Skip to content
woodfrog
AI
Public sectorUse case, no client result

Answer citizen questions in their own language, at any hour.

Am I eligible? Which documents do I need? Where is my application? When is the office open? The same questions arrive all day, by phone, at the counter and by email, in many languages. Your staff repeat the same answers, citizens queue or give up, and the answer can depend on who they ask.

One citizen questionSchematic
3×Target return in 12 months
Week 1Fixed-fee audit
6 weeksTo the first live result
3Senior people on the work
TypeUse case: what we build, not a client story
ForDepartments and agencies that answer citizen questions about schemes, permits, certificates and applications
EngagementWeek 1 audit, then build in about six weeks, fixed scope, fixed fee
Runs onClaude, your published rules, your application records, chat and WhatsApp
Before and after

From queues and repeat calls to one clear answer.

Before
  • The same four questions fill the phone lines, counters and inbox
  • Citizens who speak Marathi or Hindi wait for the right person
  • Answers change depending on who picks up
  • Help stops when the office closes
After
  • Routine questions answered on chat or WhatsApp, at any hour
  • Each citizen answered in their own language
  • One answer, from your published rules, with the source shown
  • Anything unclear passed to a person, with a summary
The problem

The same questions, all day, in every language.

Most questions a department receives are not hard. They are about who qualifies, what to bring, how far an application has got and when to come in. But there are thousands of them, they arrive on every channel at once, and many are in a language only some of your staff speak. So the people who should be processing applications spend their day answering the phone.

People seated in a row, waiting their turn
Citizens wait for answers that are already written down somewhere
What was broken
  • Eligibility, documents, status and office hours make up most of what citizens ask.
  • The questions come by phone, at the counter and by email, so each one is answered several times over.
  • Only some staff speak each language, so citizens queue for them or give up.
  • Rules change, and not everyone hears at once, so two counters can give two different answers.
  • Nothing is recorded, so nobody can see which questions come up most or which rule confuses people.
You might be here if
  • Your phone lines are busiest with questions your website already answers.
  • Citizens come back a second time because a document was missing.
  • Staff ask colleagues to take calls in Marathi, Hindi or another language.
  • You hear that one office told a citizen something different from another.

Why it mattered. Every repeated answer is time taken from processing applications. Every wrong answer becomes a second visit, a complaint or a grievance.

The approach

Start with your questions and your rules, not with the agent.

In Week 1 we gather a sample of real questions from your phone logs, counters and inbox, and sort them by topic and language. Next to each we put the published rule, circular or record that answers it. That shows which questions an agent can answer from your own sources, and which must always reach a person. Then we build the agent on Claude for those questions only. If it cannot point to a source, it does not answer. It passes the question on.

A helpdesk officer with a headset, smiling while she listens
A person takes every question the agent is unsure of
What comes in
Citizen questionsPhone, counter, email
Many languagesEnglish, Hindi, Marathi, more
woodfrog method
Sort real questionsTopic, language, source
Agree what goes to a personWritten down, with your team
Build on ClaudeYour rules only, tested on your questions
What changes
Routine questions answered
Every answer sourced
Unsure cases handed over
One answer, in every language
Deliberately not doneLetting the agent decide eligibility or approve anything, or answer from anywhere but your own published rules
Your staff own the rules and take the hard cases. Claude answers, cites and summarises.schematic
The solution

A helpdesk that answers from your rules, and knows when to stop.

What we build
  • A helpdesk agent on Claude, on chat and WhatsApp, that answers in English, Hindi, Marathi and the other Indian languages your citizens use.
  • Answers drawn only from your department's published rules, circulars, document lists and office details. Every answer shows the rule it came from.
  • Application status looked up from your own records, once the citizen gives their application number, and explained in plain words.
  • A confidence check on every answer. Anything the agent is unsure of goes to a person, with a summary of the question and what the citizen has already said.
  • A simple way for your team to update the rules the agent reads, so a new circular changes every answer on the same day.
  • A log of every question and answer, with a daily list for a person to review and a stop button for the whole agent.
woodfrog is an Anthropic Certified Services Partner, and the agent runs on Claude. Citizens are told they are talking to an assistant, and any citizen can ask for a person at any point. In our public-sector case study, eight AI agents went live in citizen services with a trace to source, rule checks, a confidence score and a stop button each, and 0 audit gaps. Read how those controls work →
How it would runThe agent goes live on one service or one set of questions first, then widens. Operate is open-ended and drawn at one week only to fit.
The impact

What we would measure.

Where every question goesThree outcomes. None of them is a citizen left waiting with no answer.
Schematic, no values
The split depends on your questions and your rules. We measure it in Week 1, before anything is built.
What we would measure on a real engagementWhat we would measure
Questions answered by the agentBy topic, by language and by hour
Questions passed to a personAnd how long the citizen waited for them
Calls and counter visits for routine questionsBefore and after, on your own logs
Answers with a source shownAnd how often a source was out of date
Return on the workTarget: at least 3× within 12 months, measured on your own numbers
A woman checks an answer on her phone
An answer in her own language, at any hour
Answer every citizen, in their own language.A short call, or record a message. A person replies within one business day.
Contact usCertified Services Partner, Claude Partner Network