Companies must tell customers by the first direct interaction that they are communicating with an AI system, unless this is already obvious from the context. The notice belongs at the start of a call, in the opening chat view, or before an AI feature in a customer portal. It should state what the system handles and how to reach a person.
Legal status reviewed on July 14, 2026. This article provides operational guidance and is not a substitute for legal advice. The European Commission guidance published in May 2026 was still presented as draft guidance on the review date.
When must customers know that they are communicating with AI?
Article 50(1) of the EU AI Act applies to AI systems intended to interact directly with natural persons. The provider must design and develop the system so that the person is informed that the interaction is with an AI system rather than a human. The exception applies only when this fact is already obvious to a reasonably well-informed, observant, and circumspect person, taking the circumstances and use context into account. The transparency obligations become applicable on August 2, 2026.
Use AI with clear rules and responsibilities
KrambergAI helps companies establish practical AI compliance structures for internal rules, data handling, approvals, responsibilities and responsible use in daily work.
Structured guidance · Responsible implementation · Made in Germany
For a mid-sized service organization, this can cover far more than a conventional website bot. It may include an AI phone receptionist, an after-hours service line, a conversational appointment assistant, an automated warranty intake, a parts inquiry assistant, a digital quotation intake, or an interactive feature within a customer portal.
An HVAC contractor might use an assistant to collect the property address, equipment type, fault symptoms, and callback number. An electrical contractor may ask for the site, circuit or device involved, urgency, and availability. A construction or traffic-management company may collect project location, work window, scope, responsible contact, and requested documentation. A marina or yacht service provider might record berth number, vessel type, engine details, and the requested maintenance work.
In these situations, the customer is not simply completing a static form. The system interprets unrestricted language, identifies the request, asks follow-up questions, and changes its response based on the conversation. A natural voice, full sentences, and a personal greeting can make the artificial nature of the interaction difficult to recognize.
Businesses should therefore use the “obvious from context” exception cautiously. A small icon, a product name, or a generic assistant label may not be sufficient when the overall interaction resembles a human service conversation. Germany’s Federal Network Agency states that individuals must generally be informed about the interaction unless the AI nature is already evident from the context.
Who is responsible for providing the AI notice?
The legal distinction between provider and deployer is particularly important under Article 50. Paragraph 1 formally addresses the provider of the interactive AI system. The provider is not necessarily the developer of the underlying language model. It may also be the organization that develops, commissions, brands, or puts the complete customer-facing system into service under its own name.
A mid-sized company using an off-the-shelf phone assistant from a software vendor will often be a deployer. The vendor should provide the technical mechanism for presenting the notice. The deploying company should still verify that the notice is enabled in its implementation, appears at the required point, uses wording that matches the actual workflow, and remains active after configuration changes.
A different assessment may apply to proprietary or white-label solutions. If a company commissions a custom assistant, determines its intended purpose, places its own brand on the service, and puts it into operation under its own name, it may have to assess whether it has become the provider of that system.
This question cannot be answered by looking only at the foundation model. The complete solution may include a voice platform, language model, company knowledge base, CRM connection, routing logic, booking workflow, and automated actions. The organization controlling that combined product may hold responsibilities that differ from those of the model developer. The Commission’s draft guidance and published legal analyses emphasize this separation between provider and deployer roles.
Contracts should therefore assign responsibility for the technical notice, the approved customer wording, testing after releases, accessibility, logs, and evidence that the notice was presented. The vendor should not assume that the customer will add the message, while the customer assumes that the platform handles it automatically.
From the customer’s perspective, the interaction takes place on the company’s phone number, website, or portal. A dispute between the company and its technology provider does not improve the customer experience. Technical responsibility and operational ownership need to work together.
Where is the boundary between an AI assistant, automation, and a human representative?
Not every automated process is an AI system, and not every use of AI creates a direct interaction between a person and AI.
A conventional telephone menu normally follows predefined rules: “Press one for service, press two for billing.” If the system does not interpret unrestricted language, generate answers, or adapt the conversation through AI, automation alone does not create an Article 50(1) notice requirement.
The situation changes when callers describe their requests in their own words and the system determines the service category, urgency, customer record, or next action. A deterministic workflow may still run after the information is captured, but the customer-facing intake has already involved direct interaction with AI.
The same distinction applies to online interfaces. A form that moves through fixed fields is not necessarily an AI chatbot. A system that understands free-text requests, identifies intent, generates individualized replies, or creates its own follow-up questions is much more likely to qualify as an interactive AI system.
