How to handle citizen complaints with AI

What AI is good for here

Offices that deal with the public receive complaints by letter, email, WhatsApp, web portals and in person. Many are similar. An AI assistant can help with three tasks:

• Sorting. Suggesting a category for each complaint, such as delay, staff conduct, billing or service quality.
• Summarising. Turning a long or emotional letter into a short, neutral statement of the problem and what the person is asking for.
• Drafting replies. Producing a polite, clear first draft that an officer then checks, corrects and signs.

AI does not decide anything. Whether a complaint is upheld, what action is taken and what the citizen is told are decisions for an accountable officer. See human oversight.

Keep personal data safe

Complaints are full of personal data: names, CNIC numbers, addresses, phone numbers, account numbers and sometimes health or family details. Follow these rules:

1. Use only approved tools. Do not paste complaints into a public chatbot unless your department has approved it for this data. Follow your department's information security instructions.
2. Remove identifying details first. Replace them with placeholders such as "[Complainant]", "[CNIC]" or "[Address]". Share only what the task needs. See data minimisation.
3. Be careful with small details. A village name, a job and a date together can identify someone. See re-identification.
4. Never upload complaints about sensitive matters, such as harassment, corruption allegations or whistleblowing, to any tool not cleared for them.

Step by step

1. Remove personal details from the complaint and keep the original on file.
2. Ask for a neutral summary. One or two sentences on the problem, the outcome requested and any dates mentioned.
3. Ask for a suggested category from your office's own list. Do not let the tool invent new categories.
4. Check the summary against the original. Make sure nothing the citizen said has been lost, softened or changed.
5. Decide the action yourself, using the facts on the file.
6. Ask for a draft reply that states your decision, the reasons and the next step. Give it the tone and format your office uses.
7. Check the reply line by line. Confirm every fact, date, rule and promise. AI can invent rules and timelines that sound real. See hallucination.
8. Put the real details back in yourself and sign it as your own.

Watch for bias and tone

Summaries can quietly change how a complaint reads. AI may treat an angry letter as less credible or a complaint written in Roman Urdu or a regional language as less serious. Translation errors can change meaning. See bias. Read the original whenever the matter is serious or the summary seems thin.

Replies should be respectful and plain. Avoid wording that blames the citizen. Never let a draft promise action, compensation or a deadline your office has not agreed.

What to check before sending

• Does the reply answer what the citizen actually asked?
• Is every fact correct and every rule cited checked in the original?
• Does it say what happens next and how the citizen can take the matter further? See redress.
• Has all placeholder text been replaced?
• Would you be comfortable if this reply were read out in public?

What not to use it for

• Deciding whether a complaint is upheld or rejected
• Ranking complaints by who the complainant is
• Automatic replies sent without an officer reading them

For ready-made wording, see the reply to a citizen complaint prompt and the summarise citizen feedback prompt.