In short: AI agents in customer service pay off where many similar requests come in, the answers are clear and a person takes over as soon as things get difficult. Most projects do not fail because of the technology. They fail because nobody chose the right requests, nobody measured and nobody planned the handover to the team. According to Gartner, only 24 per cent of customer service leaders could show a positive financial return from their AI use cases. [1] This guide takes you through eight steps from contact analysis to rollout, with a cost example, KPIs and the legal duties.
We rewrote this post in October 2026. The first version from February 2026 promised recognition of regional accents, processing on edge servers, a commitment that data never leaves the DACH region, and proactive service models for market-leading companies. We cannot back up these claims, or they do not describe what EchoCall offers today. They have been removed. New are an eight-step plan, a cost example with our current prices and figures from Gartner, Forrester and Bitkom, each with a source.
This guide is for small and medium-sized businesses that want an AI to answer some of their calls or website chats. What an AI phone agent costs and how it works technically is covered in AI Phone Agent 2026: What It Costs, How It Works and How to Roll It Out. What changed in the market in 2025 and 2026 is in our timeline The Rise of AI Voice Agents 2025 to 2026: What Really Changed in Technology, Prices, Funding, Law and Analyst Forecasts. This post deals with the question that comes first: where do you use an AI agent, and how do you introduce it so that it lasts?
What AI agents can and cannot solve in customer service
An AI agent differs from a classic chatbot in two ways. It understands free speech instead of asking users to pick menu options. And it can act: enter an appointment, open a ticket, look up an order status or hand a conversation to a person. We explain the difference in AI Agents vs. Chatbots: What's the Difference? (2026).
Typical problems AI agents solve in customer service:
- Availability: calls and chats outside opening hours, over lunch or when every line is busy.
- Recurring questions: opening hours, prices, delivery times, required documents, who is responsible for what.
- Simple tasks: booking or moving an appointment, taking a callback request, recording a request in a structured way.
- Pre-qualification: asking for the request, its urgency and contact details before a person takes over.
- Languages: answering requests in languages nobody on the team speaks.
AI agents reach their limits with goodwill decisions, angry customers with a long history, questions that carry liability such as medical, legal or tax advice, and tasks for which no reliable data exists. These cases belong in the handover from day one.
Why you need a strategy: what the 2026 figures show
AI in customer service is no longer a side topic in 2026. According to Bitkom, 57 per cent of companies in Germany with 20 or more employees use AI, up from 36 per cent a year earlier. AI is used most often in customer contact, for example to handle enquiries (72 per cent). [2] AI agents that act on their own, however, are only starting out. Of the companies that use AI or are planning or discussing it, 11 per cent already use such agents, 29 per cent plan to and 31 per cent are discussing it. [2]
At the same time, analyst figures show how often projects fall short:
- 91 per cent of customer service leaders feel pressure from executive leadership to implement AI. Gartner surveyed 321 service leaders in October 2025. [3]
- Service functions invested a median of 12 per cent of their 2025 budget in AI. Yet only 24 per cent could show a positive financial return from their AI use cases. [1]
- More than 40 per cent of agentic AI projects will be cancelled by the end of 2027, according to Gartner, because of escalating costs and unclear business value. [4]
- Only 20 per cent of service leaders have actually reduced staff because of AI. Gartner expects that by 2027 half of the companies that attributed headcount reduction to AI will rehire staff for similar work, under different job titles. [5]
- Forrester predicts that in 2026 a third of companies will harm customer experiences with frustrating AI self-service, because cost pressure leads to premature deployments. [6]
The lesson is not to avoid AI. The lesson is to start small, measured and with a handover. Bitkom president Ralf Wintergerst puts it roughly like this: an agent that acts on its own needs reliable data, connected systems and rules on what it may do. [2] That is what the following eight steps are about. The developments behind these figures are covered in Customer Service Trends 2026: Eight Shifts in Conversational AI, Voice AI and AI Agents, with Sources. Where German and European businesses stand on AI overall is described in Digital Transformation 2026: Where Businesses Stand on AI and What Voice AI Changes on the Phone.
Step 1: Take a contact inventory
Before you talk to providers, record for four weeks what actually comes in. A few details per request are enough:
| Field | Example |
|---|---|
| Channel | phone, website chat, email |
| Time | Tuesday, 12:40 |
| Request in one sentence | "move my appointment" |
| Duration | 3 minutes |
| Outcome | solved, callback needed, forwarded |
| System access needed | calendar, inventory system, CRM or none |
Count the calls nobody answered as well. Most phone systems show missed calls in the call log. What these gaps cost is calculated in Missed Calls Cost Revenue: What Being Unreachable Really Costs Your Business in 2026 and How an AI Answering Service Changes That.
