Citizen-facing policy assistant

A citizen-facing policy assistant explains published rules on benefits, permits, and entitlements in plain language. It is not a decision-maker. Quality means every substantive sentence traces to an approved policy paragraph. If the pack has no paragraph, that part of the reply stays empty. The bot does not invent eligibility, does not set a payment, and does not close a case.

Contact-centre leads need this for the questions that already have a desk-aid answer: waiting periods, which form to use, whether a permit can be renewed online, how to report a change of address. Those answers live in handbooks, gazetted circulars, and FAQs that legal has cleared. Retrieval-augmented generation restates that text so a caller can follow it. It is a different job from automated eligibility determination, which applies rules to a file. This page is about explaining the rule. The file still belongs to a human.

Hosting can sit in the same class of government platforms you already use, including Microsoft, Salesforce Government Cloud, CivicPlus, and Accela. Treat those products as a channel and a workspace. They do not make the corpus official. Your policy pack, citation format, and escalation rules do.

Name the sources before anyone types a prompt. Include the current benefit handbook, permit conditions, circulars still in force, and the public FAQ legal has cleared. Exclude campaign pages, news posts, staff chat exports, and any explainer a journalist or advocacy group wrote. If a blog post is in the index, the model will retrieve it and speak as if it were the rule. Answering from a blog is a quality failure even when the blog is roughly right.

Version every file. When a rate, waiting period, or form number changes, replace the paragraph and withdraw the old one from retrieval. A chat that still ranks last year's circular will give last year's answer with this year's confidence.

Chunk by heading and numbered paragraph, not by scanned page. The cite the citizen sees ("Income Support Manual, chapter 7, paragraph 12") is the product. Mixed chunks produce mixed cites, and mixed cites cannot be audited.

Do not pad the index to look complete. A gap in the handbook is an instruction to stay quiet, not a search problem to fix with extra documents.

Assign an owner. Policy, not IT, signs each file into the pack and signs it out when it is superseded. If two circulars conflict, retrieval will pick one at random from the model's point of view. Resolve the conflict in the source, then reindex.

The reply pattern is restatement, citation, stop. A safe shape looks like this: you may apply for the winter fuel supplement if you already receive the core living payment. Source: Income Support Manual, chapter 7, paragraph 12. That sentence explains published policy. It does not say you qualify, and it does not name a dollar figure unless the approved table prints one.

If the citizen asks how much, and the pack contains a published rate table, cite the table and the date it took effect. If the table says "as assessed," shows a range, or is silent, do not invent a dollar amount. Invented figures become promises. Promises become complaints.

Here is one illustrative exchange, not a measured rollout. A caller writes: I moved in with my sister last month. Do I still get the single-rate housing supplement, and how much is it this fortnight?

Retrieve the household-composition and residency rules. Quote the paragraph that defines single occupancy and the paragraph that says a change of address must be reported. If a rate table in the pack lists a published single rate, you may show that table with its cite. You must not say you still get it. You must not say you will be paid a specific amount this fortnight. Whether the household still counts as single occupancy is a determination. The amount after a change of circumstances is a determination. Both go to a caseworker.

If the pack has no paragraph on sharing a dwelling with a sibling, leave that part blank. Say the published rules do not cover this living arrangement, and offer a transfer or callback. Do not analogize from a shared-tenancy blog, a neighbouring jurisdiction's FAQ, or a training anecdote.

Empty stays empty. A missing rule is not permission to interpolate. "We do not have an approved paragraph on that" is a complete, high-quality answer.

Silence shows up in predictable places: new household types, overlapping permits, transition between old and new schemes, and anything that depends on facts the chat cannot see, such as income evidence, medical certificates, or tenancy contracts. The assistant may list documents the application completeness checker would look for, but only if that list sits in the approved pack. It must not say the claim will be granted once the documents arrive.

Sample transcripts the way you sample calls. Flag replies with no paragraph cite, invented waiting periods, and invented amounts. Those are defects whether or not the citizen said thanks. A supervisor who only reads the glowing ones will miss the invented dollar figure that has already gone out.

The chatbot does not determine the case. When the citizen asks "Do I qualify?", "Can you approve this?", or "What will I be paid?", restate any published criteria you can cite, then route.

Put the handoff in the work queue, not only in a footer disclaimer. Pass the thread, the reason (determination requested), and the citations already shown so the caseworker does not restart the conversation. If the contact centre already uses a case workspace from that platform class, pass the thread there. Do not make the citizen repeat themselves.

Treat the chat as a record of what was asked and what was cited. It is not the decision. If supervisors start treating a fluent bot reply as an approval, you have built a shadow determination process. That is the failure mode where the organisation believes the chat.

Once a human has decided, other delivery steps can use the same approved text. Proactive entitlement alerting can tell an existing claimant that a published rule change might affect them. A multilingual communication generator can turn the cited paragraph into a letter in the citizen's language. Both should run from the signed pack, not from an unsourced chatbot paraphrase.

Keep the public bot and the officer desk aid on the same corpus. If they diverge, citizens will quote the assistant against the person on the phone. One pack, one cite format, one path to a human.