At checkts.gpa.at, an AI legal chatbot that GPA, the Austrian trade union for private-sector employees, launched with Honeyfield and Studio C, 88 percent of users open the cited paragraph beneath each answer and 12 percent click through to the Federal Legal Information System (RIS) — the rate of source engagement the engineering team treats as the deployment's defining metric.
"Du brauchst kein besseres Modell, du brauchst einen besseren Loop." That sentence, from the Honeyfield team's post-deployment retrospective, captures the central design decision: the model is not trusted to recall citations from memory, so the chatbot cannot name a paragraph until it has searched the RIS and matched each reference against the search results. Citations that do not match are flagged and logged.
GPA wanted low-threshold first-tier legal information — not formal legal advice — for both members and non-members, sitting between a Google search for "kündigungsfrist österreich" and a booked lawyer appointment. Production analytics show the use case landed on hard topics: 27 percent of topic clicks are about overtime, 17 percent about termination, 16 percent about mobbing, 15 percent about salary questions, with roughly three quarters of usage in those four areas.
The deployment launched on checkts.gpa.at in 2026. Rather than a single end-of-project acceptance, the team ran continuous calibration with GPA's in-house legal experts throughout the test phase: where the agent's RIS research was too shallow, the response was deeper pre-answer searching, not better prompt wording.
The system is built in four layers. At the agent layer, every content question triggers a live RIS search before any paragraph can be cited, and each citation is verified against the search results. The legal integration was extracted into ris-mcp, an open-source MCP server (github.com/Honeyfield-Org/ris-mcp-ts) wrapping the Bundeskanzleramt's public RIS API with 13 tools — federal law, all nine federal states' law, and jurisprudence from 16 court jurisdictions. Honeyfield's AgentHub orchestrates the loop with auto-rollback on a failed health check; an AI Gateway adds PII detection, per-IP rate limits and content security policy. More than 80 governance rules are wired as structured code: mandatory disclaimers on notice-period questions, mandatory handoff on mobbing, prohibition of concrete legal recommendations, and the citation rule.
Under each answer sits a thumbs-up/down button, an expandable cited paragraph and a one-click path to the original RIS page. On termination, mobbing and insolvency topics the bot is structurally prevented from continuing without offering a transfer to GPA legal experts, and analytics show two thirds of handoffs originate from the sidebar, the rest from mobile and widget. The disclaimer UX — modal on first visit, persistent banner at top, warning emoji at the end of each answer — keeps the product from claiming to be legal advice. The example question "Wie lange ist meine Kündigungsfrist nach 8 Jahren im Job?" returns a structured list of statutory notice periods, cites § 20 Angestelltengesetz (Salaried Employees Act) and adds the note "3 Rechtsquellen geprüft" above the citation.
"Das System ist fertig im Sinne von produktionsreif, nicht fertig im Sinne von ausgelernt." The same architecture — RIS-grounded citation, structured governance, expert-in-the-loop calibration — is being extended to other regulated domains where a hallucinated paragraph hurts the same way: insurance, HR, tax advisory. The ris-mcp server is publicly available on GitHub for any team building on the RIS.