ChatGPT in Customer Service 2026: What Works, Where It Gets Risky and When You Need Your Own AI Agent

In short: ChatGPT is a good tool for staff in customer service: draft replies, summaries and translations, always with a person who checks before anything is sent. It gets risky as soon as customers talk directly to a general language model, or customer data is copied into private accounts. Your own AI agent answers from your knowledge base, acts in your systems, hands over to your team and also works on the phone.

We revised this post in October 2026. The previous version claimed that ChatGPT was the most-searched term in Germany in 2026, and told the story of a US car dealership whose chatbot allegedly offered a car for one dollar. We found no verifiable source for either. We also removed unverifiable details of OpenAI's plans. On EchoCall, it listed model profiles, extra messaging channels, a latency figure and a language count. The current prices are further down. WhatsApp is currently not part of what you set up yourself in the EchoHub. Today the EchoCall chat agent runs on your website, with handover to your team.

What ChatGPT is good for in customer service

AI has arrived in German businesses. According to the Federal Statistical Office, 26 per cent of companies with ten or more employees used AI technologies in 2025. Of these, 35 per cent used AI to generate natural language [1]. In customer service, this usually means someone on the team drafts replies with ChatGPT.

As a tool for staff, not as a point of contact for customers, a general language model is a real help in day-to-day service:

• Draft replies: three bullet points become a friendly, complete email that a person checks before sending.
• Tone: an angry complaint gets a calm, factual reply, even if it is the third one that day.
• Summaries: a long ticket history shrinks to five lines before a colleague takes over.
• Translations: understanding an enquiry in Polish and translating the reply back.
• Text snippets: writing macros and FAQ answers, making help articles easier to understand.

In all these cases, a person who knows the facts stays in the loop and approves the reply. The model supplies the language, the team supplies the facts. As long as this division of labour holds, the risk is manageable.

What problems can AI agents solve in customer service?

An AI agent goes one step further. It talks directly to your customers and completes tasks itself. OpenAI describes agents as systems that independently accomplish tasks on a user's behalf, with access to tools and within defined guardrails [6]. In customer service, the best candidates are requests that come up often and end with a clear action:

• recurring questions about opening hours, delivery times, returns or prices, including evenings and weekends,
• booking, moving and cancelling appointments,
• information the agent looks up in a system, such as the order status in your shop,
• recording a request, opening a ticket with all the details and routing it to the right place,
• absorbing call peaks when all lines are busy,
• enquiries in several languages.

Less suitable are complaints with potential for conflict, goodwill decisions and anything that needs a legal assessment. Here the agent should hand over to a person. How an agent differs from a chatbot is explained in What Is an AI Agent? Definition, How It Works and Examples.

Where it gets risky: five problems when customers talk directly to ChatGPT

The difficulties start when a tool for staff becomes a channel for customers: a chat window on your website that passes enquiries straight to a general model, or a team that copies customer enquiries, data and all, into a chat window.

1. Hallucinations: the model invents what it does not know

In its report on ChatGPT, the taskforce of the European data protection authorities writes that, due to the probabilistic nature of the system, the training approach leads to a model that may also produce biased or made-up outputs. At the same time, end users are likely to take these outputs as factually accurate [2].

In customer service, that is exactly the expensive mistake. Air Canada's chatbot told a customer he could book at full price and claim back the difference to a reduced fare within 90 days. That was not true. In February 2024, the Civil Resolution Tribunal of British Columbia ordered the airline to pay the fare difference of 812 Canadian dollars, plus costs and interest. The tribunal rejected the argument that the chatbot was a separate legal entity responsible for its own actions [3]. So plan as if you will have to stand behind every answer your bot gives.

2. No access to your data

ChatGPT does not know your opening hours, your current price list, and certainly not Mrs Miller's order status. You can paste it into a prompt, but then you maintain a second, handwritten knowledge base that dates with the next price change. A customer service agent needs a knowledge base fed from your website and your documents and refreshed regularly. How to do that without programming is shown in How to Build an AI Agent Without Code in 2026.

