A customer service chatbot is a virtual assistant on your website, app or WhatsApp that answers repeat questions — order status, returns, product details — from your own data. A good one does not guess: what it cannot resolve goes to a live agent with the full history. Success depends on getting the scope and hand-over rules right.
Where the real workload in customer service sits
A large share of customer messages boils down to the same handful of questions: "Where is my order?", "How do I return this?", "Will this fit my device?", "Where can I download my invoice?". Agents spend much of the day retyping the same answers. Messages that arrive after hours wait until morning, and during promotions customers sit in a queue. Hotels see the same pattern in guest messages and municipalities in citizen questions; see AI in hotels and hospitality and AI in local government.
Classic decision-tree bots are not much help. They march customers through menus and reply "Sorry, I didn't understand that" to anything unexpected. Dropping a general-purpose language model straight onto your website creates the opposite problem: it chats fluently, but it does not know your returns policy — and it will not tell you that it doesn't know.
The right design sits between these extremes: a bot that uses a language model's ability to understand and write, but takes its facts only from your catalogue, FAQs and order system. When a customer writes "my parcel hasn't arrived, I've been waiting three days", such a bot asks for the order number, reads the shipping status from the system, states the delay plainly and, if the customer is still unhappy, passes the conversation to an agent. The customer is not lost in a menu, and the agent does not have to ask for the same details again.
The cost of a badly built bot — or no bot at all
Customers expect fast answers on digital channels. In Türkiye, TurkStat's 2026 Survey on ICT Usage in Households shows internet use among 16–74-year-olds at 92.3%, with 60.0% buying goods or services online. A customer who ordered online expects to get answers in the same channel, without waiting.
Businesses are putting AI to work here first. According to TurkStat's Artificial Intelligence Statistics 2025, 46.5% of Turkish enterprises using AI used it for marketing or sales — the most common purpose in the survey.
A bot that is not grounded in your own content, however, can mislead customers. In Vectara's document-summarisation benchmark, as reported in the Stanford HAI AI Index Report 2026, even the top 15 models evaluated introduced unsupported information into document summaries at rates between 1.8% and 5.4%, and documented AI incidents in the AI Incident Database rose from 233 in 2024 to 362 in 2025. In a customer-facing bot, a single wrong returns rule or invented promotion becomes a complaint an agent spends hours putting right.
Which questions should a chatbot handle?
Not every conversation belongs with a bot. The table below shows the simple split we use when drawing the scope:
| Conversation type | Example | Bot or agent? | Why |
|---|---|---|---|
| Information request | "How long does delivery take?" | Bot | The answer is documented and asked often |
| Status check | "Where is my order?" | Bot, with a system connection | Read live from the order system |
| Product choice | "Which model suits me?" | Bot, then sales if undecided | Catalogue comparison is possible |
| Troubleshooting | "The device won't turn on" | Bot step by step, then a ticket | Follows the manual's flow |
| Complaint or returns dispute | "It arrived damaged" | Agent | Needs judgement and empathy |
| Personal data or payment changes | "Update my card details" | Agent or secure form | Needs identity checks and security |
The rule is simple: questions whose answer is written down in a document or a system go to the bot; conversations that need judgement, flexibility or identity checks go to a person. We look at the same split for voice channels in call centre speech analytics.
Seven steps to launching a customer service chatbot
- Classify the last three months of conversations. Pull the 20 most common question types from email, WhatsApp and phone logs. The bot starts with that list, not with everything. Grouping Turkish messages full of typos and abbreviations reliably is a job for Turkish natural language processing.
- Prepare the knowledge source. FAQs, delivery and returns terms, product catalogue, user manuals. Contradictory or outdated documents break bot answers directly, so agree on a single, current source.
- Connect the systems. For order status, invoices and service tickets the bot needs to talk to your ERP, e-commerce and support tools; API integration explained shows how that works.
- Ground every answer. The bot should answer only from the texts in its knowledge source, and never guess when the answer is not there. The technique is called RAG; its architecture is explained in retrieval augmented generation for the enterprise, and other safeguards in reducing AI hallucinations.
- Write the hand-over rules. When does a conversation pass to an agent? When the customer asks for a person twice, when frustration is detected, when the bot fails twice, or when the topic involves a complaint, payment or personal data. The full conversation history must travel with it, so the customer never has to start again. Customers should also know from the start that they are talking to AI; if you serve customers in the EU, that transparency rule is an obligation under the EU AI Act.
- Build in data protection from day one. Chat logs contain personal data. Link to your privacy notice in the chat window, write the retention period into your data retention and disposal policy and do not ask for personal data the bot does not need. In the UK and EU that means the GDPR; in Türkiye the equivalent is the KVKK (Law No. 6698), which likewise requires notice and proportionate processing. For automated decisions and transfers abroad under Türkiye's rules, see AI and KVKK.
- Measure and feed it. Track weekly the share of conversations resolved without an agent, the hand-over rate and the unanswered questions. Every unanswered question points to a gap in the knowledge source. To work out the return, use the method in measuring AI project ROI.
What decides the quality of the hand-over is whether the bot and the agents see the same customer record. Our CRM and sales pipeline guide explains why customer history belongs in one place. If you run a portal for dealers or trade customers, the bot can sit there too; see our B2B dealer portal guide.
Common mistakes
- Passing the bot off as a person. Customers should know they are talking to a bot. One that introduces itself as an agent damages trust twice over when it gets something wrong; saying up front that it is a virtual assistant, and showing the way to a person, is enough.
- Hiding the route to an agent. Making hand-over difficult artificially inflates the bot's "resolution rate" while lowering customer satisfaction. A person should always be one step away.
- Preparing the knowledge source once and leaving it. As promotions, prices and delivery terms change, the bot goes stale. Keeping the source current must be someone's explicit job, and the list of unanswered questions should be reviewed every week.
- Opening every channel at once. It is less risky to pilot on one channel, usually the website, settle the hand-over rules and knowledge source, and then extend to WhatsApp, the mobile app or self-service kiosks in shops and reception areas.
- Measuring success by conversation count alone. A bot can hold plenty of conversations, but if the customer's problem is not solved, that is a postponed complaint rather than a success.
How we approach this at Digital Bridge
In our NLP chatbot projects we build a bot that knows your business rather than a generic one:
- Answers from your own knowledge. The bot uses the language skills of large language models — GPT, Claude, Gemini or open-source models running in-house — but bases its answers on your catalogue, FAQs, user manuals and past support tickets.
- Channel-independent. It answers around the clock through a website chat window, your mobile app and WhatsApp Business, in Turkish and English.
- Connected to your systems. Through API integration it reads order status and support tickets from your e-commerce, ERP and support tools, and can connect to your CRM as well.
- It leaves what it doesn't know to people. Conversations it cannot resolve go to a live agent with their full history; for faults it walks the customer through troubleshooting and turns unresolved cases into a support ticket with the details already collected.
- Measurable. Conversation reports and a list of unanswered questions show where the bot falls short. We address personal data through our data protection compliance service.
If the assistant will only answer staff questions about internal procedures, an enterprise LLM assistant is the better starting point. We do not sell packaged products; after a needs analysis we provide a written proposal setting out scope, phases and cost.
Next step
Start by taking 200 random customer messages from the last three months and labelling them using the table above. You will see how many belong with a bot. Then contact us and we will review the list with you and scope a small pilot. To see where else AI could help your business, read AI in business: where to start and browse our complete Artificial Intelligence guide.