Public service intake

A citizen-facing intake bot with the audit trail public bodies need. A free public tier answers procedure questions on WhatsApp in the citizen's language (Arabic, French, English); an internal caseworker agent classifies requests and prepares case files for human officers. Every dispatch lands in the append-only intent ledger.

Why this shape

• Public tier ≠ privileged tier. The intake agent reads procedure docs and collects structured summaries — it holds no case data and no internal channel. A prompt-injected citizen message has nothing privileged to reach.
• Auditable by construction. When a case is contested, agentx ledger replay reconstructs exactly what was dispatched, when, and why — the compliance story most chatbot deployments can't tell.
• Humans decide. The caseworker agent prepares and queues; officers approve. Uncertainty routes to human review, not to a guess.
• Self-hosted. Citizen conversations stay on infrastructure the institution controls — no third-party SaaS processing.

This is the free/paid split in practice: the public intake tier costs little and serves everyone; the internal ops tier is where budget and value concentrate.

Start

Ready-to-run config: examples/civic/

cp examples/civic/agentx.json agentx.json
echo "TG_INTERNAL_BOT_TOKEN=..." >> .env
mkdir -p workspaces/{intake,caseworker}
agentx daemon start        # pair WhatsApp from the dashboard QR

Drop the service's procedure docs into workspaces/intake/references/.

Grow

• Formal multi-step applications with citizen forms → BPM — grant application
• FAQ wiki that compounds from real questions → shared wiki