Human customer service supported by AI is different. A representative may use AI to retrieve equipment records, summarize a service history, identify a relevant manual, or draft a response. When the representative reviews the information, decides what to say, and communicates with the customer personally, the Commission’s draft interpretation generally does not treat the customer as interacting directly with the AI system.
Hybrid processes require closer examination. An email may be drafted entirely by AI but approved by a person before sending. A live-chat representative may copy an AI-generated response without meaningful review. A voice assistant may handle the first half of a call and then transfer it to an employee. The business should examine who receives the input, who interprets it, who generates the response, and whether a person can intervene before the message reaches the customer.
Calling a process “automated” does not settle the issue. The actual architecture and customer journey matter more than the internal product label.
At what point must the notice appear?
Article 50(5) requires the relevant information to be provided no later than the first interaction or first exposure. It must be readily noticeable, distinguishable from surrounding information, and presented in an accessible format. A notice delivered at the end of the conversation, hidden in general terms, placed only in a privacy policy, or available solely on a separate legal page is too late for the interactive moment.
On a phone call, the notice belongs in the opening statement before the AI assistant asks for the customer’s issue, name, address, account number, or service details. A short company greeting can come first, but the caller should understand the nature of the conversation before providing substantive information.
In a website chat, the information should be visible when the chat window opens or appear as the assistant’s first message. A neutral chat icon that reveals the AI identity only after several exchanges is not a dependable implementation.
Within a customer portal, the placement depends on the scope of the AI feature. If the entire experience is conversational, the opening view should identify the AI system. If the portal also contains invoices, documents, project status, and other conventional functions, the notice should appear directly beside and within the specific AI module.
A portal button labeled “Ask the AI Assistant” can contribute to the disclosure. The module should still provide a brief opening explanation before the customer enters contract information, fault descriptions, technical documents, or personal data.
Returning users also need appropriate notice. A disclosure shown during initial registration may no longer be memorable when a customer uses a new AI feature months later. A persistent module label combined with an opening message is more dependable than a one-time onboarding pop-up.
The Commission’s draft guidance discusses text, visual indicators, and spoken messages depending on the communication channel. For voice systems, a spoken statement at the start of the interaction is particularly relevant. Generic references to an “assistant,” technical descriptions of language models, links to terms, or machine-readable information alone are not treated as adequate customer-facing notice in the draft interpretation.
What should an AI phone notice say?
“This call is handled by AI” is memorable, but it can create uncertainty. The statement does not explain whether AI is merely collecting the request, conducting the entire call, recording the conversation, making a decision, or committing the company to an outcome.
A stronger opening identifies the conversational partner, describes the task, and explains how human support is available.
General phone reception:
“Welcome to [Company]. You are speaking first with our AI assistant. It will collect your request and route it to the appropriate team.”
A field-service organization can make the scope more specific:
Service and repair intake:
“Welcome to the [Company] service line. Our AI assistant will collect your service issue, equipment details, and callback number. It can transfer the request to our service team when needed.”
An after-hours line may use:
Outside office hours:
“Our office is currently closed. An AI assistant will collect your request and contact details so our team can prepare a callback.”
When live transfer is available, it can be introduced at the beginning rather than after the customer becomes frustrated:
Phone assistant with human transfer:
“You are speaking with our AI assistant. Say ‘representative’ at any time to be transferred to our customer service team.”
The wording must reflect the system’s actual capability. An assistant should not promise a transfer if it can only create a callback ticket. In that case, the message should say that the system will prepare a callback request.
The opening should remain brief. AI identity, data protection information, and call-recording notices serve different purposes. If the call is recorded, fully transcribed, analyzed for additional purposes, or retained beyond ordinary case handling, the company must separately assess the applicable privacy and communications requirements. An Article 50 notice does not replace those disclosures.
The assistant should also avoid overstating authority. If it can suggest appointment windows but cannot make a binding booking, the opening or later confirmation should state that the service team will confirm the appointment. If it collects quotation information but cannot issue a contractually binding price, it should not tell the customer that the quotation is final.
What should a chatbot transparency notice say?
A chatbot notice should be available without requiring the customer to search for it. Labels such as “Service,” “Online Help,” “Virtual Colleague,” or “Digital Assistant” do not reliably explain whether the response comes from a person, a scripted workflow, or an AI system.
