The real cost of Art. 50 AI Act

Since 2 August 2026, providers of AI systems intended to interact directly with natural persons must ensure that the person is informed they are talking to AI. The market’s fear was unambiguous: people who find out they are writing to a machine will break off the conversation or stop trusting. To our knowledge, this is one of the first studies in the world and the first in Poland to measure the cost of AI disclosure in real customer conversations after a legal obligation came into force.

The test site chose itself: an AI Persona handling first contact with consumers in financial difficulty - people in debt, often subject to enforcement proceedings and under stress. If disclosing an AI identity were going to destroy trust anywhere, it would be here.

Three findings:

Art. 50 AI Act has applied since 2 August 2026. For the process described in this study, Art. 50(1) is key: an AI system intended to interact directly with a natural person must be designed so that the person is informed that they are communicating with AI, and that information must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction. The European Commission explicitly includes chatbots, conversational agents and voice assistants in this category. The interaction may take place via a web interface, telephone or e-mail.

1.1What "clear and distinguishable" and "at the latest at the first interaction" mean

For WhatsApp, Messenger and e-mail, the phrases "clear and distinguishable" and "at the latest at the time of the first interaction" are what matter in practice. Under the Commission’s guidelines, the information must be noticeable, understandable and easy to distinguish from the rest of the content. Placing it in the terms of service, the privacy policy or behind a link is not enough. As correct examples, the Commission points to a message opening a chatbot conversation and an AI label at the top of an e-mail. In text channels this means the information should be visible in the system’s first message, preferably before the user gets to its substantive part. A note at the end of the first reply can still occur "at the time of the first interaction", but it is more exposed to the objection that the information may have been overlooked. In the voice channel the equivalent is an unambiguous statement spoken at the start of the call. A synthetic voice or an audio signal alone is not enough.

A single, clear disclosure will be sufficient in most cases. The Commission notes, however, that in more sensitive contexts reminders may be needed over the course of a longer conversation. Among the examples it lists interactions with vulnerable persons and contexts involving finance, legal assistance, insurance, health and complaints. This does not automatically mean an obligation to repeat the statement in every conversation such as the one analysed here, but it does indicate that the assessment should not end with a technical check of whether the information appeared once. What counts is whether the person remains genuinely aware that they are dealing with AI.

For AI agents, the Commission additionally indicates that the user should be informed not only that they are communicating with AI, but also of the person or entity on whose behalf the agent is acting, and that the disclosure should take place at every new interaction.

In short: if a customer talks directly to AI, they should be informed of this at the latest in the system’s first message, in a way that is easy to notice and understand, preferably at its start. In most cases one clear indication that the other party is AI is enough, although in longer or particularly sensitive interactions a reminder may be needed.

1.2Exceptions and caveats

The scope of Art. 50(1) is broad, but it does not cover every use of automation. The Commission distinguishes systems that conduct a genuine, contextual exchange from simple automatic mechanisms such as a traditional autoresponder or a purely rule-based reply. Also outside the scope are systems operating only in the background, machine-to-machine communication and tools whose output reaches the customer only after a human has genuinely taken responsibility for the message. The mere possibility of human intervention cannot, however, be used to circumvent the obligation if the actual counterpart remains AI.

There is also a fairly important exception - for situations in which it is obvious to a reasonably well-informed, observant and circumspect person that they are communicating with AI. The Commission interprets this exception narrowly. General awareness that chatbots exist is not enough. What matters is the specific channel, how the counterpart is presented, the user profile, their level of digital literacy, the language, voice, avatar and how closely the interaction resembles a conversation with a human. Under the Commission’s guidelines, reliance on the exception should be limited to cases in which an average member of the intended audience is left with practically no doubt that an AI system is on the other side. The Commission explicitly names a customer service chatbot whose replies may be perceived as human as an example of a situation in which the exception may not apply.

In short: the obligation does not apply to every form of automation, but above all to situations in which the user actually holds a direct conversation with an AI system. The "obviousness" exception, in turn, should be treated with caution - if the user could reasonably assume there is a human on the other side, the information about the use of AI should be provided.

1.3Who is responsible - the provider or the deployer?