3. Data protection: customer data in the chat window

This is where it gets concrete for many teams. In 2024, the conference of German data protection authorities published guidance on AI [4]. For day-to-day service, these points matter most:

• Staff should only use AI through business accounts, not private ones.
• Training on inputs and the input history should be switched off by default.
• Controllers must check whether the provider uses inputs and outputs for training. Applications without training are preferable.
• A cloud provider is often a processor, in which case an agreement under Article 28(3) GDPR is required.
• Removing names and addresses is regularly not enough to anonymise data.
• Stricter rules apply to health data and other special categories under Article 9 GDPR.

The EDPB report also notes that OpenAI refers to prompts as content, states that it uses them for training and offers an opt-out [2]. Check how this applies to your plan in your contract and settings. Either way, a customer enquiry with name, address, order number and complaint text does not belong in a private account. What an AI provider for customer service should deliver is covered in the GDPR guide for voice agents.

4. No actions

Used as a normal chat window, the model can explain how to book an appointment but cannot enter one in your calendar. It can describe what a return involves but cannot create one. Without a connection to calendar, CRM, ticketing system or shop, the conversation ends with "Please contact our team". The difference between a chatbot that answers and an agent that acts is covered in AI Agents vs. Chatbots: What's the Difference?.

5. No phone

Many urgent requests still come in by phone. A chat window does not help the caller who wants to move an appointment at half past five. Taking the phone seriously requires a voice agent that responds quickly enough for the conversation to feel natural, and that can transfer to a person.

Can AI agents be used in customer service in compliance with the GDPR?

Yes, if the conditions are right. The key points at a glance:

1. Contract: anyone passing personal data to an AI provider usually needs a data processing agreement under Article 28 GDPR. If data is processed outside the EU, the rules for transfers to third countries under Chapter V GDPR also apply [4].
2. Training: find out whether conversation content is used to train models. The German data protection conference recommends applications without training [4].
3. No automated final decisions: under Article 22 GDPR, people have the right in principle not to be subject to a decision based solely on automated processing with legal or similarly significant effects [4]. An agent may book appointments. A person decides on a cancellation or a goodwill gesture.
4. AI notice: Article 50 of the AI Act has applied since 2 August 2026. People who interact directly with an AI system must be informed of this [5]. In the EchoCall chat widget, the notice appears automatically. For phone agents, you write it into the first message yourself, and the editor reminds you with an info box. More in EU AI Act 2026: What the AI Disclosure Duty Means for Voice and Chat Systems.
5. Responsibility: you remain responsible for what your agent tells people. That responsibility cannot be passed on to the model provider.

ChatGPT as a chat window vs. your own customer service agent

A customer service agent also uses a language model. The difference lies in everything built around it: knowledge base, tools, limits, handover and contracts.

What your own AI agent does differently

1. It answers from your knowledge base. At EchoCall you upload PDF, TXT, Markdown or CSV files, add single URLs with a re-read interval, have the whole website crawled with a refresh interval, or write free text. Each agent has its own knowledge base.
2. It acts in your systems. Calendars such as Google Calendar, Outlook or Cal.com, CRMs such as HubSpot or Salesforce, ticketing systems such as Zendesk or Freshdesk, shops such as Shopify or Shopware 6, plus Zapier, Make and custom webhooks.
3. It knows its limits. In chat, your team takes over in the Live Inbox whatever the agent should not handle. On the phone, the voice agent transfers to a phone number.
4. It talks. The voice agent answers calls, with system tools such as voicemail detection, keypad tones and automatic language detection, which can be switched off. EchoCall supports more than 30 languages. Since October 2026, visitors can also speak via microphone in the chat widget.
5. It runs on a defined infrastructure. EchoCall runs the platform in IONOS data centres in Frankfurt and Karlsruhe. The EchoHub has a zero-PII mode.

Which chat use cases work well is covered in AI Chat Agent Use Cases 2026.

When is ChatGPT enough, and when do you need an agent?

ChatGPT is enough when a member of staff checks every reply before it goes out, when no personal data is entered or a business account with a contract and training switched off is used, and when the job is wording, summarising or translating.

