Government agency enquiry and case desk

The problem

A citizen-facing agency handles enquiries across a portal form, a hotline, and WhatsApp. The policy that governs the answers sits in circulars, scheme criteria, and FAQ pages that change every budget cycle. Officers answer the same eligibility questions repeatedly, and each answer means opening two or three systems to check a case status. When an answer turns out to be wrong, there is no record of what produced it, so the agency cannot show how it was reached. Generic chatbots were ruled out: a confident wrong answer about benefit eligibility becomes a public accountability problem, and the vendor could not show where any answer came from.

The solution

One enquiry agent per scheme, grounded in the agency's own policy documents and given read-only access to the case system. It answers eligibility and status questions with a citation to the clause it used, handles routine transactions, and routes anything involving discretion or appeal to an officer with the context already assembled.

How Chocolate Factory does it

- Knowledge bases per scheme, with retrieval instructions, so every answer cites the governing clause and a policy change lands by re-syncing.

- The case system connects read-only: status lookups work, writes are impossible, and the flag is visible on the connection before an agent touches it.

- One workflow: chat trigger, a Branch across factual, transactional, and discretionary, then a hold before any write-back.

- The same agent serves the portal, WhatsApp, and Teams through /chat, with a project and scoped API keys per division.

- Per-run traces show the retrieval and the clause behind each answer, with tokens attributed by project.

What changes

An officer's first action becomes reviewing a drafted answer with the clause already cited, instead of hunting for the clause. Eligibility answers become reproducible, because the trace shows which version of which circular produced an answer on the day it was given. Policy changes take effect by re-syncing a knowledge base, so nobody has to retrain staff on a new script. Enquiries that arrive out of hours get a grounded answer or a queued case with context attached. And because spend is attributed by project, extending to a new channel becomes a decision with arithmetic behind it.