An effective opening message may state:
General website chatbot:
“You are chatting with [Company]’s AI assistant. It answers questions using approved company information and can pass your request to our team.”
For appointment intake:
Appointment and callback assistant:
“This conversation is handled by an AI assistant. It will collect your preferred appointment and contact details. Our team will provide the final confirmation.”
For quotation preparation:
Quotation request:
“Our AI assistant will help you assemble the information needed for your request. It does not issue a binding quotation. An employee will review the details and contact you.”
For technical support:
Technical service chat:
“You are communicating with an AI assistant. It will collect the fault description and equipment details, but it does not replace an on-site technical safety assessment.”
A short notice can be supported by a detailed information page describing data processing, retention, sources, contact options, and escalation. The link should supplement the opening message rather than replace it.
The interface should not rely solely on color, an animation, or an icon. The information should be available to screen readers and remain understandable when visual styling is removed. Messages written by the AI assistant and messages written by an employee should also remain distinguishable throughout the conversation.
If the system switches between AI and human support within the same window, the interface should indicate the current speaker. A generic company avatar that remains unchanged throughout the transition can make that distinction difficult for customers.
How can an AI notice be placed inside a customer portal?
A customer portal often combines AI and non-AI functions. Customers may download invoices, review project status, access service reports, submit documents, confirm appointments, and use an AI knowledge assistant. A broad notice on the login page may therefore describe too much while still failing to inform the customer at the actual point of interaction.
The AI feature should be identified directly:
Portal knowledge assistant:
“AI assistant for questions about your approved project and service documents.”
When the module opens, a second layer can explain its role:
“This feature uses AI to locate and summarize information from the documents available to your account. Verify technical, contractual, and scheduling details before relying on them.”
A service-intake feature may say:
Digital service intake:
“An AI assistant will structure your information and create a service ticket. Dispatch and appointment confirmation remain with our service team.”
A project portal for construction, traffic management, engineering, or technical services should not imply that the AI can approve change orders, safety measures, project milestones, prices, or schedule changes when those decisions remain with the responsible project manager.
A suitable notice is:
Project information assistant:
“The AI assistant answers questions using the project information available in this portal. Your project manager continues to confirm approvals, change orders, and schedule changes.”
The notice can appear in the module header, immediately above the input field, and in the first system message. This repetition connects the function name, the customer’s action, and the beginning of the interactive session.
For portals used by subcontractors, property managers, dealers, or external technicians, the wording should also reflect the authorized data scope. The assistant should not suggest that it has access to the company’s entire knowledge base when it can only retrieve documents released to that account.
Which types of wording should businesses avoid?
Businesses should avoid language that disguises the artificial nature of the interaction. “Your digital contact,” “online service team,” “virtual colleague,” or a human first name without further explanation may lead customers to believe that an employee is communicating with them.
Purely technical wording is also ineffective for this purpose. “This system uses a large language model” identifies a technical component without telling the customer whether the current response is being generated by AI.
The notice should not exaggerate the assistant’s authority. A system that only collects a complaint should not claim that it is “processing the complaint” if no substantive review occurs. A chatbot showing available appointment windows should not announce a confirmed booking when human approval is still required.
Broad disclaimers can be equally unhelpful:
“All answers may be wrong. Use this service at your own risk.”
That sentence does not explain the system’s role, the appropriate use of its answers, or where the customer can obtain a binding response. A more useful limitation is:
“The AI assistant supports general inquiries. An employee confirms binding technical, contractual, and pricing information.”
The opening should also avoid unnecessary technical detail. Customers do not need a description of model parameters, hosting architecture, retrieval technology, or orchestration layers before submitting a callback request. They need to know that AI is involved, what it is doing, and how the process continues.
A human name is not automatically prohibited, but the name should be paired with an AI designation. “Anna, [Company]’s AI assistant” is less likely to be mistaken for an employee than “Anna from the service team.”