After four weeks, group the requests and sort them by frequency. Often a few requests make up a large share of all contacts. Check this against your own figures. Do not record names, the request is enough.
Step 2: Choose the right requests
Not every frequent request suits the start. Rate your main requests against five criteria: volume, how repeatable the answer is, available data, the risk of a wrong answer and the action required.
| Request | Suitability for the start | What the agent needs |
|---|---|---|
| Opening hours, directions, prices, documents | very good | an up-to-date knowledge base |
| Booking, moving or cancelling an appointment | good | calendar integration, clear rules |
| Order or delivery status | good | connection to the shop or inventory system |
| Callback request, recording a request | good | ticket or email action |
| Complaint, goodwill | record only, then hand over | handover to the team |
| Cancellation, contract change | with caution | identification, handover |
| Medical, legal or tax advice | not suitable | handover to qualified staff |
Start with two or three requests from the top rows. Gartner advises service leaders to begin with targeted, thoroughly tested deployments and to prioritise reliability over reach. [7] The reason is in the same customer survey: only 27 per cent of customers would try a chatbot again after a negative experience. [7]
Step 3: Choose your channels: phone, website chat or both
Your inventory answers the channel question. Where do the chosen requests arrive?
- Phone: callers usually want an answer straight away. A voice agent answers calls, responds to questions, carries out actions and transfers the call when needed.
- Website chat: good for questions that come up while people browse your website, for example about products, prices or delivery times.
- Both: if the requests overlap, both agents should work from the same content so that phone and chat give the same answers.
In a Gartner survey of 199 service leaders, live chat and phone are their most valuable channels today. In two years they expect GenAI chatbots to lead, followed by live chat and GenAI voicebots. [8] Even so, plan for the channel your customers use today.
Step 4: Prepare knowledge, data and integrations
An AI agent is only as good as the content it draws on. According to Gartner, 58 per cent of service leaders aim to upskill agents into knowledge management specialists. [3] There is a reason for that. Outdated prices or contradictory FAQs lead to wrong answers, and wrong answers cost trust.
Three questions for this step:
- Which sources are authoritative? Price list, FAQs, service descriptions, process rules. Remove contradictions before the agent reads them.
- Which actions does the agent need? Book an appointment, look up data, open a ticket, notify the team. Each action needs an interface and a rule on when it is allowed.
- Which data may the agent see? Give it only what it needs for the request. The principle of data minimisation in Article 5(1)(c) GDPR requires this too. [9]
In EchoCall you upload knowledge as PDF, TXT, Markdown or CSV (voice agents also read JSON), add individual URLs, have your website crawled or write free text. Actions run through integrations such as Google Calendar, Outlook, Cal.com, HubSpot, Salesforce, Zendesk, Freshdesk, Shopify or Shopware 6, through Zapier, Make and custom webhooks, or through the REST API. Integrations are actions during a conversation, not additional channels.
Step 5: Governance, GDPR and the AI disclosure under the AI Act
Governance sounds like something for corporations. A mid-sized business needs four decisions:
- One responsible person for content, greeting, rules and approvals.
- A change process: who may change the knowledge base and rules, and who checks first?
- Fixed review dates: weekly during the pilot, then monthly spot checks of conversation logs.
- A handover rule: when and where does the agent hand over to a person?
The handover is not an emergency plan, it is part of the service. According to Gartner, 87 per cent of customers say it is essential to be able to reach a human agent when a company uses generative AI in customer service. [10] In a Bitkom survey, 62 per cent of online shoppers with a problem would most like to contact a person who is quickly reachable, while 36 per cent wanted a chatbot. [11]
Data protection. Using AI agents in customer service in line with the GDPR is possible. You need a legal basis, a data processing agreement, clarified transfers to third countries, retention periods and a privacy notice. For customer support using artificial intelligence, the data protection impact assessment is also on the mandatory list of the German supervisory authorities. [12] All points, with the legal text, are in our guide AI and Data Security 2026: The GDPR Guide for Voice Agents and AI Chat in Customer Service.
AI disclosure. The AI Act has applied generally since 2 August 2026. Under Article 50(1), people must be informed that they are interacting with an AI system unless this is obvious. Under paragraph 5 the information must be given at the latest at the time of the first interaction. [13] The duty falls first on the provider of the system. How platform and business implement it together, and what good example sentences sound like, is explained in EU AI Act 2026: What the AI Disclosure Duty Means for Voice and Chat Systems. In EchoCall the chat widget shows an AI notice automatically. For phone agents you write the notice into the first message yourself.
