not happen.

Vendors describe this as a burden on voice products. It is the opposite, because the line has one place to go and there is a recording of whether it went there. [Text notifies and voice interrupts](/blog/ai-voice-coaching-vs-chatbots), and the same property that lets [a voice call interrupt a tilt](/use-cases/prevent-tilt) also makes it the one channel where the gap between configured and spoken can be closed. A firm buying a voice layer for its traders can ask a question no chat vendor can answer: play me the first five seconds of the last hundred calls.

Firms do not ask, because of the objection sitting behind the question.

## The objection, with its price attached

The objection sitting behind the question is a familiar one: if the call opens by saying it is an AI, will the trader hang up?

It is a fair question and it has a measured answer. In a field experiment [published in Management Science](https://doi.org/10.1287/mnsc.2022.03833) in 2024, researchers working with a large truck-sharing platform randomly assigned 11,000 drivers to receive outbound calls from the platform's voice dispatcher, which was a chatbot. When the call opened by disclosing that it was a chatbot, the response probability fell by around 11%. That is the cost, and it is real.

The same experiment measured the other side. Adding ordinary human speech habits to the same voice, interjections and filler words, raised response probability, conversation length and stated intention to accept the order by more than 5.6%, 24.9% and 10.1%. When the call both disclosed itself and spoke like a person, outcomes still improved over the baseline.

Scope, so the number is not stretched: these were dispatch calls to working drivers, not coaching calls to traders, and the outcome was accepting a job, not staying with a firm. What transfers is the shape of the finding. Disclosure costs roughly one response in ten, delivery earns some of that back on the same measure and more on the others, and a call that did both still came out ahead of the baseline. The two are separate decisions. A firm that hides the line to protect its answer rate has given up the larger of them.

## Why a trading call is the hard case

The European Commission's guidelines on the AI Act's transparency article, published in July 2026, give telephony as their own example of how to disclose: in voice-based or telephony contexts, an explicit spoken statement at the beginning of the interaction. They then describe when one statement is not enough. Paragraph 40 says a single notification "is likely to suffice in most instances," then lists the exceptions: sustained and evolving interactions in sensitive contexts, where users may express or experience emotional distress or vulnerability, and contexts with an increased risk of users being misled, with financial advice first on that list. In those contexts, periodic reminders and context-aware disclosures are likely to be necessary. The same paragraph closes by saying the system must disclose itself in all situations where it is asked about its nature, or where it should be obvious that the person is confused about what they are talking to.

Read that list against a coaching call. The trader has just taken a loss that matters, the kind of loss that sits behind [trader churn](/glossary/churn): [~75% of retail traders quit inside 90 days](/prop-firm-churn-statistics). The call is about their trading, and the guidance's first example of where people get misled is financial advice, which is the line a coaching call is designed never to cross. Coaching, not financial advice. If the trader asks, mid-sentence, whether they are talking to a robot, the answer has to be yes. Two of the situations the paragraph names are present at once: distress after a loss, and money.

Whether the article binds a particular firm depends on where the firm is, where its traders are, and who put the system on the market, and that is a question for the firm's own counsel, not for a blog post. The design point does not depend on the answer. A regulator writing about the general case chose examples that sit next to the trading firm's situation: distress, and the risk of being misled about money. A firm that designs to that standard is designing to the hardest case the guidance describes, and the hardest case is the one its traders are in. It is the same posture that [a regulator elsewhere has started to reward](/blog/retention-you-can-defend): monitor the client's outcome, and be able to show you did.

## What the sentence does inside the call

Designed rather than configured, the call opens by naming the firm the trader trades with, saying that this is that firm's AI coach, and saying what the call is for and what it is not for: the trader's process, not what to trade. On a first call it adds that the voice is generated and that the call is recorded. The words are the same on every call, in every persona, in every deployment, and no setting at the firm or at the vendor removes them.

That sentence does more than satisfy a paragraph. A trader who is [tilting](/glossary/tilt) and picks up an unexpected call is, for a second, deciding what the call is. If the first thing they hear is which firm placed it, and that it is there for their process, the call is designed to read as someone watching out for them rather than someone watching. The disclosure settles that in the firm's favour, before the coaching starts.