How do calls, chats, and customer portals compare?
| Customer interface | Typical function | Appropriate notice point | Practical wording | Human transition |
|---|---|---|---|---|
| AI phone assistant | Capture requests, classify cases, prepare callbacks | Opening statement before the first substantive question | “You are speaking with our AI assistant.” | Voice command, live transfer, or stated callback process |
| Website chatbot | Answer questions, collect leads, structure service requests | Open chat window and first assistant message | “You are chatting with [Company]’s AI assistant.” | Visible transfer button or in-chat handoff |
| Customer portal | Search documents, explain status, create service tickets | Directly beside the AI function and when the dialog opens | “This feature uses AI to process your request.” | Contact route to the assigned service or project team |
| Fixed phone menu | Route callers through keypad choices | Article 50(1) is generally not triggered merely by deterministic automation | “Select the department you need.” | Rule-based routing |
| Human representative using background AI | Search knowledge, summarize records, draft replies | Usually no Article 50(1) notice when the employee conducts the communication | Do not present automated output as an independent human speaker | Employee remains responsible |
The table is an operational guide rather than a substitute for assessing the individual system. Hybrid services require particular attention to whether the customer is interacting directly with AI or with an employee who is merely using an internal tool.
How should a handoff from AI to a person work?
A handoff should include more than technical routing. The customer should know when the AI interaction ends and a person begins handling the conversation.
On a phone call, the assistant may say:
“I will now transfer you to Maria Schneider on our service team. The information you have already provided will be available to her.”
In chat, a visible system message can state:
“A customer service representative is taking over from this point. Your previous information will remain attached to the case.”
When no live transfer is available:
“I have created a callback request using the information you provided. A team member will contact you during service hours.”
This transition reduces repeated questions and prevents customers from assuming that an automated statement constitutes final approval. It also helps the employee understand which statements came from the system and what still requires professional review.
Escalation rules should reflect the industry. An HVAC assistant must respond differently to a potential gas leak, major water loss, or heating failure than to a routine maintenance request. A traffic-management service needs rules for damaged barriers, an unsafe work zone, or an urgent change to traffic control. A yacht service assistant may need special handling for fuel leaks, bilge-water alarms, electrical hazards, or a vessel taking on water.
The transparency notice should therefore be linked to the operating model. Customers need to understand not only that AI is involved, but also when the assistant is appropriate, when a person takes over, and when an emergency contact is required.
What should a business document for implementation?
The customer-facing sentence is only one part of compliance. For each digital interface, the company should document whether direct AI interaction occurs, who acts as provider and deployer, where the notice appears, which wording is approved, and how the configuration is tested.
The record should also describe actual system authority. Can the assistant create a binding appointment? Can it open a work order, change a delivery address, approve a return, provide pricing, or access contract documents? Does it only create a note for a callback? Can it hand the conversation to an employee?
System changes should trigger reassessment. A fixed form may later gain generative answers. A phone assistant that originally collected only names and callback numbers may later gain access to customer records, generate pricing statements, or schedule field technicians. The previous notice may no longer describe the expanded role.
Testing should be performed from the customer’s perspective. The business should verify that the opening message is played on new calls, remains visible on mobile chat interfaces, works with assistive technology, matches all supported languages, and does not disappear after a software update.
Companies should also test the “obvious from context” assumption with actual users. Internal project teams already know that a system is AI-based and may underestimate how human the interaction appears to a first-time customer.
The Commission’s public guidance was still identified as draft guidance on the review date and is not legally binding. It nevertheless provides an important indication of how timing, accessibility, media choice, and user perception may be assessed. Final interpretation remains with the competent authorities and courts.
Bring AI into daily operations in a structured way
The KrambergAI AI Introduction helps companies select suitable use cases, prepare workflows and integrate AI solutions into everyday operations in a controlled and practical way.
Structured implementation · Practical guidance · Made in Germany
Why does the wording matter commercially as well as legally?
Transparency affects whether customers accept the interface and whether they complete the service process.
In Salesforce’s international State of the Connected Customer research, 72 percent of customers said it was important to know whether they were communicating with an AI agent. Separate Salesforce research found that 45 percent would be more likely to use an AI agent when a defined escalation path to a person was available.
Deploying an assistant does not automatically improve the experience. In PwC’s 2025 Customer Experience Survey, 30 percent of consumers said AI had made getting customer service assistance better, while 38 percent said it had made the experience worse.
The practical conclusion is that the opening notice should do more than disclose the use of AI. It should tell the customer what the assistant can accomplish, where its authority ends, and how the customer can obtain human assistance.
This can reduce friction. A caller who knows that the assistant is collecting information for a repair ticket will not expect a final technical diagnosis. A portal user who is told that an employee will confirm the appointment is less likely to treat a proposed time as a binding commitment.