You need your own agent as soon as one of these statements applies:

• Customers should talk or write directly to the AI without a person approving every reply.
• The AI should get something done: appointments, tickets, order information.
• The phone is a relevant channel.
• You process health, financial or other sensitive data.
• You want to know what the AI said and close gaps systematically.

From ChatGPT experiment to your own agent in one week

Teams that already use ChatGPT have taken the most important step: they know which requests keep coming back. OpenAI advises starting small, validating with real users and growing capabilities over time [6]. A realistic plan:

• Day 1, try it: in the live demo you enter your website address. In about 20 seconds an agent is built from your public pages, and you can call it or write to it in the browser without signing up. You will see straight away which answers are missing.
• Day 2, tidy up the knowledge base: the prompts and snippets your team collected for ChatGPT become your FAQ. Add your website, price list and terms.
• Day 3, rules and limits: what the agent does, what it does not do and when it hands over. Check the AI notice.
• Day 4, connect one tool: usually the calendar or the ticketing system.
• Day 5, go live: install the website widget with a script snippet or the WordPress plugin, with generous handover to the team at first.
• Week 2: read the conversations, close gaps and decide whether to add the phone.

What your own AI agent costs

At EchoCall you book packages, net per month [7]:

For the voice agent, that is €0.30 per minute, and there is no setup fee. Phone numbers and telephony costs are not included in the package prices. In the smallest chat package, one conversation works out at €0.40. To try it out there is a trial: 30 voice minutes and 30 chat conversations for 14 days, and the credit card is verified with €1. You can work through your own scenario with the ROI calculator.

Frequently asked questions

Can I enter customer enquiries into ChatGPT?

With personal data, only under clear conditions: through a business account, with a data processing agreement, with training switched off and only if the transfer is legally covered. Real anonymisation is harder than it looks. According to the German data protection conference, removing names and addresses is regularly not enough.

Is ChatGPT GDPR-compliant?

That depends on the plan, the contract and how you use it. For your account, check whether a data processing agreement is in place, whether inputs are used for training and where the data is processed. Private accounts are not a solution for customer data.

What does the AI Act require for AI in customer service?

Mainly the transparency duty in Article 50, in force since 2 August 2026: people must be told that they are talking to an AI. The EchoCall chat widget does this automatically. On the phone, you write the notice into the agent's first message.

Can I embed ChatGPT as a chatbot on my website?

Technically yes, via the provider's programming interface. But then you have to build the knowledge base, limits, actions, handover to people, AI notice, logging and contracts yourself. That is exactly what a platform for customer service agents takes off your hands.

What happens if the agent cannot answer a question?

You define that in the instructions. It makes sense for the agent to state the gap openly and hand over: in chat to your team in the Live Inbox, on the phone by transferring to a number. Wrong answers cannot be ruled out completely. So read the conversations regularly and close gaps in the knowledge base.

Sources

1. Federal Statistical Office (Destatis), companies using artificial intelligence technologies in 2025 (as of 24 November 2025), in German: destatis.de
2. European Data Protection Board, Report of the work undertaken by the ChatGPT Taskforce (May 2024): edpb.europa.eu
3. Moffatt v. Air Canada, 2024 BCCRT 149, case summary: wikipedia.org
4. Datenschutzkonferenz (German data protection conference), Orientierungshilfe Künstliche Intelligenz und Datenschutz, version 1.0 (6 May 2024), in German: datenschutzkonferenz-online.de
5. Regulation (EU) 2024/1689 (AI Act), Article 50: eur-lex.europa.eu
6. OpenAI, A practical guide to building agents: cdn.openai.com
7. EchoCall pricing: echocall.de/en/pricing

All links were checked on 6 October 2026. This post is not legal advice.

Read next

• What Is an AI Agent? Definition, How It Works and Examples
• AI Chat Agents vs. Traditional Chatbots: Differences, Costs and When a Rule-Based Bot Is Enough
• How to Build an AI Agent Without Code in 2026
• AI and Data Security 2026: GDPR Guide for Voice Agents