Benefits navigation & public-facing chat

A retrieval-grounded navigation copilot built on verify-before-use principles — the cautious end of the design spectrum.

A benefits-advice copilot that never talks to the public: every AI draft is routed to a separate supervisor to approve, edit, or reject before it reaches even the adviser — the domain's positive counterexample, carrying a developer-run RCT and a live question about whether the every-message gate survives scaling to a whole national network.

Britain's biggest public test of generative AI in government: over roughly 18 months and two gated pilots, more than 10,000 people asked a retrieval-grounded assistant 26,000 questions about tax, benefits and visas — and GDS published the whole learning curve, from a 2023 version it held back for missing its accuracy bar to a self-reported 76 percent and then 90 percent measured accuracy, then launched it in the GOV.UK app in May 2026. The staged-gate counterexample to public chatbots that skipped the gates — with the honest catch that the operator grades its own accuracy.

The most heavily studied welfare-agency chatbot anywhere: Frida is the anonymous front door to Norway's national labour-and-welfare agency, and its payload is the moment a citizen can ask for a human. About four in five conversations end without one — but that is a completion rate, not an accuracy rate, and the escalation itself is a dial the agency turns, not the citizen: about one in five conversations reached a person under free choice, and about 30 percent when NAV removed the explicit human-chat option. The conversations that cross the boundary arrive degraded, and no per-answer accuracy rate has ever been published. The handover-boundary counterexample: the governance question is who sets how easy it is to reach a person, and whether anyone reads the conversations that ended without one.

Estonia's national network of public-sector chatbots — a federated 'chatbot of chatbots' in which each institution runs its own assistant, a central classifier routes between them, and one shared knowledge module feeds them all — and the atlas's clearest real instance of a governance question that is about the shared link, not the node.

A whole-of-government retire-and-replace: Singapore decommissioned Ask Jamie, the decade-old scripted assistant embedded on 70-plus agency websites as independent per-agency answer engines, and migrated government chatbots onto a small number of centrally provided large-language-model engines - a single decision that swapped roughly 80 uncorrelated error sources for one shared generative layer now fielding, per the government's own figures, over 800,000 citizen queries a month across 60-plus agencies. The domain's clean natural experiment in error-correlation structure, and its retire-and-replace lifecycle case.

A federal collection agency deployed a scripted chatbot, human live chat, and voice bots to deflect balance-due inquiries off its phone line, then expanded them and made live chat permanent — while, an inspector-general audit found, having no performance measures for the program at all and telemetry so unreliable it reported one assistor handling 603 chats at once against a system cap of three: the domain's monitoring-absence case, where the danger is not a wrong decision but an unmeasured one, and the only functioning oversight is an external, episodic audit.

France's flagship sovereign AI assistant for one-stop benefits counters gave the Prime Minister a wrong answer at its own launch, was quietly shelved within about eighteen months, and was formally denied generalization in January 2026 in its current form. Its distinctive feature is an absence: no error rate, usage or override figure was ever published, so the error-detection channel was the workforce and the unions, not a metric, and a cost-accounted review now gates its successor.

A consumer benefits app grounds its AI help on a state-verified deposit record it reads but never writes to, and escalates every AI dead-end to a human — a read-only design that severs the memory loop, on the strength of vendor-published pilot results only.

A nonprofit-built web form - not a government system, not a scoring model - that came to carry more than 70% of California's online SNAP applications while making no eligibility decisions at all: the domain's positive anchor for a critical navigation node whose safety came from a statutory determination floor it never touched, and whose single point of dependency was resolved by a planned, dated transfer into a state-owned portal rather than an abrupt collapse.

A twenty-year benefits-navigation nonprofit dissolved by its own board in 60 days — a node deletion that orphaned data-sharing links and left referral partners with nowhere to hand off.