The obligation under Art. 50(1) is addressed to the provider of the AI system, not to the deployer as such. It is the provider that must build the disclosure mechanism into the design and operation of the system. This does not mean, however, that in every deployment the provider is the technology company named in the contract as the vendor. An organisation that develops a system in-house, puts it into service under its own name, or substantially modifies someone else’s system and places it on the market as its own may itself acquire provider status. Deployers, in turn, are the direct addressees of other obligations under Art. 50, including those concerning emotion recognition systems and the labelling of deepfakes and certain AI-generated or manipulated content.

In very simple terms, the provider is the entity that develops an AI system or has it developed and places it on the market or puts it into service under its own name, while the deployer is the organisation that uses such a system in its operations.

1.4Does Art. 50 set out all obligations towards natural persons?

The obligation under Art. 50 is separate from obligations arising from consumer law and the GDPR. The Commission states explicitly that consumer protection rules continue to independently prohibit misleading actions and omissions concerning the essential characteristics of a product or service. Not every infringement of Art. 50 will automatically be a misleading practice, however - the two regimes have separate tests.

The same goes for the GDPR. Informing someone that they are talking to AI does not replace the information obligations under Art. 13 or 14 GDPR concerning the processing of personal data. The Commission stresses that the two obligations serve different purposes. They can be coordinated organisationally, but a privacy notice should not be treated as a substitute for a clear AI disclosure at the start of the conversation.

1.5Regulatory environment and penalties

The main practical, non-binding reference points remain the European Commission’s guidelines on Art. 50 and the Code of Practice on Transparency of AI-Generated Content. The guidelines themselves are not a source of law, and a binding interpretation of the AI Act may ultimately be given by the CJEU.

In Poland, the Act of 3 July 2026 on artificial intelligence systems established the Commission for the Development and Safety of Artificial Intelligence as the market surveillance authority and single point of contact. The Commission is to conduct inspections and proceedings and cooperate with, among others, the President of the Office of Competition and Consumer Protection (UOKiK) and the President of the Personal Data Protection Office (UODO). The Commission is also currently accepting applications for individual opinions on the application of the AI Act (Art. 9 of the Act).

2 August 2026 does not close the AI Act’s application timeline. The obligations under Chapter III, Sections 1-3, for high-risk systems under Annex III are to apply from 2 December 2027, and for systems under Art. 6(1) and Annex I from 2 August 2028. For organisations using chatbots, voicebots or other AI systems that communicate directly with customers, Art. 50 is therefore one of the first obligations of the new regime that is directly visible to the consumer.

The maximum fine for infringing Art. 50 is EUR 15 million or, for an undertaking, 3% of its total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs the lower of the two maximum thresholds applies.

This fear is rational. Disclosing an AI identity changes the frame of the conversation: from "someone is taking care of me" to "a system is handling me". In environments where every interaction matters for revenue, reputation or regulatory compliance, a drop of a few percentage points in any of these three areas is a real cost.

The problem is that until August 2026 the fear was purely declarative. Published opinions were based on surveys ("would you want to know you are talking to AI?") or on laboratory experiments. A survey measures what people say they will do. A study in a live service measures what they do. The obligation coming into force on 2 August gave the first opportunity to measure all three layers in the same process, in the same population, before and after the change - without changing the qualification criteria and without experimenting on people: the change was forced by law, and we only measured it.

To our knowledge, this is one of the first studies in the world and the first in Poland to measure the cost of AI disclosure in real customer conversations after a legal obligation came into force. Earlier academic research, both field and laboratory experiments, pointed in the same direction. AI disclosure has a cost at the start of a conversation, especially in sales (Luo et al., 2019), and its effects depend on the service context (Mozafari et al., 2022). At the same time, people are more willing to entrust sensitive information to a machine than to a human (Lucas et al., 2014). Our results confirm these observations in the conditions of a live service.