A well-designed notice can therefore become part of the service promise. It establishes expectations before misunderstandings occur and supports a more efficient handoff when the request exceeds the assistant’s role.
Which sources support the figures used in this article?
Salesforce: State of the Connected Customer
https://www.salesforce.com/eu/resources/research-reports/state-of-the-connected-customer/
Salesforce: AI Connected Customer Research
https://www.salesforce.com/news/stories/ai-customer-research/
PwC: Customer Loyalty with Agentic AI
https://www.pwc.com/us/en/technology/alliances/library/salesforce-customer-experience-survey.html
Which resources provide useful further reading?
EUR-Lex: Regulation (EU) 2024/1689, the Artificial Intelligence Act
https://eur-lex.europa.eu/eli/reg/2024/1689/oj
European Commission: Draft Guidelines on Article 50 Transparency Obligations
https://digital-strategy.ec.europa.eu/en/library/draft-guidelines-implementation-transparency-obligations-certain-ai-systems-under-article-50-ai-act
European Commission: Guidelines and Code of Practice on Transparent AI Systems
https://digital-strategy.ec.europa.eu/en/faqs/guidelines-and-code-practice-transparent-ai-systems
Does every chatbot need an AI notice?
No. A static contact form or rule-based decision tree is not necessarily an AI system merely because the process is automated. When the chat interprets unrestricted language, identifies customer intent, generates individualized answers, or creates context-dependent follow-up questions, the business should assess whether direct AI interaction occurs and whether Article 50(1) applies.
Is the label “digital assistant” sufficient?
The term is often too broad. A digital assistant could be a person, a scripted workflow, or an AI system. An explicit label such as “AI assistant,” combined with a short description of its role, is more informative. Customers should understand at the beginning that an AI system is generating or controlling the responses they receive.
Must the notice be repeated on every phone call?
The notice should appear during each new interaction so that first-time and occasional callers receive it. A statement on the website or during an earlier call does not reliably inform a person starting a later conversation. A short opening message should therefore identify the AI assistant before it asks for the issue, account information, address, or other details.
May an AI phone assistant sound human?
A natural voice is not generally prohibited. The more human the voice, name, and conversational behavior appear, however, the less a company should rely on the argument that the AI identity is obvious. An explicit opening statement helps prevent callers from mistaking the assistant for an employee or providing information based on a false assumption about the speaker.
Must every AI interface provide access to a human?
Article 50(1) does not expressly require an immediately available representative for every ordinary AI interaction. A human route may nevertheless follow from the business process, other legal duties, or the seriousness of the request. Complaints, contractual decisions, emergencies, payment disputes, and safety-related issues should have a defined escalation process that customers can understand and use.
Where should the notice appear in a website chat?
The notice should be visible in the open chat window and appear no later than the first assistant message. Suitable placement includes the area above the input field or the opening message itself. A statement available only in the privacy policy, legal notice, terms of use, or a separate help page does not reliably reach the customer at the first interaction.
Is a registration notice enough for a customer portal?
A one-time registration notice may not be sufficient when the AI feature is used later or added after the account was created. The function should be labeled as an AI assistant at the point of use. When it opens, the portal should explain its task, information scope, and which decisions or confirmations remain with an employee.
Must a human service representative disclose background AI support?
Not automatically. When an employee uses AI internally for knowledge retrieval, call summaries, or writing suggestions and personally decides what to communicate, the customer is still interacting with a person. The assessment may change when AI-generated replies are delivered directly or passed through without meaningful review, making the human representative largely nominal.
Does the AI notice replace a privacy notice?
No. The Article 50 notice identifies the nature of the conversational partner. Privacy information addresses matters such as the controller, processing purposes, legal basis, retention, recipients, and individual rights. If calls are recorded, transcribed, analyzed, or retained for additional purposes, the applicable requirements must be assessed and addressed separately from the AI identity notice.
May an AI assistant use a human name?
A human name is not automatically prohibited, but it should not conceal the system’s artificial nature. “Anna from the service team” may create the impression of an employee. “Anna, [Company]’s AI assistant” provides a more accurate description. The name, voice, profile image, greeting, and interface should not combine to create a misleading human identity.
What should happen when AI transfers the conversation to an employee?
The change should be apparent to the customer. The assistant can announce that an employee is taking over and explain which information will be transferred. A chat can display a system message, while a phone assistant can provide a short spoken notice. This establishes when a person begins handling the communication and becomes responsible for subsequent statements.