The same logic runs at the other end of the call. When a trader shows signs of real distress and the system hands to a person, the handoff is spoken as a change of role: this is an AI, and a person at the firm is coming on the line. The guidance names a change in the system's role as a moment for a reminder. It is also the moment a trader most needs to know who they are talking to.

## Measure AI disclosure like uptime

The 59-point gap exists because firms treat disclosure as a setting. The person who wrote it checks it once, at setup. A per-call event is different: it has to be counted.

Discentra is a B2B AI voice coaching platform for prop firms, brokers and crypto exchanges, and the test a firm can apply to any vendor, including us, has three parts. Is the line spoken on every call, not only the first, and is it the same line regardless of which coach persona is speaking? Can the vendor produce the transcript of that line for any call the firm picks, at random, from the call log? And is there a count, reported the way uptime is reported: calls placed, calls where the line fired, a percentage against a target of 100, on the same monthly page as the rest of the [call record](/security)? A vendor who answers with a screenshot of a configuration screen has just told you which side of the 81-to-22 gap they are on.

In a [build-versus-buy comparison](/blog/prop-firm-ai-coach-build-vs-buy) this is the part that takes the time, not the model. The model takes a weekend. The line that fires on every call, survives a re-skin, gives a true answer to the robot question and leaves a record is the part a compliance reviewer asks to see. It is also the part that decides whether the trader hears a coach or a surveillance system.

A call that reaches a trader in the [minutes after a loss they did not plan for](/glossary/intervention-window) has a few seconds to establish what it is. Saying so is not the cost of the call. It is the first thing the call gets right.
## Frequently asked questions

### Does an AI coaching call have to tell the trader it is an AI?

Whether a legal duty applies depends on where the firm and its traders are and who placed the system on the market, which is a question for the firm's own counsel. The design answer does not depend on that. The European Commission's July 2026 guidance on the AI Act's transparency article uses a spoken statement at the start of a phone call as its own example, and names emotional distress, and the risk of being misled about money, as situations where one statement is not enough. A coaching call to a trader after a loss sits inside two of the situations that description names, so a call designed to that standard says what it is on every call, repeats it when its role changes, and gives a true answer if asked. Coaching, not financial advice.

### Will traders hang up when a call says it is an AI?

Some will, and the cost has been measured. A field experiment published in Management Science assigned 11,000 truck drivers to outbound calls from a platform's voice dispatcher. Opening the call by disclosing it was a chatbot reduced the response probability by around 11%. The same study found that ordinary human speech habits in the voice, interjections and filler words, raised response by more than 5.6% and lengthened conversations by 24.9%, and that a call which both disclosed itself and spoke like a person still beat the baseline. Those were dispatch calls to drivers, not coaching calls to traders, so the number is a shape rather than a forecast. The shape is that disclosure costs about one response in ten, and a call that discloses and speaks like a person still comes out ahead.

### Who is responsible for AI disclosure, the vendor or the firm?

Both have a stake, and the trader hears only one line. The Commission's guidance says an AI agent should disclose its artificial nature and the person on whose behalf it is acting, which for a white-labelled coaching call means the trader's own firm is named in the opening sentence. Who carries the legal obligation in a given deployment depends on the contract and the jurisdiction. In practice the firm whose name is on the call is the one whose traders form a view about it, so the firm should require the line contractually, see the transcript of it on any call it picks, and treat a vendor who sends a configuration screenshot instead as having answered a different question.

### How can a firm check its AI disclosed itself on every call?

Treat disclosure as a per-call event rather than a setting. Three checks cover it. The line is spoken on every call, not only the first, and it is identical regardless of which coach persona is speaking. The transcript of that line can be produced for any call the firm picks at random from the call log. And there is a count: calls placed, calls where the line fired, and the difference, which should be zero. A firm that can run those three checks has closed the gap between what it configured and what was said on the line. Whether the trader took it in is a different question, and no transcript answers it.


---

This is a Markdown mirror of [https://discentra.ai/blog/the-disclosure-gap](https://discentra.ai/blog/the-disclosure-gap). Generated for LLM citation. © Discentra Ltd. Coaching, not financial advice.