Step 6: Set KPIs before the first call
Without a baseline there is no proof. Record where you stand before the start and measure the same values afterwards:
| KPI | What it shows | Where the data comes from |
|---|---|---|
| Availability | share of answered calls and chats | phone system, chat log |
| Resolution rate without handover | requests the agent completes on its own | conversation logs |
| Handover rate and reason | where the agent reaches its limits | conversation logs |
| Repeat contacts | whether the answers hold up | phone system, CRM |
| Time to resolution | speed from the customer's point of view | ticket system |
| Satisfaction | quality from the customer's point of view | short follow-up survey |
| Cost per resolved request | cost-effectiveness | invoices, staff costs |
A warning about the resolution rate: it is only good if the requests are really solved. An agent that holds on to customers instead of passing them on looks good in the statistics and still annoys people. Gartner advises building chatbots as connectors to human support, not as containment traps. [7] That is why repeat contacts and satisfaction always belong in the same report.
Step 7: Run a pilot with a clear exit rule
A pilot lasts four to eight weeks, runs in one channel and covers the two or three requests from step 2. Proceed like this:
- Internal test: your team calls or writes, with real questions from the inventory and with difficult questions the agent should hand over.
- Limited start: for example only outside opening hours, only for one location or only on one page of the website.
- Weekly review: read conversation logs, correct the knowledge base, sharpen the rules.
- Decision against fixed criteria: expand, improve or stop.
Set the exit rule in advance. An example: if repeat contacts rise after four weeks, or satisfaction falls below the baseline, you improve or stop.
You can get a first impression without a project. Our live demo builds an agent from the public pages of your website in about 20 seconds, which you can call in the browser or write to, without signing up. For the pilot there is a trial: 30 voice minutes and 30 chat conversations for 14 days, and the credit card is verified with €1.
Step 8: Rollout and change in the team
After a successful pilot, expand step by step: more requests, more hours, then the second channel. The team is decisive. According to Gartner, nearly 80 per cent of organisations plan to move at least some agents into new roles, and 84 per cent plan to add new skills to the agent role. [3]
In practice three tasks emerge:
- Taking over: staff take over conversations the agent hands on. In EchoCall your team takes over chats in the Live Inbox, and the voice agent transfers calls to a phone number.
- Maintaining: someone keeps the knowledge base and rules up to date.
- Reviewing: someone reads logs and reports where customers get stuck.
Say early on what the agent is for. If the team sees it as relief, it reports errors. If the team sees it as a cost-cutting programme, it stays a foreign body. Gartner's rehiring forecast above shows that staff reduction as the main goal rarely works. [5]
Cost example: how AI lowers customer service costs and where it does not
The following example rests on assumptions. Insert your own figures, most easily in the ROI calculator.
Assumptions: a business receives 600 calls a month with an average length of 2.5 minutes. The voice agent answers all calls, handles routine questions and forwards the rest.
| Item | Calculation | Result |
|---|---|---|
| Call minutes | 600 × 2.5 minutes | 1,500 minutes |
| EchoCall voice plan | 1,500 minutes for €450 | €450 a month |
| Cost per call | €450 divided by 600 | €0.75 |
| Setup fee | €0 |
Phone number and telephony come on top because they are not included in the plan price. For website chat, 300 sessions cost €120 and 1,000 sessions cost €400, so €0.40 per session in both cases.
Compared with staff: if an hour of service work costs you €40 including all overheads (an assumption, insert your own figure), a call minute costs around €0.67, without preparation, follow-up or idle time between calls. The agent only saves money where it really completes requests. If a call is handed over, you pay both.
Compared with outcome pricing: some chat platforms charge per resolved case. Intercom charges $0.99 per outcome for its AI agent. [14] Zendesk charges €1.50 per automated resolution with a committed volume, or €2.00 pay as you go. [15] Minutes, sessions, outcomes and telephony are billed very differently. Always compare offers in the same unit: cost per resolved request.
A frequently cited example of large effects is Klarna. In February 2024 the company reported that its AI assistant had held 2.3 million conversations in its first month, handled two thirds of its customer service chats and done the work of 700 full-time agents. Time to resolution had dropped from 11 minutes to under 2. [16] These are the company's own figures, not independently verified. They are not a benchmark for a mid-sized business.
Where EchoCall fits into this strategy
EchoCall is a platform for AI agents on the phone and in website chat. In concrete terms:
- Voice agent for calls, with transfer to a phone number, automatic language detection and more than 30 languages.
- Chat agent for your website, embedded with a script snippet or the WordPress plugin, with a domain whitelist. Your team takes over conversations in the Live Inbox. Since October 2026 visitors can also talk in the widget. A notice card appears before the microphone starts, and the minutes come from the voice allowance.