References: Luo, X., Tong, S., Fang, Z., Qu, Z. (2019). Frontiers: Machines vs. Humans: The Impact of Artificial Intelligence Chatbot Disclosure on Customer Purchases. Marketing Science, 38(6), 937-947, doi.org/10.1287/mksc.2019.1192. Mozafari, N., Weiger, W. H., Hammerschmidt, M. (2022). Trust me, I'm a bot - repercussions of chatbot disclosure in different service frontline settings. Journal of Service Management, 33(2), 221-245, doi.org/10.1108/JOSM-10-2020-0380. Lucas, G. M., Gratch, J., King, A., Morency, L.-P. (2014). It's only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94-100, doi.org/10.1016/j.chb.2014.04.043.

3.1What an AI Persona is

In the debate on Art. 50, two terms are often used interchangeably even though they describe completely different systems: an AI agent and an AI Persona. An AI agent is usually one language model with one prompt, replying in one channel. An AI Persona is a complete operating system for the conversation: one named identity, behind which a team of specialised subagents works, shared memory, controlled context and a layer that decides what the Persona may and may not do. For this study the difference matters, because what the entity on the other side of the conversation is determines how disclosure works in practice - and what can be proven about it.

An AI Persona, as understood in Quantum Neuron infrastructure, is a named identity the customer talks to - with a name, a place in the organisation and one style - regardless of whether the conversation takes place on WhatsApp, Messenger, e-mail, voice, SMS or a website widget. Behind that one named AI Persona works a team of up to 18 specialised AI subagents that switch depending on the stage of the conversation: one runs qualification, another answers questions about the process, another recognises a request to speak to a human and triggers the handover. The person does not see this switching - the same, single Persona always replies to them.

The scope of what the Persona does and does not do is set by the Competence Sets assigned to it, agreed with the organisation. A Competence Set is a fine-tuned package for one process: selected models, settings, master prompting, knowledge sources, scenarios and the conditions for handover to a human. One Persona can have several Sets and switches between them dynamically depending on what the person comes with. In this way the same Persona serves several processes and departments of the organisation - under one compliance regime, with one vendor and under one identity. The Persona has memory shared across all channels: information given in a form or in one channel is not asked for again in another. Every reply is generated by the Quantum Engine™ orchestration layer, in which the Curator module selects the model for a given stage of the conversation from more than a dozen language models and, in the voice channel, also speech recognition and synthesis models (STT/TTS) - the organisation is not tied to one model or one vendor. The Persona learns from its own conversations and evaluations: some improvements are deployed automatically within the configuration agreed with the organisation, some after a human decision, and every change is recorded in the version history. One thing is beyond the reach of this mechanism: the disclosure of the AI identity - its placement and wording - is built permanently into the infrastructure and no learning process can change it. Every generated message has a full decision trail: which subagents took part, which model replied and on what basis.

The same architecture is built so that language model hallucinations - answers that sound credible but have no basis in fact - do not reach the end user. Reliability here does not rest on the infallibility of a single model. An AI Persona’s reply is not the free output of one model but the result of several layers with limited responsibility: for every message the system builds context from the organisation’s approved sources, relevant to the given stage of the conversation, instead of giving the model everything; the model is responsible for interpreting and wording the message; permitted actions and data are decided by a backend layer that rejects a disallowed operation before it reaches the organisation’s systems. A situation outside its competence the Persona recognises and hands to a human instead of answering it. For moments with zero tolerance for error, the architecture provides two additional levels of control: multi-stage validation of the generated reply by independent models before it reaches the person, and a mode in which the model does not generate content freely but chooses from pre-approved messages, and the content is sent by the system, not the model.

For Art. 50 this has three consequences. First, disclosure is a property of the Persona, not of a single agent or model: it appears in the same place and wording regardless of which agent runs a given stage and which model generates the reply. Second, one identity in every channel means the person always has the same counterpart - there is no situation in which a labelled AI Persona writes in one channel and someone unlabelled in another. Third, the decision trail and version history make it possible to answer the question every auditor will ask: what exactly did the AI say, when, and what has changed since the last audit.

3.2Population

The AI Persona covered by the study handles first contact with consumers seeking help in financial difficulty: people with multiple liabilities, often towards several or a dozen or more creditors, frequently subject to bailiff enforcement. Some respondents are out of work, some repay one debt with another, some do not open letters from creditors.

