Gains:
- It designs a secure response flow that classifies citizen questions, produces draft responses, and delegates them to humans when necessary.
- It measures chatbot quality on a topic-by-topic basis with coverage, accuracy, turnover and resolution rate indicators and finds blind spots.
- It determines which citizen questions can be answered automatically and which must be left to human authority.
The point where the public service comes into most contact with the citizen is the service desk: telephone, box office, e-mail, call center, institution's website and social media. A citizen asks "How do I pay my property tax?", "Where do I get a residence certificate?", "What documents are required for this document?" he asks. Many of these questions are recurring; The answers are obvious; but the volume is high and limited to business hours. Chatbot (software assistant that produces automatic answers to written or voice questions) and AI-supported response systems meet this repetitive burden and direct the officer to tasks that require real expertise. In this unit, you'll learn how to use AI in citizen services end-to-end—from question classification to response generation, multilingualism, accessibility, and the most critical aspect, preventing misinformation.
Why is the citizen question difficult?
The citizen question is difficult in three respects. First, accuracy: an incorrect list of documents brings the citizen to the institution in vain; An incorrect period will result in loss of rights. Secondly, equality: two people asking the same question cannot be given different answers; Moreover, everyone should understand it, regardless of education level, language or disability. Thirdly, privacy: citizens often write down their personal information when asking questions ("my TR ID is this, how is my application?") and this data must be protected.
The key to positioning AI within these three constraints is this: AI is an interface that conveys pre-approved accurate information to the citizen in appropriate language; It does not produce legislation on its own. The way to achieve this is to connect the chatbot to the organization's approved knowledge base (FAQs, regulatory summaries, service guides) — this is called RAG (retrieval-augmented generation) where the AI pulls the relevant piece from a trusted document repository and responds solely based on it before producing an answer.
Tip: Always include the source in the chatbot's response: "This information is based on [service guide / circular number], [official page] for current status." An unsourced answer is both unverifiable and reduces the citizen's trust.
Step by step: setting up a citizen response flow
- Determine the scope. On which issues will the chatbot respond (tax, civil registration procedures, license) and on which issue will it say "I am directing you to the relevant unit"? It is preferable to delegate it to humans rather than making up an answer to a question that is out of scope.
- Prepare the approved knowledge base. Each answer must be based on a dated and sourced text that is controlled by the institution. A responsible person should keep track of the updates.
- Write down tone and language rules. Polite, simple, does not use jargon, short sentences. A formal but not cold language.
- Set the personal data rule. The chatbot should provide personalized information, such as transaction status, only after secure authentication; should not request personal data when providing general information.
- Define escalation points. Complaints, objections, questions that have legal consequences or unclear questions should be conveyed to people.
- Monitor and improve. Questions answered incorrectly should be reviewed weekly and the knowledge base updated.
three mini cases
Case 1 — The load decreased, the officer turned to expertise. A metropolitan municipality's call center received ~48,000 calls per month; of these, 61% were about five recurring topics. The RAG-based chatbot took on these five topics; call center staff were divided into complex complaints and appeals; The average waiting time decreased from 7 minutes to 2 minutes.
Case 2 — Misinformation trap. An unsourced chatbot pilot repeated an old amount saying "passport fee is 1,200 TL this year"; whereas the mortar had been updated. For 3 days, citizens came with the wrong amount. Lesson: the chatbot should only rely on dated, up-to-date and sourced knowledge base; should not be left to free production.
Case 3 — Accessibility gain. An organization simplified chatbot responses to a B1 (basic-intermediate) language level and made them suitable for screen readers for the visually impaired. The rate of citizens coming to the box office because they did not understand the service guide decreased by 28%; In complaints, feedback with the theme "I don't understand" dropped to one third.
Four copyable templates
1) Source-dependent response generation:
Your role: a polite and plain-spoken public service consultant. Answer only based on the CERTIFIED INFORMATION text I have provided below. If you are asked something that is not in the text, say "I direct you to [unit] for precise information on this subject" and do not make up the answer. Add the source at the end of the answer. APPROVED INFORMATION: [paste text] QUESTION: [citizen question]
2) Language simplification:
Translate the official text below into plain Turkish that a secondary school graduate can easily understand. Explain jargon, shorten sentences, make a step-by-step list. The meaning and legal content should not change. Mark the unclear area as "needs clarification". TEXT: [paste official text]
3) Multilingual response (with verification notes):
Translate the following approved Turkish answer into [target language].Recognize the official terms correctly; Leave the term you are not sure of in parentheses with its Turkish version. Add a note that the translation should be checked by an expert. ANSWER: [text]
4) Escalation classifier:
Classify the following citizen message: (A) general inquiry -> auto-reply appropriate, (B) personalized transaction status -> authentication required, (C) complaint/appeal/legal outcome -> handover to human. Give just one letter and one sentence justification. If uncertain, safest option (C). MESSAGE: [text]
Weak prompt / Strong prompt
Weak: "Inform citizens about property taxes."