- Interfaces: REST API, WHMCS plugin and a white-label portal for resellers.
- Hosting in IONOS data centres in Frankfurt and Karlsruhe, plus a Zero-PII mode in the EchoHub.
- Prices from €75 for 250 voice minutes and from €40 for 100 chat sessions, with no setup fee. Phone numbers and telephony come on top. All plans are on the pricing page.
If you would like to discuss your contact inventory or a pilot with us, get in touch.
Frequently asked questions about AI agents in customer service
What problems can AI agents solve in customer service?
They answer calls and chats when nobody has time, respond to recurring questions, book appointments, record requests in a structured way and hand over to people. Goodwill decisions and advice that carries liability remain a job for people.
Can AI agents be used in customer service in line with the GDPR?
Yes, if the legal basis, data processing agreement, third-country transfers, data protection impact assessment, retention periods and privacy notice are settled. Your business remains the controller. The details, with the legal text, are in our GDPR guide for voice agents and AI chat.
What does the AI Act require for AI in customer service?
For chat and voice agents, Article 50 of the AI Act matters most. People must learn at the latest at the first interaction that they are talking to an AI, unless this is obvious. The duty has applied since 2 August 2026. [13]
How long does it take to introduce an AI agent?
The contact inventory takes about four weeks, the pilot four to eight weeks. Plan two to three months until the decision on the rollout.
Do AI agents replace customer service staff?
Rarely, so far. According to Gartner, only 20 per cent of service leaders have actually reduced staff because of AI. [5] It is more realistic that AI takes over routine questions while staff handle handovers, difficult cases and the upkeep of knowledge.
How does AI reduce customer service costs?
Mainly by fully resolving routine requests and catching missed calls. In EchoCall a voice minute costs €0.30. With 600 calls of 2.5 minutes each that is €450 a month, plus telephony. Whether that is cheaper than today depends on your resolution rate. Work it out in the ROI calculator.
Sources
Gartner blocks automated access to its press releases. We therefore checked the Gartner figures against verbatim reprints.
- Gartner, survey of 3,566 B2B and B2C customers (February and March 2026) and figures on AI budgets in service, reprint at IT-Online dated 9 July 2026: it-online.co.za
- Bitkom, press release of 14 September 2026, survey of 603 companies with 20 or more employees (in German): bitkom.org
- Gartner, survey of 321 customer service and support leaders (October 2025), reprint at IT-Online dated 19 February 2026: it-online.co.za
- Reuters, 25 June 2025, Gartner forecast on agentic AI projects, via Yahoo Finance: ca.finance.yahoo.com
- Gartner, rehiring forecast and survey of 321 service leaders, reprint at IT-Online dated 4 February 2026: it-online.co.za
- Forrester, press release of 28 October 2025, Predictions 2026 for B2C marketing, CX and digital: forrester.com
- Gartner, survey of 3,566 customers (February and March 2026) on chatbots after negative experiences, reprint at IT-Online dated 4 September 2026: it-online.co.za
- Gartner, survey of 199 service leaders (April and May 2026), reprint at IT-Online dated 28 August 2026: it-online.co.za
- Regulation (EU) 2016/679 (General Data Protection Regulation), Article 5: eur-lex.europa.eu
- Gartner, customer survey (February and March 2026) on generative AI in service, reprint at IT-Online dated 5 August 2026: it-online.co.za
- Bitkom, press release of 22 May 2025, survey of 1,006 internet users aged 16 and over, including 978 online shoppers (in German): bitkom.org
- Datenschutzkonferenz (German data protection conference), list of processing operations requiring a DPIA, version 1.1, item 11 (in German): datenschutzkonferenz-online.de
- Regulation (EU) 2024/1689 (AI Act), Articles 50 and 113: eur-lex.europa.eu
- Intercom, pricing page, Fin at $0.99 per outcome: intercom.com
- Zendesk, pricing page, automated resolutions (in German): zendesk.de
- Klarna, press release of 27 February 2024: klarna.com
All links were checked on 6 October 2026. This post is not legal advice.
Read next
- Customer Service Trends 2026: Eight Shifts in Conversational AI, Voice AI and AI Agents, with Sources
- Digital Transformation 2026: Where Businesses Stand on AI and What Voice AI Changes on the Phone
- AI and Data Security 2026: The GDPR Guide for Voice Agents and AI Chat in Customer Service
- AI Phone Agent 2026: What It Costs, How It Works and How to Roll It Out
- ChatGPT in Customer Service 2026: What Works, Where It Gets Risky and When You Need Your Own AI Agent