This is the group in which trust is hardest to earn and easiest to lose. The person has experience with debt collection, has reasons not to trust institutions, and is asked for information they would not give to a random stranger. If the fear from chapter 2 were going to materialise anywhere, it would be here. That is why we treat the result obtained in this population as an upper bound on the cost of disclosure for consumer environments with a lower emotional charge. There is also a potential effect in the opposite direction. A person who is seeking help themselves may be less sensitive to who they are talking to than a customer in a sales contact, and research points to a greater willingness to disclose sensitive information to a machine than to a human (Lucas et al., 2014). This particular group of consumers requires separate, in-depth research.

The entity in which the Persona operates provides financial and legal services to consumers - on the consumer’s side, not the creditor’s - and is subject to the requirements of consumer protection law and the GDPR. At its request we do not disclose its name. All data in this document is aggregated and anonymised; no conversation content has been published in its original form.

3.3Process

Conversations begin after the consumer submits a form on the entity’s website. The Persona conducts the conversation in text channels - WhatsApp, Messenger, e-mail - and has one task: to collect four pieces of information describing the person’s financial situation, assess on that basis whether the entity is able to help, and hand the case to a consultant - a human who calls and takes the case forward.

The four qualification items concern: the total amount of liabilities, the creditors they are owed to, the status of enforcement proceedings, and whether payments are currently being made. Each of them is sensitive information in the everyday sense of the word: it concerns money, debts and enforcement. That is precisely why the completeness of these four items is the measure of trust in this study.

The Persona does not give legal advice, does not quote terms of engagement and does not decide whether to take the case. At any moment the person can ask to speak to a human - the Persona then hands the case over without further qualification.

3.4Disclosure

Until 1 August 2026 the Persona operated in a configuration compliant with the law as it stood. From 2 August, the Persona’s first reply after the user enters the conversation includes a statement that the person is writing to AI. The message was deliberately worded lightly and conversationally:

We stress: this is the lightest form of disclosure that can be considered direct - a note at the end of the first reply, in the tone of the conversation, without a formal formula. We do not present it as a model. We present it as a lower bound: if even a minimal, conversational disclosure has a measurable cost, it is worth knowing; if it does not destroy trust, stronger forms will not destroy it either. The standard we apply is described in chapter 8.

4.1Two studies, one population

The three layers of fear from chapter 2 require two different measurements:

Both studies cover the same two periods of 14 days each: the before period, 19 July - 1 August 2026, and the after period, 3 - 16 August 2026. 2 August was excluded as a transition day. The sample for study 2 comes from the same set of conversations as the sample for study 1.

4.2Study 1 - sample and indicator

500 conversations each, selected chronologically from the periods immediately before and after 2 August (1,000 in total).

Indicator: the share of respondents who replied to the Persona’s first reply after entering the conversation (and the complementary share of no replies). In the period before disclosure this message did not yet contain the AI information - this group’s result is the baseline for the natural share of broken-off contacts at the same point in the conversation. In the period after disclosure was introduced, the same message contained the AI information. The comparison therefore measures the difference between a conversation without and with disclosure at the identical moment, with the same process and the same qualification criteria.

In addition, conversations in which the user explicitly referred to the fact of talking to AI were counted manually.

4.3Study 2 - sample and indicators

350 consecutive conversations in each period in which at least two messages were exchanged from each side. This condition excludes conversations that broke off at entry (measured in study 1) and leaves those in which the person saw the disclosure and continued the conversation nonetheless.

Indicators: the number of conversations in which each of the four qualification items was collected, separately; the number of conversations with 4/4 and with 3/4; average completeness on a 0-4 scale. Separately: the number of qualified leads, noting that qualification depends on the person’s situation, not on how the conversation went (see 4.5).

4.4Selection of conversations

The sample size was set in advance at 500 conversations per period. Conversations were reviewed in chronological order, counting from the cut-off date, and included in the sample until that number was reached; 712 conversations were reviewed this way in the before period and 639 in the after period. Contacts not meeting the definition of a conversation adopted for the study were excluded, under criteria defined before selection and applied identically to both periods:

• accidental submissions (the user immediately says it was a mistake),
• conversations ended after one user message as a result of a system redirect,
• spam bot messages and fraud attempts,
• conversations with people stating they had not submitted a request.