Strong: "Your role is that of a polite public advisor. Answer ONLY based on the approved text of information below; make up what is not in the text, refer to the relevant department if necessary. Write the answer simply, concisely and step by step. Do not ask for personal data. Add the source and a note 'check official page for current amount' at the end. Approved information: [text]. Question: [question]."
Difference: strong prompt confines response to approved source, prevents fabrication, defines tone and confidentiality; The result is of a quality that can be shown directly to citizens.
Automatic response or handover to human?
Status
proper way
Why
Document list, working hours, general process
Automatic response (RAG)
Information is fixed and approved
“How is my application doing?”
Authentication + system query
personal data
Complaint, objection, grievance
handover to human
Appreciation and empathy required
Question with legal consequences
handover to human
Responsibility and justification
Out of scope/uncertain
handover to human
Avoiding the risk of fabrication
Measuring chatbot quality: numbers, not intuition
A citizen chatbot is not managed by saying "it seems to be working fine"; is measured. Track chatbot performance in public service with four indicators. Coverage rate (percentage of the question that the chatbot answered correctly without handing it off to the human); accuracy rate (compliance of the answers given with the legislation and current practice, with manual sample control); turnover rate (questions transferred to the human — if it is too low, the chatbot may be confident about the wrong topics); satisfaction/solution rate (resolving the citizen's problem at the first contact). Tracking these indicators on a weekly dashboard makes it visible where the chatbot is producing hallucinations and which topics need to be delegated to the human.
Mini case — blind spot caught. A provincial directorate's chatbot gave a good overall coverage rate of 78 percent; However, on the subheading “rights of disabled citizens,” the accuracy rate was only 41 percent because the model produced an answer based on an unupdated regulation. Without subject-based measurement, this blind spot would remain hidden within the average. The team temporarily closed this thread to the "always delegate to human" rule and reopened it after the content was updated.
A weekly template for checking response quality by sample:
Task: Below are 20 answers and citizen questions that the chatbot gave last week. Evaluate each answer: 1) Is it in compliance with the legislation? (Yes/No/Doubtful — specify item/source)2) Is there any missing or misleading information?3) Should it have been delegated to a human?Output: Table (question number | suitability | risk note | recommendation).List those whose suitability is questionable to the top.Note: This is a draft audit; The final decision belongs to the inspection officer.
Tip: The chatbot's ability to say "I don't know, I'm transferring you to an official" is not a weakness, but a quality assurance. In public, a confident wrong answer is much more costly than an honest takeover.
Common mistakes
- Releasing the chatbot to free production. A bot that is not connected to an approved knowledge base will provide outdated or fabricated information.
- Not citing sources. Citizens cannot verify, civil servants cannot inspect; The bug spreads silently.
- Processing personal data without authentication. Answering the question "What is the status of my application?" without secure verification is a KVKK risk.
- Not defining tipping points. Leaving complaints and objections to the bot will increase the victimization.
- Not simplifying the language. Accurate information that is not understood is null and void for the citizen; accessibility is a must.
- Not establishing a feedback loop. If incorrectly answered questions are not monitored and corrected, the same mistake will be repeated.
In summary
In citizen services, AI responds to repetitive questions quickly, simply and 24/7, directing the officer to jobs that require real expertise. The secret is in three things: linking answers to the approved knowledge base (RAG), adding sources to each answer, and delegating complaint/objection/personal action questions to humans. A properly designed chatbot reduces the load; An unsourced and uncontrolled bot spreads false information quickly.
Application task
Write down the 5 most frequently asked citizen questions in your unit and their approved answers. Generate chatbot answers for these five questions using the “Resource based answer generation” template. Then, with the "Delegation classifier" template, mark which questions should be delegated to the human. Simplify at least one answer to level B1.
checklist
- [ ] Chatbot only depends on approved, dated, sourced knowledge base.
- [ ] Each answer cites the source.
- [ ] Personal data is not processed without secure authentication.
- [ ] Transfer to human for complaint/objection/legal question defined.
- [ ] The language is simple and accessible (disability and linguistic diversity are taken into account).
- [ ] There is a weekly feedback and update cycle for incorrect answers.