Multilingual by necessity

At enterprise scale, a citizen assistant that works well in one language serves a fraction of the people it should. Multilingual AI is a requirement, not a feature.

People crossing a busy city street
Photo: Abhishek R on Unsplash

Government services are among the highest-volume conversations any institution has. Citizens and staff ask the same questions again and again: about eligibility, documents, deadlines and status. In a country of many languages and scripts, answering those questions well in only one or two languages leaves most people behind.

Grounded, not generative guesswork

The temptation is to put a general-purpose chatbot in front of the public. For government use, that is risky. An assistant must answer from approved documents, show which document it used and say clearly when it does not know.

For public services, a cited answer is worth more than a fluent one.

One knowledge baseTextSpeechDocumentsLocal languages
One knowledge base, answered in the language people use.

What a good citizen assistant needs

  • Answers grounded in approved circulars, rules and FAQs, with citations.
  • Understanding of questions in many languages, including mixed-language queries.
  • Clear escalation to a human official for anything sensitive or unclear.
  • An audit trail of every answer, so accuracy can be checked and improved.
  • Deployment on infrastructure the department controls.

Staff benefit first

Some of the quickest wins come from serving staff, not citizens. Document intelligence that can search, summarise and cross-check files, circulars and correspondence saves officials hours each week and builds confidence in the technology before it faces the public.

Measure it like a service

We judge these systems on answer accuracy, response time, time saved and audit-trail coverage. Those measures keep the focus where it belongs: on faster, fairer service.

Want to explore how this applies to your organisation? We share our approach with qualified organisations in a private briefing.