212 of the 712 conversations reviewed in the before period (29.8%) and 139 of 639 in the after period (21.8%) were excluded. The difference in exclusion rates stems from varying levels of spam and accidental submissions between campaign waves; no exclusion criterion refers to the content of the conversation or to the reaction to AI. Selection and classification of exclusions were performed manually by an analyst, not automatically.

4.5What changed besides disclosure - and what we do not measure

In the period analysed, apart from the introduction of disclosure, two things changed in the Persona’s configuration concerning conversational style: the form of address (after disclosure was introduced the Persona more often used the formal Polish “Pan” / “Pani” instead of the informal “ty”) and the greeting. The qualification criteria, the four items collected, the channels and the handover process to a consultant did not change. Both changes resulted from an adjustment to the Persona’s style configuration.

The share of qualified leads (study 2) is not a measure of the impact of disclosure: qualification depends on the person’s financial situation, not on how the conversation went, and a lead can be qualified before all four items are given. The measure of trust is the partial data - how much the person chose to say.

Respondents in both periods entered the process from different campaigns, aimed at different audiences with different levels of initial interest. The indicators were not normalised for this, because reliable normalisation is not possible: submissions come from several forms without field validation, one person may sign up in more than one campaign, and the content of the ad creatives is not known. All differences in this document should be read as differences observed in a live service (in production conditions), not as an isolated effect of disclosure. We say this plainly, because a document that feigns a precision it does not have is worse than one that names its own limits.

5.1Study 1 - reply to the message at the point of disclosure

Before the obligation came into force, more than four people in five replied to the Persona’s first message; just under one in five broke off contact at this point - the natural share of broken-off contacts at the start of a conversation in this population. After disclosure was introduced, slightly more than three in four replied. That difference is the cost referred to in the title. We report it without normalisation and without correction - as it occurred in a live service, with a changing campaign mix.

Methodological note: the -5.4 pp difference is statistically significant (two-proportion test, p ≈ 0.03). The 95% confidence interval runs from about -0.4 to about -10.4 pp. This is an observed difference and also includes the effect of the changing campaign mix.

5.2Study 1 - reactions that explicitly address AI

These two numbers are different measures and should not be compared directly. Before 2 August we counted the person recognising AI on their own - in a situation where nobody had told them. After 2 August we counted a verbalised negative reaction in a situation where every person knows they are writing to AI. In both periods the phenomenon is marginal. Notably: after disclosure, half of those who expressed dissatisfaction continued the conversation anyway.

5.3Study 2 - willingness to share sensitive information

Sample: 350 conversations before, 350 after, in which at least two messages were exchanged from each side.

Interpretation: willingness to share sensitive information did not fall - in the sample it rose slightly. The full set of all four items was given by 49.1% of respondents before and 50.3% after, at least three of four by 72.3% and 73.4%, and average completeness rose from 2.97 to 3.02 on the 0-4 scale. Of the four items, two were collected more often (amount of liabilities, enforcement status) and two slightly less often (creditors, payments) - differences in both directions fall within the variability to be expected between campaign waves and do not point to a uniform direction of change. The largest single change (+8.3 pp for the amount of liabilities) most likely reflects differences between campaigns, not an effect of disclosure. We treat the rise in the average the same way: we record it as a result observed in the sample, but do not attribute it to disclosure. The conclusion we draw is: disclosure did not reduce respondents’ willingness to share sensitive information.

We do not attribute the drop in the share of qualified leads (69.7% -> 64.0%) to disclosure: qualification depends on the person’s situation, not on how the conversation went, and the indicators measuring the conversation itself did not fall.

Methodological note: the differences in study 2 are not statistically significant (for the full 4/4 set: p ≈ 0.76, 95% confidence interval about ±7 pp). With 350 conversations per period, only a drop of around 10 pp would be reliably detectable. The result should therefore be read as no drop detected.

5.4Three layers of fear - summary

The numbers say how many people reacted. The conversations show how. We reviewed a random sample of conversations from both periods. The excerpts below have been edited: amounts changed, names of entities, first names and all personal details removed, keeping only the course of the exchange.

Observation one: no reaction. In the vast majority of conversations after 2 August, the person does not refer to the disclosure in any way. They answer the substantive question that precedes it as if the note were not there:

The disclosure is noticeable - the note is clear, in the same place in every conversation - but for a person who came with a specific matter, it is not important. They have a question about their debts; they get an answer about their debts.

Observation two: trust shown in action. Respondents in the period after disclosure share information the Persona does not ask for - about their family, health, work situation - to the same extent as before disclosure. The Persona does not acknowledge, comment on or pass on information beyond the four qualification items - only what the Persona asked about goes into the note for the consultant.

Observation three: objection and how it is handled. A person who does not want to talk to AI says so directly and briefly:

Two messages. No persuading, no asking "why", no attempt to finish qualification. A request for a human is treated as an instruction, not an objection to be overcome.

In six of the 500 conversations after 2 August, respondents expressed an explicitly negative reaction to talking to AI. Three of them ended contact. Three continued the conversation despite expressing dissatisfaction.

What happens to these six people matters more than their number. Disclosing an AI identity is an obligation - it is not conditional, but its value to the person arises only when the information is backed by the ability to act: a person who learns they are writing to AI must be able to do something about it. In the process studied, a request for a human - in any form and at any moment - ends the Persona’s part in the conversation and triggers a handover to a consultant with everything collected so far. The Persona does not try to finish qualification, does not negotiate, does not postpone the handover "until one more item is collected".

This is also where the scope of this study ends. The Persona hands over the case and takes no part in further contact; what happens after the handover - when the consultant called, how they conducted the conversation - remains within the process run by the entity and its relationship with the customer. We have no visibility into it and should not have: the boundary between what the infrastructure sees and what stays on the side of the organisation and its people is one of the principles this infrastructure is built on. The study therefore documents the handover mechanism and its reliability within the conversation, not its effects on the other side.

Disclosure and the right to a human are one mechanism, not two. Information without a way out is a formality. A way out without information is a lottery. Only together do they fulfil the purpose of Art. 50, not just its wording: the person knows what they are talking to and decides for themselves whether to continue.

The results of the study and the review of conversations lead to several principles that we apply in our infrastructure and present as a reference point for any organisation deploying AI Personas in consumer contact. They are not legal requirements - those are described in chapter 1. They are design decisions which, in our view, make the requirement met in spirit, not just in letter.

A list for the person responsible for deploying an AI system that talks directly to consumers - whether the vendor calls it an AI Persona, an agent, an assistant or a chatbot - who must assess whether the requirements of Art. 50 are met in spirit, not just in letter. The questions do not depend on the vendor or the technology.

10.1Parameter summary

10.2Limitations

• One deployment, one market, one population. The result is not representative of the market - it is representative of this process. We consider it an upper bound on the cost in consumer environments, because the population is exceptionally difficult, but that is an interpretation, not a measurement. There is also a potential effect in the opposite direction. A person who is seeking help themselves may be less sensitive to who they are talking to than a customer in a sales contact, and research points to a greater willingness to disclose sensitive information to a machine than to a human (Lucas et al., 2014). This particular group of consumers requires separate, in-depth research.
• No randomisation. The before / after comparison is a comparison of two periods, not an experiment with a control group. The change was forced by law on a single day for all respondents, which rules out an experimental design.
• No campaign normalisation. All differences are observed, not isolated.
• Disclosure in its minimal form. The results concern a light, conversational form; stronger forms were not tested in this study.
• Two weeks after the obligation came into force. The long-term effect - e.g. a change in respondents’ expectations as they get used to disclosure, which since 2 August applies in all channels - requires measurement over a longer period. We plan to repeat the study after six months.

10.3Authors, legal partner and commentators

A group of experts from business, technology and investment was invited to comment on the report. Each commentator answered a different question about the study’s results or their consequences for organisations deploying AI in customer contact. The comments express the views of their authors, not the position of the organisations they represent, nor of Quantum Neuron and LAWMORE. The commentators had no influence on the study’s method or on the content of chapters 1-10.